Study Curriculum (control plane)

# Study Curriculum (control plane)

AI Summary

Purpose:

• Backlog of study chapters the daily routine pulls from. One row = one planned chapter.

Key points:

• Status: pending | draft | published. Routine picks the top pending (or a user-requested item).
• Topics seeded from the repo tag vocabulary (the user's real stack).
• After generating, the routine flips the row to published with a date and adds the chapter to the series' series.json.
• When no pending row remains after pulling, the scheduled routine uses the dynamic AI/DB fallback policy below instead of stopping.

Relevant when:

• The daily study routine runs, or the user requests a specific next chapter.

Priority queue

User-requested items go to the TOP and are taken first.

Remember public Data Engineering gaps — requested 2026-08-07

• [x] 1. Entity resolution and MDM: matching, review bands, golden record, provenance, and rollback — published 2026-08-07
• [x] 2. Full Load and CDC handoff: Debezium, Kafka Connect, upsert, and reconciliation — published 2026-08-07
• [x] 3. Amazon S3 Tables operations: compaction, snapshot retention, and orphan-file removal — published 2026-08-07
• [x] 4. PyIceberg observability: snapshot, manifest, data-file, and small-file metrics — published 2026-08-07
• [x] 5. Lake Formation governance: IAM, table/column/row access, ownership, and audit — published 2026-08-07
• [x] 6. AWS Glue PySpark Full Load: replica isolation, JDBC partitioning, type conversion, and verification — published 2026-08-07
• [x] 7. StarRocks with Iceberg serving: low-latency OLAP, materialized views, freshness, and reconciliation — published 2026-08-07

Dynamic AI/DB fallback policy

Use this policy only when the freshly pulled curriculum has zero pending rows.

• The 04:00 KST job selects a recent AI systems or AI engineering topic.
• The 05:00 KST job selects a recent database or data-platform topic.
• The 06:00 KST job selects the most technically consequential AI/DB topic not already covered that day.
• "Recent" means announced or materially updated within the last 90 days. Widen to 180 days only when no qualified topic exists, and state that decision in the chapter.
• Prefer primary evidence: official documentation, release notes, upstream repositories, standards, and papers. Use at least two sources, including at least one primary source.
• Do not publish rumor roundups, product marketing rewrites, unsupported benchmark claims, or a topic already covered without a material new technical change.
• Search ai/study/curriculum.md, human/study/content/**, and ai/wiki/** before selecting a topic.
• Write a durable Korean technical blog chapter, not a news summary. Explain what changed, architecture or mechanism, operational impact, tradeoffs and limits, and a verification or adoption checklist.
• Record the generated topic as a completed row under Dynamic fallback publications and publish it in either the ai-frontier or database-frontier series. Create a complete series.json when the selected series does not exist.

DBA / Data Platform Engineer deepening track — requested 2026-06-26

Intent:

• Treat the learner as an operator who must keep databases and data platforms reliable, observable, recoverable, and cost-aware.
• Prefer chapters that connect internals to incident response, capacity planning, migration safety, and platform design tradeoffs.
• Keep examples grounded in MySQL, PostgreSQL, Kafka, Airflow, Spark/Flink, ClickHouse, Redis, Kubernetes, Terraform, and observability.

Recommended next order:

1. database-reliability — incidents, SLOs, backup/restore drills, HA, replication lag, capacity.
2. sql-performance-engineering — query plans, indexing, join strategy, cardinality, workload profiling.
3. mysql-advanced-operations — replication, online DDL, PITR, connection storms, upgrade planning.
4. postgresql-advanced-operations — vacuum, bloat, WAL, logical replication, autovacuum tuning.
5. data-platform-operations — ingestion contracts, data quality, lineage, governance, platform runbooks.
6. lakehouse-table-formats — Iceberg/Delta/Hudi, file layout, compaction, schema evolution.
7. streaming-data-platform — Kafka/Flink operational semantics, exactly-once boundaries, replay.

Series backlog

kubernetes (DB/Infra) — ☸

• [x] 1. 컨테이너 기초 — published 2026-06-10
• [x] 2. 클러스터 아키텍처 — published 2026-06-10
• [x] 3. Pod 스케줄링과 리소스 — published 2026-06-10
• [x] 4. 네트워킹과 Service — published 2026-06-10
• [x] 5. 스토리지와 PV/PVC — published 2026-06-10
• [x] 6. ConfigMap/Secret과 설정 — published 2026-06-10
• [x] 7. 헬스체크와 롤아웃 전략 — published 2026-06-10
• [x] 8. 오토스케일링(HPA/VPA) — published 2026-06-10
• [x] 9. 관측성(로깅·메트릭) — published 2026-06-10
• [x] 10. 운영 트러블슈팅 패턴 — published 2026-06-10

mysql-core (MySQL) — 🐬

• [x] 1. 인덱스와 InnoDB 스토리지 엔진 — published 2026-06-15
• [x] 2. 트랜잭션과 잠금 — published 2026-06-15
• [x] 3. 쿼리 옵티마이저와 실행계획 읽기 — published 2026-06-15

mysql-binlog (MySQL) — 🐬

• [x] 1. 바이너리 로그란 무엇인가 — published 2026-06-10
• [x] 2. binlog 포맷(STATEMENT/ROW/MIXED) — published 2026-06-10
• [x] 3. 복제와 GTID — published 2026-06-11
• [x] 4. CDC와 binlog 기반 동기화 — published 2026-06-11
• [x] 5. ZSTD 압축과 8.4 변화점 — published 2026-06-11
• [x] 6. binlog 운영·보존·모니터링 — published 2026-06-11

airflow (데이터파이프라인) — 🌊

• [x] 1. DAG/Task/Operator 개념 — published 2026-06-11
• [x] 2. 스케줄링과 backfill — published 2026-06-11
• [x] 3. XCom과 태스크 통신 — published 2026-06-12
• [x] 4. 실행기(Executor) 비교 — published 2026-06-12
• [x] 5. 운영(재시도·SLA·알림) — published 2026-06-12

vulnerability-collection (보안) — 🔎

• [x] 1. CVE/OSV 데이터 모델 — published 2026-06-12
• [x] 2. OS 패키지 vs 라이브러리 취약점 — published 2026-06-12
• [x] 3. 버전 비교와 fixed_version 처리 — published 2026-06-12
• [x] 4. 증분 수집 전략 — published 2026-06-13

kafka (데이터파이프라인) — 📨

• [x] 1. Kafka 아키텍처와 핵심 개념 — published 2026-06-13
• [x] 2. 프로듀서와 컨슈머 설계 — published 2026-06-13
• [x] 3. 파티셔닝과 오프셋 관리 — published 2026-06-13
• [x] 4. 내구성과 복제 메커니즘 — published 2026-06-13
• [x] 5. Kafka Streams와 ksqlDB — published 2026-06-13
• [x] 6. 운영과 모니터링 — published 2026-06-13

data-platform (DB/Infra) — 🏗️

• [x] 1. 데이터 플랫폼 큰 그림 — published 2026-06-14
• [x] 2. OLTP와 OLAP: 운영 DB와 분석 DB를 나누는 이유 — published 2026-06-14
• [x] 3. Data Lake, Lakehouse, Medallion Architecture — published 2026-06-14
• [x] 4. Data Warehouse와 모델링 기본기 — published 2026-06-14
• [x] 5. Data Mart, Semantic Layer, BI — published 2026-06-14
• [x] 6. ETL, ELT, CDC, Orchestration — published 2026-06-14
• [x] 7. Governance, Catalog, Lineage, Data Quality — published 2026-06-14
• [x] 8. DBA·DB 플랫폼 엔지니어 운영 체크리스트 — published 2026-06-14

postgresql (MySQL) — 🐘

• [x] 1. PostgreSQL 아키텍처 개요 — published 2026-06-14
• [x] 2. MVCC와 트랜잭션 관리 — published 2026-06-14
• [x] 3. 인덱스 설계와 실행계획 읽기 — published 2026-06-14
• [x] 4. 잠금(Lock)과 데드락 처리 — published 2026-06-14
• [x] 5. WAL과 복제(Streaming Replication) — published 2026-06-14
• [x] 6. 백업·복구(pg_dump, PITR) — published 2026-06-14
• [x] 7. 성능 튜닝과 쿼리 최적화 — published 2026-06-15
• [x] 8. 운영 트러블슈팅 패턴 — published 2026-06-15

redis (DB/Infra) — ⚡

• [x] 1. Redis 아키텍처와 핵심 자료구조 — published 2026-06-15
• [x] 2. 지속성(RDB/AOF)과 메모리 관리 — published 2026-06-15
• [x] 3. 복제와 센티넬(고가용성) — published 2026-06-16
• [x] 4. 클러스터 모드와 샤딩 — published 2026-06-16
• [x] 5. 캐싱 패턴과 운영 트러블슈팅 — published 2026-06-16

clickhouse (DB/Infra) — 🏛️

• [x] 1. ClickHouse 아키텍처와 핵심 개념 — published 2026-06-16
• [x] 2. MergeTree 엔진 패밀리와 인덱스 — published 2026-06-16
• [x] 3. 분산 아키텍처: 샤딩과 복제 — published 2026-06-16
• [x] 4. 쿼리 최적화와 실행계획 읽기 — published 2026-06-16
• [x] 5. 스키마 설계 패턴과 Materialized View — published 2026-06-16
• [x] 6. 운영과 모니터링 — published 2026-06-16

elasticsearch (DB/Infra) — 🔍

• [x] 1. Elasticsearch 아키텍처와 핵심 개념 — published 2026-06-17
• [x] 2. 역 인덱스 심화와 분석기(Analyzer) 설계 — published 2026-06-17
• [x] 3. 매핑(Mapping)과 스키마 설계 전략 — published 2026-06-17
• [x] 4. 쿼리 DSL 심화와 집계(Aggregation) — published 2026-06-17
• [x] 5. 성능 튜닝과 운영 트러블슈팅 — published 2026-06-18

spark (데이터파이프라인) — ⚡

• [x] 1. Apache Spark 아키텍처와 핵심 개념 — published 2026-06-17
• [x] 2. RDD·DataFrame·Dataset API 비교 — published 2026-06-17
• [x] 3. 파티셔닝과 셔플 최적화 — published 2026-06-18
• [x] 4. Spark SQL과 Catalyst 옵티마이저 — published 2026-06-18
• [x] 5. 구조적 스트리밍(Structured Streaming) — published 2026-06-18
• [x] 6. 운영과 성능 모니터링 — published 2026-06-18

dbt (데이터파이프라인) — 🔧

• [x] 1. dbt 기초와 아키텍처 — published 2026-06-18
• [x] 2. 모델, 소스, Ref 참조 — published 2026-06-18
• [x] 3. 테스트와 문서화 — published 2026-06-18
• [x] 4. 증분 모델(Incremental Models) — published 2026-06-18
• [x] 5. 배포와 CI/CD — published 2026-06-19

prometheus-grafana (DB/Infra) — 📊

• [x] 1. Prometheus 아키텍처와 메트릭 수집 — published 2026-06-19
• [x] 2. PromQL과 알림 규칙 — published 2026-06-19
• [x] 3. Grafana 대시보드 구성과 시각화 — published 2026-06-19

terraform (DB/Infra) — 🏗️

• [x] 1. Terraform 기초와 IaC 개념 — published 2026-06-19
• [x] 2. State 심화 관리와 원격 백엔드 — published 2026-06-19
• [x] 3. 모듈 설계와 코드 구조화 — published 2026-06-19
• [x] 4. 프로바이더와 리소스 의존성 관리 — published 2026-06-20
• [x] 5. CI/CD 통합과 팀 협업 전략 — published 2026-06-20
• [x] 6. 보안 강화와 Terraform 운영 패턴 — published 2026-06-20

opentelemetry (DB/Infra) — 🔭

• [x] 1. OpenTelemetry 개요와 아키텍처 — published 2026-06-20
• [x] 2. 트레이싱(Tracing)과 스팬(Span) 심화 — published 2026-06-20
• [x] 3. 메트릭 API와 계측(Instrumentation) — published 2026-06-20
• [x] 4. 로그와 구조화된 로깅 — published 2026-06-20
• [x] 5. OTel Collector 구성과 파이프라인 — published 2026-06-20
• [x] 6. 실전 운영과 성능 고려사항 — published 2026-06-20

apache-flink (데이터파이프라인) — 🌊

• [x] 1. Apache Flink 아키텍처와 핵심 개념 — published 2026-06-21
• [x] 2. DataStream API와 상태(State) 심화 — published 2026-06-21
• [x] 3. 윈도(Window)와 시간 의미론 실전 — published 2026-06-21
• [x] 4. Flink SQL과 Table API — published 2026-06-21
• [x] 5. 체크포인트·세이브포인트와 Exactly-Once — published 2026-06-22
• [x] 6. 운영과 성능 모니터링 — published 2026-06-22

grpc (DB/Infra) — 🔗

• [x] 1. gRPC와 Protocol Buffers 개요 — published 2026-06-21
• [x] 2. 스트리밍 RPC 패턴 심화 — published 2026-06-21
• [x] 3. 인터셉터·메타데이터·에러 처리 — published 2026-06-22
• [x] 4. 서비스 디스커버리와 로드밸런싱 — published 2026-06-22
• [x] 5. 보안(TLS/mTLS)과 인증 — published 2026-06-22
• [x] 6. 실전 마이크로서비스 설계 패턴 — published 2026-06-22

golang (DB/Infra) — 🐹

• [x] 1. Go 기초 문법과 설계 철학 — published 2026-06-22
• [x] 2. 고루틴(Goroutine)과 채널(Channel) — published 2026-06-22
• [x] 3. 인터페이스와 타입 시스템 — published 2026-06-22
• [x] 4. 에러 처리 심화와 패닉·복구 — published 2026-06-23
• [x] 5. 표준 라이브러리와 net/http — published 2026-06-23
• [x] 6. 테스트와 벤치마킹 — published 2026-06-23

python (DB/Infra) — 🐍

• [x] 1. Python 기초와 설계 철학 — published 2026-06-23
• [x] 2. 자료구조·컬렉션·이터레이터 심화 — published 2026-06-23
• [x] 3. 함수형 도구와 클로저·데코레이터 — published 2026-06-23
• [x] 4. 클래스와 메타프로그래밍 — published 2026-06-23
• [x] 5. 비동기 프로그래밍(asyncio) — published 2026-06-23
• [x] 6. 타입 힌팅과 정적 분석 도구 — published 2026-06-23

linux (DB/Infra) — 🐧

• [x] 1. 커널 아키텍처와 시스템 콜 — published 2026-06-24
• [x] 2. 프로세스와 스레드 관리 — published 2026-06-24
• [x] 3. 메모리 관리와 가상 메모리 — published 2026-06-24
• [x] 4. 파일시스템과 VFS — published 2026-06-24
• [x] 5. 네트워킹 스택 — published 2026-06-24
• [x] 6. 성능 분석 도구(perf/strace/eBPF) — published 2026-06-24

docker (DB/Infra) — 🐳

• [x] 1. Docker 아키텍처와 컨테이너 런타임 — published 2026-06-25
• [x] 2. Dockerfile 최적화와 멀티스테이지 빌드 — published 2026-06-25
• [x] 3. 네트워킹과 볼륨 심화 — published 2026-06-25
• [x] 4. 보안 강화: rootless·seccomp·AppArmor — published 2026-06-25
• [x] 5. Docker Compose와 멀티컨테이너 앱 — published 2026-06-25
• [x] 6. 레지스트리와 이미지 관리 — published 2026-06-25

database-reliability (DB/Infra) — 🧯

• [x] 1. 데이터베이스 운영자의 책임 범위: 가용성, 내구성, 성능, 비용 — published 2026-06-26
• [x] 2. 장애 분류와 SLO: 접속 장애, 지연, 데이터 손상, 복제 지연 — published 2026-06-26
• [x] 3. 백업 전략: full/incremental, logical/physical, snapshot, 보존 정책 — published 2026-06-26
• [x] 4. 복구 전략: RTO/RPO, PITR, 복구 리허설, 복구 후 검증 — published 2026-06-26
• [x] 5. 고가용성 패턴: primary-replica, failover, quorum, split-brain 방지 — published 2026-06-26
• [x] 6. 용량 계획: CPU, memory, IOPS, connection, disk growth 예측 — published 2026-06-26
• [x] 7. 변경 관리: DDL, 인덱스, 파라미터 변경의 위험도와 롤백 플랜 — published 2026-06-26
• [x] 8. 데이터베이스 사고 대응 런북: 탐지, 완화, 원인분석, 재발방지 — published 2026-06-27

sql-performance-engineering (DB/Infra) — 🧠

• [x] 1. SQL 성능 튜닝 사고법: latency, throughput, cardinality, selectivity — published 2026-06-27
• [x] 2. 실행계획 읽기 공통 패턴: scan, seek, join, sort, aggregate — published 2026-06-27
• [x] 3. 인덱스 설계 심화: composite, covering, partial, functional index — published 2026-06-27
• [x] 4. 조인 최적화: nested loop, hash join, merge join과 통계 정보 — published 2026-06-27
• [x] 5. 느린 쿼리 분석: slow log, pg_stat_statements, Performance Schema — published 2026-06-27
• [x] 6. 쓰기 성능: batch, transaction size, lock contention, hot row — published 2026-06-27
• [x] 7. 페이징과 대량 조회: offset pagination, keyset pagination, cursor — published 2026-06-27
• [x] 8. 튜닝 검증: before/after 측정, plan regression, 부작용 확인 — published 2026-06-27

mysql-advanced-operations (MySQL) — 🐬

• [x] 1. MySQL 운영 아키텍처: InnoDB, redo/undo, buffer pool, doublewrite — published 2026-06-28
• [x] 2. Replication 운영: GTID, semi-sync, lag 원인, replica read 안전성 — published 2026-06-28
• [x] 3. Online DDL과 스키마 변경: instant/inplace/copy, gh-ost, pt-osc — published 2026-06-28
• [x] 4. 백업·복구 실전: mysqldump, xtrabackup, binlog PITR — published 2026-06-28
• [x] 5. 잠금과 장애: metadata lock, gap lock, deadlock, long transaction — published 2026-06-28
• [x] 6. Connection storm과 pool 운영: max_connections, thread, timeout — published 2026-06-28
• [x] 7. 업그레이드 계획: 5.7→8.0/8.4, charset/collation, optimizer 변화 — published 2026-06-28
• [x] 8. MySQL 운영 대시보드: 핵심 지표와 알림 임계값 — published 2026-06-28

postgresql-advanced-operations (DB/Infra) — 🐘

• [x] 1. PostgreSQL 운영 아키텍처: process model, shared buffers, WAL — published 2026-06-28
• [x] 2. Vacuum과 bloat: autovacuum 튜닝, freeze, xid wraparound 방지 — published 2026-06-29
• [x] 3. WAL·Checkpoint 튜닝: write amplification, archive, replication slot — published 2026-06-29
• [x] 4. Logical replication과 CDC: publication/subscription, slot lag, 재동기화 — published 2026-06-29
• [x] 5. Partitioning 운영: 범위/리스트 파티션, detach/drop, pruning — published 2026-06-29
• [x] 6. Connection 관리: pgbouncer, transaction pooling, prepared statement 주의점 — published 2026-06-29
• [x] 7. PITR과 standby 운영: base backup, timeline, failover/rejoin — published 2026-06-29
• [x] 8. PostgreSQL 운영 대시보드: 핵심 지표와 알림 임계값 — published 2026-06-29

data-platform-operations (데이터파이프라인) — 🏗️

• [x] 1. 데이터 플랫폼 운영 모델: source, ingestion, storage, transform, serving — published 2026-06-29
• [x] 2. 데이터 계약(Data Contract): schema, SLA, owner, breaking change — published 2026-06-29
• [x] 3. 데이터 품질: completeness, freshness, uniqueness, validity, reconciliation — published 2026-06-30
• [x] 4. Lineage와 Catalog 운영: column-level lineage, ownership, 영향도 분석 — published 2026-06-30
• [x] 5. 배치 파이프라인 운영: backfill, retry, idempotency, checkpoint — published 2026-06-30
• [x] 6. 비용 관리: storage lifecycle, compute sizing, query cost attribution — published 2026-06-30
• [x] 7. 개인정보·보안: masking, encryption, access control, audit log — published 2026-06-30
• [x] 8. 플랫폼 런북: 장애 triage, 지연 원인분석, 재처리 기준 — published 2026-06-30

lakehouse-table-formats (데이터파이프라인) — 🧊

• [x] 1. Lakehouse가 필요한 이유: file lake의 한계와 transaction log — published 2026-06-30
• [x] 2. Apache Iceberg 핵심: snapshot, manifest, partition evolution — published 2026-06-30
• [x] 3. Delta Lake 핵심: transaction log, time travel, optimize/vacuum — published 2026-06-30
• [x] 4. Apache Hudi 핵심: copy-on-write, merge-on-read, incremental pull — published 2026-07-01
• [x] 5. 파일 레이아웃 최적화: small files, compaction, clustering, partitioning — published 2026-07-01
• [x] 6. Schema evolution과 compatibility: add/drop/rename, reader/writer contract — published 2026-07-01
• [x] 7. Query engine 연동: Spark, Trino, Flink, ClickHouse 외부 테이블 — published 2026-07-01
• [x] 8. 운영 체크리스트: metadata growth, compaction 주기, rollback 전략 — published 2026-07-01

streaming-data-platform (데이터파이프라인) — 🌊

• [x] 1. 스트리밍 플랫폼 설계: event time, processing time, late data — published 2026-07-01
• [x] 2. Kafka 운영 심화: partition sizing, consumer lag, rebalance, retention — published 2026-07-01
• [x] 3. Schema Registry와 호환성: Avro/Protobuf/JSON Schema, evolution policy — published 2026-07-01
• [x] 4. Exactly-once의 실제 경계: producer idempotence, transaction, sink semantics — published 2026-07-01
• [x] 5. Replay와 재처리: offset reset, DLQ, poison message, idempotent sink — published 2026-07-02
• [x] 6. Flink 상태 운영: state backend, checkpoint, savepoint, rescaling — published 2026-07-02
• [x] 7. 스트리밍 관측성: lag, watermark, throughput, backpressure, end-to-end latency — published 2026-07-02
• [x] 8. 장애 대응 패턴: broker failure, bad deploy, schema break, downstream outage — published 2026-07-02

orchestration-advanced (데이터파이프라인) — 🕰️

• [x] 1. 오케스트레이션 설계 원칙: DAG 경계, idempotency, dependency contract — published 2026-07-02
• [x] 2. Airflow 운영 심화: scheduler, executor, metadata DB, pool, queue — published 2026-07-02
• [x] 3. Backfill 안전장치: data interval, catchup, replay window, 중복 방지 — published 2026-07-02
• [x] 4. Sensor와 event-driven orchestration: deferrable operator, dataset scheduling — published 2026-07-02
• [x] 5. 실패 처리: retry budget, alert fatigue, partial failure, compensation — published 2026-07-02
• [x] 6. 배포와 테스트: DAG validation, local test, CI, rollback — published 2026-07-03
• [x] 7. Metadata DB 운영: vacuum/analyze, cleanup, migration, connection pool — published 2026-07-03
• [x] 8. 운영 대시보드: SLA miss, queue delay, task duration, scheduler health — published 2026-07-03

data-modeling-analytics (데이터파이프라인) — 📐

• [x] 1. 분석 모델링 기본: grain, fact, dimension, slowly changing dimension — published 2026-07-03
• [x] 2. Star/Snowflake schema와 wide table tradeoff — published 2026-07-03
• [x] 3. Metric layer 설계: 정의 충돌, semantic model, ownership — published 2026-07-03
• [x] 4. dbt 모델링 심화: staging/intermediate/mart, incremental, snapshot — published 2026-07-03
• [x] 5. 품질 테스트: source freshness, accepted values, relationship, custom test — published 2026-07-03
• [x] 6. 변경 영향도: lineage 기반 리뷰, downstream dashboard 보호 — published 2026-07-03
• [x] 7. 성능 최적화: partition, clustering, materialization, aggregate table — published 2026-07-04
• [x] 8. 모델 운영: deprecation, migration, documentation, consumer communication — published 2026-07-04

security-compliance-data (보안) — 🛡️

• [x] 1. 데이터 보안 기본: classification, least privilege, separation of duties — published 2026-07-04
• [x] 2. 인증·인가: RBAC/ABAC, service account, credential rotation — published 2026-07-04
• [x] 3. 암호화: at-rest, in-transit, KMS, envelope encryption — published 2026-07-04
• [x] 4. 개인정보 처리: masking, tokenization, pseudonymization, retention — published 2026-07-04
• [x] 5. 감사와 추적성: audit log, query log, access review — published 2026-07-04
• [x] 6. 백업 보안: immutable backup, ransomware 대응, restore 권한 분리 — published 2026-07-04
• [x] 7. 규정 대응 실무: evidence collection, control mapping, exception 관리 — published 2026-07-04
• [x] 8. 데이터 플랫폼 보안 사고 대응: 유출 탐지, 격리, 영향도 분석 — published 2026-07-05

distributed-systems (DB/Infra) — 🌐

• [x] 1. 일관성 모델 기초: linearizability, sequential consistency, causal consistency, eventual consistency — published 2026-07-05
• [x] 2. CAP 정리와 PACELC: 분산 시스템 트레이드오프 실전 — published 2026-07-05
• [x] 3. 복제와 합의: Raft, Paxos, 리더 선출 — published 2026-07-05
• [x] 4. 분산 트랜잭션: 2PC, Saga, TCC — published 2026-07-05
• [x] 5. 벡터 클락과 인과성: 이벤트 순서와 분산 추적 — published 2026-07-05
• [x] 6. 분산 스토리지 패턴: 샤딩, 파티셔닝, 데이터 지역성 — published 2026-07-05
• [x] 7. 장애 감지와 복구 패턴: heartbeat, circuit breaker, bulkhead — published 2026-07-06
• [x] 8. 운영자를 위한 분산 시스템 판단 기준 — published 2026-07-06

sre-practices (DB/Infra) — 🚨

• [x] 1. SRE 기초와 에러 버짓: 신뢰성 공학의 원칙 — published 2026-07-06
• [x] 2. SLI/SLO/SLA 실전: 정의, 측정, 에러 버짓 소비 — published 2026-07-06
• [x] 3. 토일(Toil) 제거와 자동화: SRE의 핵심 효율화 원칙 — published 2026-07-06
• [x] 4. 인시던트 관리와 포스트모텀: 장애 대응의 체계적 접근 — published 2026-07-06
• [x] 5. 온콜 운영과 번아웃 방지: 에스컬레이션, 로테이션, 알림 피로 — published 2026-07-06
• [x] 6. 카오스 엔지니어링: 실패 주입과 복원력 검증 — published 2026-07-06

storage-and-io (DB/Infra) — 💾

• [x] 1. 스토리지 시스템 기초: 블록·파일·오브젝트 스토리지와 I/O 스택 — published 2026-07-07
• [x] 2. NVMe·SSD·HDD 심화: 내부 구조, 마모 관리, ZNS — published 2026-07-07
• [x] 3. 파일시스템 선택과 마운트 최적화: ext4, XFS, Btrfs — published 2026-07-07
• [x] 4. 클라우드 블록 스토리지: EBS, Persistent Disk, Azure Disk 운영 — published 2026-07-07
• [x] 5. 오브젝트 스토리지 운영: S3 라이프사이클, 멀티파트, 성능 패턴 — published 2026-07-08
• [x] 6. RAID와 스토리지 이중화: 패리티, 핫스페어, 복구 — published 2026-07-08
• [x] 7. 스토리지 성능 벤치마킹: fio, sysbench, 실측 기반 용량 계획 — published 2026-07-08
• [x] 8. 데이터베이스별 스토리지 튜닝 체크리스트 — published 2026-07-08

ci-cd-data-platform (데이터파이프라인) — 🔄

• [x] 1. CI/CD 파이프라인 기초: 데이터 플랫폼 팀의 배포 자동화 — published 2026-07-07
• [x] 2. dbt CI/CD 심화: Slim CI, state, 환경 분리, 아티팩트 관리 — published 2026-07-07
• [x] 3. Airflow DAG 배포 자동화: GitSync, 패키지 배포, 버전 관리 — published 2026-07-08
• [x] 4. 스키마 변경 안전 배포: 하위 호환, 마이그레이션, Contract 검증 — published 2026-07-08
• [x] 5. 데이터 파이프라인 테스트 전략: 단위·통합·E2E 테스트 — published 2026-07-08
• [x] 6. GitOps와 데이터 플랫폼: ArgoCD, FluxCD, Infrastructure as Code — published 2026-07-08
• [x] 7. 배포 후 모니터링: SLO 연동, 자동 롤백 트리거, 배포 이력 — published 2026-07-08
• [x] 8. 팀 운영: 브랜치 전략, 리뷰 프로세스, 배포 권한 관리 — published 2026-07-09

cloud-native-databases (DB/Infra) — ☁️

• [x] 1. 클라우드 관리형 DB 서비스 개요: 자체 운영 vs 매니지드 트레이드오프 — published 2026-07-09
• [x] 2. AWS RDS와 Aurora 심화: Multi-AZ, 읽기 복제본, Aurora 스토리지 레이어 — published 2026-07-09
• [x] 3. Google Cloud SQL과 AlloyDB 심화: HA, 컬럼형 엔진, 연결 관리 — published 2026-07-09
• [x] 4. Azure Database 서비스: MySQL/PostgreSQL Flexible Server 운영 — published 2026-07-09
• [x] 5. 클라우드 DB 마이그레이션: 온프레미스→클라우드 이관 전략과 실행 — published 2026-07-09
• [x] 6. 클라우드 DB 비용 최적화: 인스턴스·스토리지·I/O 비용 관리 — published 2026-07-09

networking-fundamentals (DB/Infra) — 📡

• [x] 1. TCP/IP와 OSI 모델 — 패킷이 데이터베이스까지 가는 길 — published 2026-07-10
• [x] 2. TLS/HTTPS 동작 원리와 인증서 관리 — published 2026-07-10
• [x] 3. DNS 심화: 조회 과정, 캐싱, TTL, 도메인 설계 — published 2026-07-10
• [x] 4. 로드밸런서: L4/L7, 헬스체크, 세션 어피니티 — published 2026-07-10
• [x] 5. 방화벽과 보안 그룹: iptables, nftables, VPC 규칙 — published 2026-07-10
• [x] 6. 네트워크 성능 분석: iperf3, fio 연동, 병목 진단 — published 2026-07-11

query-federation (데이터파이프라인) — 🔀

• [x] 1. Trino 아키텍처와 분산 쿼리 실행: 쿼리 페더레이션의 원리 — published 2026-07-10
• [x] 2. Trino 커넥터 심화: Iceberg, MySQL, Kafka 커넥터 운영 — published 2026-07-11
• [x] 3. 쿼리 성능 튜닝: EXPLAIN ANALYZE, CBO, 파티셔닝 전략 — published 2026-07-11
• [x] 4. 보안과 접근 제어: LDAP, OAuth2, 파인-그레인드 ACL — published 2026-07-11
• [x] 5. 운영 대시보드와 장애 대응: 메모리, GC, 스필, 재시도 — published 2026-07-11
• [x] 6. DuckDB와 로컬 OLAP: 단일 프로세스 분석 엔진의 활용 — published 2026-07-11

time-series-databases (DB/Infra) — ⏱️

• [x] 1. 시계열 데이터베이스 개요: 왜 범용 DB로는 부족한가 — published 2026-07-11
• [x] 2. InfluxDB 아키텍처와 IOx 스토리지 엔진 — published 2026-07-12
• [x] 3. TimescaleDB: PostgreSQL 위에서 시계열 처리 — published 2026-07-12
• [x] 4. VictoriaMetrics: Prometheus 호환 고성능 TSDB — published 2026-07-12
• [x] 5. 시계열 데이터 모델링: cardinality, retention, downsampling — published 2026-07-12
• [x] 6. 운영과 모니터링 체크리스트 — published 2026-07-12

vector-databases (DB/Infra) — 🔮

• [x] 1. 벡터 데이터베이스 개요: 임베딩, ANN 알고리즘, 주요 엔진 비교 — published 2026-07-12
• [x] 2. pgvector 심화: HNSW 튜닝, 필터링, 대규모 운영 — published 2026-07-13
• [x] 3. Qdrant 심화: 세그먼트, 페이로드 인덱스, 클러스터 운영 — published 2026-07-13
• [x] 4. Milvus 심화: 분산 아키텍처, 인덱스 전략, K8s 배포 — published 2026-07-13
• [x] 5. 벡터 검색 성능 튜닝: ef_search, recall 측정, 벤치마킹 — published 2026-07-13
• [x] 6. RAG 파이프라인 설계: 청킹, 임베딩, 재순위화, 운영 체크리스트 — published 2026-07-13

Dynamic fallback publications

Append fallback-generated rows here only when no planned pending chapter existed before topic selection.

• [x] ai-frontier/1: LLM 추론 최적화: KV 캐시, 양자화, 투기적 디코딩의 원리와 운영 — published 2026-07-13 (dynamic fallback: AI engineering topic)
• [x] ai-frontier/2: vLLM 0.25 Model Runner V2: 비동기 추론 경로의 재설계와 마이그레이션 — published 2026-07-14 (dynamic fallback: AI systems topic; v0.25.0 released 2026-07-11)
• [x] database-frontier/1: Airflow 3.3 State Store: 재시도와 실행을 넘어 상태를 어디에 둘 것인가 — published 2026-07-14 (dynamic fallback: data-platform topic; Airflow 3.3.0 released 2026-07-06)
• [x] ai-frontier/3: PyTorch 2.13 LinearCrossEntropyLoss: 거대한 logits를 만들지 않는 학습 손실 — published 2026-07-14 (dynamic fallback: AI systems topic; PyTorch 2.13.0 released 2026-07-08)
• [x] ai-frontier/4: SGLang RadixAttention: KV 캐시를 트리로 공유하는 LLM 추론 아키텍처 — published 2026-07-14 (dynamic fallback: AI engineering topic; RadixAttention prefix caching and MLA backend architecture)
• [x] database-frontier/2: ClickHouse 26.6: 10주년 기념 릴리스의 운영자 관점 체크리스트 — published 2026-07-14 (dynamic fallback: database topic; ClickHouse 26.6 released July 2026)
• [x] ai-frontier/5: llm-d: Kubernetes가 LLM 추론 요청을 KV 캐시 기준으로 어디로 보낼지 결정하는 방법 — published 2026-07-14 (dynamic fallback: AI systems topic; llm-d CNCF Sandbox March 2026, v0.6.0 April 2026)
• [x] database-frontier/3: PostgreSQL 19 Beta: 운영자가 GA 전에 미리 알아야 할 변화 — published 2026-07-14 (dynamic fallback: database topic; PostgreSQL 19 Beta 1 released 2026-06-04)
• [x] ai-frontier/6: MCP(Model Context Protocol): LLM이 도구를 연결하는 방식의 표준화 — published 2026-07-14 (dynamic fallback: AI engineering topic; MCP stable spec 2025-11-25, AAIF governance Dec 2025)
• [x] database-frontier/4: Valkey 9: Redis 포크가 SIMD와 I/O 스레딩으로 독자적 진화를 시작한 방식 — published 2026-07-14 (dynamic fallback: database topic; Valkey 9.1 released 2026-05-19, within 90 days)
• [x] ai-frontier/7: Transformers 5.13 HfExporters: ONNX·ExecuTorch·torch.export를 하나의 파이프라인으로 묶기 — published 2026-07-15 (dynamic fallback: AI engineering topic; Transformers v5.13.0 released 2026-07-03)
• [x] database-frontier/5: TimescaleDB 2.28: 연속 집계를 작게 갱신하고 컬럼스토어 메타데이터로 답하는 방식 — published 2026-07-15 (dynamic fallback: database/data-platform topic; TimescaleDB 2.28.0 released 2026-06-16, latest patch 2.28.2 released 2026-06-30)
• [x] ai-frontier/8: MLflow 3.14: 에이전트 트레이스를 WAL로 지키고 pytest로 품질을 게이트하는 방법 — published 2026-07-15 (dynamic fallback: AI engineering topic; MLflow 3.14.0 released 2026-06-17)
• [x] ai-frontier/9: LangGraph 1.2: 장기 실행 에이전트를 위한 DeltaChannel·TimeoutPolicy·스트리밍 v2 — published 2026-07-15 (dynamic fallback: AI engineering topic; LangGraph 1.2.0 released 2026-05-12, latest patch 1.2.9 released 2026-07-10)
• [x] database-frontier/6: Apache Iceberg 1.11: 삭제 벡터·REST 카탈로그 서버 사이드 플래닝·Variant 타입으로 완성된 V3 — published 2026-07-15 (dynamic fallback: database/data-platform topic; Apache Iceberg 1.11.0 released 2026-05-19)
• [x] ai-frontier/10: Microsoft Agent Framework 1.0: AutoGen과 Semantic Kernel이 하나의 프로덕션 SDK로 수렴하는 방법 — published 2026-07-15 (dynamic fallback: AI engineering topic; MAF 1.0 GA released 2026-04-03, AutoGen maintenance mode 2025-10)
• [x] database-frontier/7: DuckDB 1.5 Variegata: VARIANT 타입과 PEG 파서로 분석 엔진의 경계를 넓히는 방법 — published 2026-07-15 (dynamic fallback: database/data-platform topic; DuckDB 1.5.0 released 2026-03-09, latest 1.5.4 released 2026-06-17)
• [x] ai-frontier/11: DeepSeek V4: mHC와 이중 압축 어텐션으로 100만 토큰 컨텍스트를 여는 방법 — published 2026-07-15 (dynamic fallback: AI engineering topic; DeepSeek V4 released 2026-04-24, within 90 days)
• [x] database-frontier/8: MySQL 9.7 LTS: Hypergraph 옵티마이저와 Enterprise 기능의 Community 개방 — published 2026-07-15 (dynamic fallback: database topic; MySQL 9.7.0 LTS released 2026-04-21, within 90 days)
• [x] ai-frontier/12: Ray 2.56 Serve LLM: 라우팅 결정과 토큰 스트림을 분리한 direct streaming — published 2026-07-16 (dynamic fallback: AI systems topic; Ray 2.56.0 released 2026-06-29)
• [x] database-frontier/9: QuestDB 9.4: posting/covering index와 Parquet sidecar로 읽기 경도 줄이기 — published 2026-07-16 (dynamic fallback: database/data-platform topic; QuestDB 9.4.0 released 2026-05-18, latest 9.4.3 patch released 2026-06-15)
• [x] database-frontier/10: Qdrant 1.18: TurboQuant와 메모리 가드레일로 벡터 검색의 용량 경계 다루기 — published 2026-07-16 (dynamic fallback: AI/database intersection topic; Qdrant 1.18.0 released 2026-05-11, latest 1.18.2 patch released 2026-06-04)
• [x] ai-frontier/13: Claude Adaptive Thinking: 토큰 예산을 버리고 복잡도를 스스로 판단하는 추론 API — published 2026-07-16 (dynamic fallback: AI engineering topic; adaptive thinking GA with Sonnet 4.6 Feb 2026, Opus 4.8 only mode; budget_tokens deprecated on Opus 4.6 / Sonnet 4.6)
• [x] database-frontier/11: Apache Kafka 4.3: 브로커 코돈·Share Group 성숙·클래식 프로토콜 퇴장 준비 — published 2026-07-16 (dynamic fallback: data-platform topic; Kafka 4.3.0 released 2026-05-22, 4.3.1 released 2026-06-25)
• [x] ai-frontier/14: Google ADK 2.0: LLM 에이전트를 그래프 노드로 바꾼 Workflow Runtime — published 2026-07-16 (dynamic fallback: AI engineering topic; ADK 2.0.0 stable released 2026-05-19, current 2.4.0 released 2026-07-07)
• [x] database-frontier/12: ScyllaDB 2026.2: Alternator Streams GA·Vector Search·Trie 인덱스·Raft 데이터 일관성 — published 2026-07-16 (dynamic fallback: database topic; ScyllaDB 2026.2 released 2026-06-29)
• [x] ai-frontier/15: Gemma 4 MTP: 드래프터가 초안을 쓰고 타깃이 검증하는 추론 가속 구조 — published 2026-07-16 (dynamic fallback: AI engineering topic; Gemma 4 MTP drafters released 2026-05-05)
• [x] database-frontier/13: Apache Flink 2.3.0: FROM_CHANGELOG·TO_CHANGELOG로 스트림과 테이블 경계를 SQL로 제어하기 — published 2026-07-16 (dynamic fallback: data-platform topic; Apache Flink 2.3.0 released 2026-06-25)
• [x] ai-frontier/16: KServe 0.20 RC0: LLMInferenceService가 모델별 라우팅을 안전하게 켜고 전역 설정을 캐시하는 법 — published 2026-07-17 (dynamic fallback: AI systems topic; KServe v0.20.0-rc0 released 2026-07-16)
• [x] database-frontier/14: Trino 482: char/varchar 의미론 수정과 Lakehouse split 제어를 한 번에 가져온 릴리스 — published 2026-07-17 (dynamic fallback: query-engine/data-platform topic; Trino 482 released 2026-06-25)
• [x] database-frontier/15: OpenSearch 3.7: 매핑 업데이트 없는 dynamic_properties, Parquet 인덱싱, WritableWarm tiering — published 2026-07-17 (dynamic fallback: search/storage-engine topic; OpenSearch 3.7.0 released 2026-06-09, within 90 days)
• [x] ai-frontier/17: Kimi K3: 2.8조 파라미터 오픈 MoE 모델이 선형 어텐션으로 100만 토큰 추론을 가속하는 방법 — published 2026-07-17 (dynamic fallback: AI engineering topic; Kimi K3 released 2026-07-16)
• [x] database-frontier/16: Apache Spark 4.1: VARIANT 쉬레딩 GA, Real-Time Mode, SQL 스크립팅으로 스트리밍과 분석을 다시 쓰는 방법 — published 2026-07-17 (dynamic fallback: data-platform topic; Spark 4.1.x series active with patches through 2026-05/06)
• [x] ai-frontier/18: Meta Llama 4 Scout & Maverick: iRoPE·조기 융합 멀티모달·MoE로 오픈 AI의 다음 기준을 정한 방식 — published 2026-07-17 (dynamic fallback: AI engineering topic; Llama 4 Scout and Maverick released 2026-04-05, within 90 days)
• [x] database-frontier/17: Elastic 9.4: GPU로 벡터 인덱싱을 12배 빠르게 하고 Prometheus TSDB를 대체하며 FIPS 140-3을 완성한 릴리스 — published 2026-07-17 (dynamic fallback: database/search-engine topic; Elastic 9.4.0 released 2026-04-30, within 90 days)
• [x] ai-frontier/19: OpenAI Agents SDK Sandbox & Harness: 에이전트를 신뢰 경계 기준으로 제어면과 실행면으로 분리한 방법 — published 2026-07-17 (dynamic fallback: AI engineering topic; OpenAI Agents SDK sandbox/harness update released 2026-04-15)
• [x] database-frontier/18: MongoDB 8.3: $scoreFusion GA와 self-managed 검색 완성 — 하이브리드 검색을 단일 데이터베이스에서 운영하는 방법 — published 2026-07-17 (dynamic fallback: database topic; MongoDB 8.3.0 released 2026-05-07)

• [x] ai-frontier/20: PydanticAI 2.12: deferred tool과 durable execution으로 장기 실행 에이전트를 운영하는 방법 — published 2026-07-18 (dynamic fallback: AI engineering topic; PydanticAI v2.12.0 released 2026-07-17, with durable-execution/deferred-tool event surface changes)
• [x] database-frontier/19: Apache Hudi 1.2: VECTOR·VARIANT·BLOB를 Lakehouse에 넣되 운영 경계를 먼저 정해야 하는 이유 — published 2026-07-18 (dynamic fallback: lakehouse/data-platform topic; Apache Hudi 1.2.0 released 2026-05-23, within 90 days)
• [x] ai-frontier/21: Transformers 5.14: MTP·ensemble verification·StaticCache로 생성 경로를 다시 맞춘 릴리스 — published 2026-07-18 (dynamic fallback: AI systems topic; Transformers v5.14.0 released 2026-07-15, within 90 days)
• [x] ai-frontier/22: DSpark: DeepSeek가 투기적 디코딩의 두 가지 약점을 동시에 해결한 방법 — published 2026-07-18 (dynamic fallback: AI engineering topic; DSpark released 2026-06-27, arXiv:2607.05147 published 2026-07-06, within 90 days)
• [x] database-frontier/20: Apache Druid 37: 세그먼트 메타데이터 캐시 기본화와 Hadoop 수집 종료로 달라진 운영 경계 — published 2026-07-18 (dynamic fallback: database/data-platform topic; Apache Druid 37.0.0 released 2026-05-08, within 90 days)

• [x] ai-frontier/23: Prefill-Decode 분리 추론: LLM 서빙에서 TTFT와 처리량을 독립적으로 제어하는 방법 — published 2026-07-18 (dynamic fallback: AI systems topic; vLLM disaggregated prefill/decode became production standard in 2026; recent papers arXiv:2507.06608 Jul 2026, arXiv:2607.02043 Jul 2026)
• [x] database-frontier/21: Apache Polaris 1.x: Iceberg REST 카탈로그 표준이 멀티엔진 Lakehouse 거버넌스 레이어가 되는 방법 — published 2026-07-18 (dynamic fallback: data-platform topic; Apache Polaris graduated to ASF TLP 2026-02-18, v1.5.0 released 2026-05)

• [x] ai-frontier/24: NVIDIA Dynamo 1.0: 데이터센터 스케일 LLM 추론 오케스트레이션 — SLO Planner·KV Block Manager·NIXL의 동작 방식 — published 2026-07-18 (dynamic fallback: AI systems topic; NVIDIA Dynamo 1.0 GA released 2026-03-16 at GTC; disaggregated inference orchestration framework with NIXL, KV Block Manager multi-tier caching, SLO Planner)
• [x] ai-frontier/25: SGLang 0.5.15: Spec V2와 Breakable CUDA Graph가 기본 경로가 되면서 달라진 운영 기준 — published 2026-07-19 (dynamic fallback: AI systems topic; SGLang v0.5.15 released 2026-07-10, with Spec V2 default path, Breakable CUDA Graph default changes, IndexShare MTP, and TopK V2 improvements)
• [x] ai-frontier/28: smolagents 1.25: 코드를 생성해 도구를 부르는 에이전트의 실행 경계와 보안 설계 — published 2026-07-19 (dynamic fallback: AI engineering topic; smolagents v1.25.0 released 2026-05-14, v1.26.0 released 2026-05-29; code-first CodeAgent paradigm, v1.25 security hardening of remote executors)
• [x] database-frontier/25: StarRocks 4.1: shared-data 클러스터의 tablet 자동 분할과 운영자가 확인해야 할 것 — published 2026-07-19 (dynamic fallback: database/data-platform topic; StarRocks 4.1 released ~April 2026, within 90 days; tablet auto-splitting, large-capacity tablets ~30GB, Fast Schema Evolution v2, cache observability)
• [x] ai-frontier/29: Qwen3.6: Gated DeltaNet 하이브리드 어텐션으로 27B가 400B를 이기는 방법 — published 2026-07-19 (dynamic fallback: AI engineering topic; Qwen3.6-27B released 2026-04-22, within 90 days; hybrid Gated DeltaNet + Gated Attention, 64 layers in 16 blocks, YaRN 262K→1M context, MTP speculative decoding, Apache 2.0)
• [x] database-frontier/26: DuckLake 1.0: JSON 파일을 버리고 SQL로 Lakehouse 카탈로그를 만든 이유 — published 2026-07-19 (dynamic fallback: lakehouse/data-platform topic; DuckLake 1.0 released 2026-04-13, 96 days ago, outside 90-day window but within 180-day fallback — no equivalent architectural departure available within 90 days; SQL-catalog lakehouse format, data inlining, deletion vectors, ACID DDL transactions)
• [x] database-frontier/22: Redis 8.8: Array 자료구조·INCREX·XNACK — 운영자가 알아야 할 기능 확장과 성능 변화 — published 2026-07-18 (dynamic fallback: database topic; Redis 8.8.0 released 2026-06; native Array data structure, INCREX window counter rate limiter, XNACK streams NACK, 60% faster full sync replication)
• [x] database-frontier/23: dbt Core 1.12: v2 parser·native private packages·기본값 전환을 운영에 반영하는 법 — published 2026-07-19 (dynamic fallback: data-platform topic; dbt Core v1.12.0 released 2026-07-16, within 90 days)
• [x] ai-frontier/26: Ollama 0.32: 로컬 모델 런처가 승인형 에이전트 셸로 바뀔 때 운영자가 확인할 것 — published 2026-07-19 (dynamic fallback: AI engineering topic; Ollama v0.32.0 released 2026-07-14 and v0.32.1 released 2026-07-17, introducing the root agent launcher and immediate guardrail follow-up fixes)

• [x] ai-frontier/27: CrewAI 1.14: 에이전트 메모리·지식·흐름을 교체 가능한 컴포넌트로 분리한 방법 — published 2026-07-19 (dynamic fallback: AI engineering topic; CrewAI v1.14.7 released 2026-06-11, pluggable memory/knowledge/RAG/flow backends, runtime state scoping, Chat API, FlowDefinition DSL split)
• [x] database-frontier/24: Delta Lake 4.3 Catalog Commits: 카탈로그가 커밋을 중개하면서 달라진 멀티엔진 쓰기의 경계 — published 2026-07-19 (dynamic fallback: database/data-platform topic; Delta Lake 4.3 released 2026-06-22, Catalog Commits GA with Unity Catalog, coordinated commits, multi-table transactions, multi-engine write support)
• [x] ai-frontier/30: Phoenix 19: OAuth2 Authorization Server와 Remote MCP로 에이전트 접근 경계를 다시 그은 릴리스 — published 2026-07-20 (dynamic fallback: AI engineering topic; Arize Phoenix 19.0.0 released 2026-07-17, introducing a built-in OAuth2 authorization server, browser-based CLI login, Remote MCP, and non-transitive API key issuance)
• [x] database-frontier/27: RisingWave 3.0: SQL 운영면과 DataFusion 기본화로 스트리밍 DB의 경계를 넓힌 방식 — published 2026-07-20 (dynamic fallback: data-platform topic; RisingWave v3.0.0 released 2026-06-11, official announcement published 2026-07-10, within 90 days)
• [x] database-frontier/28: Trino 483: PIVOT·JSON 접근자·Lakehouse S3 인증 정리 — 운영자가 먼저 볼 변화 — published 2026-07-20 (dynamic fallback: query-engine/lakehouse topic; Trino 483 released 2026-07-17, within 90 days; first-class PIVOT and JSON simplified accessor, OAuth2 domain-hint, new default Web UI, unified s3.auth-type migration across lakehouse connectors, and SPI validity bitmap change)

• [x] ai-frontier/31: Anthropic Advanced Tool Use: 프로그래매틱 도구 호출과 도구 검색으로 에이전트 컨텍스트 비용을 줄이는 방법 — published 2026-07-20 (dynamic fallback: AI engineering topic; Tool Search GA November 2025, Programmatic Tool Calling beta code_execution_20260120 January 2026, advanced tool use blog June 2026; 24% token reduction on BrowseComp/DeepSearchQA, 38% on 75-tool agent benchmark)
• [x] database-frontier/29: Apache Pinot 1.5: 멀티클러스터 페더레이션과 Multi-Stage 쿼리 엔진으로 리얼타임 OLAP의 운영 경계를 다시 그은 릴리스 — published 2026-07-20 (dynamic fallback: database/data-platform topic; Apache Pinot 1.5.0 released 2026-04-15, within 90 days; federation multi-cluster routing, MSE UNNEST and join improvements, offline upsert support, Kafka 4.x KRaft module, commit-time compaction, query resource isolation)
• [x] ai-frontier/32: NVIDIA Cosmos 3: 하나의 모델로 추론·생성·행동을 통합한 Physical AI 월드 파운데이션 모델 — published 2026-07-20 (dynamic fallback: AI systems topic; NVIDIA Cosmos 3 launched 2026-06-01 at Snowflake Summit; Mixture-of-Transformers with AR/DM shared self-attention, Edge 4B/Nano 16B/Super 64B model scales, open-weight NVIDIA Open Model License)
• [x] database-frontier/30: dbt Core v2.0: Fusion 엔진이 기본이 되면서 달라진 프로젝트 구조와 마이그레이션 기준 — published 2026-07-20 (dynamic fallback: data-platform topic; dbt Core v2.0 alpha announced 2026-06-01 at Snowflake Summit, a4 released 2026-07-06; Rust Fusion engine, 30× parse speed, Parquet artifacts, dbt State, dbt lint, dbt Docs v2, two-distribution model OSS vs proprietary)

• [x] ai-frontier/33: vLLM v0.25: MRv2 기본화·Rust 프론트엔드 성숙·PD 분리 2단계 — 운영자가 확인할 변화들 — published 2026-07-20 (dynamic fallback: AI systems topic; vLLM v0.25.0 released 2026-07-14; MRv2 default for all dense models, PagedAttention deleted, Rust frontend HTTPS/mTLS+DP supervisor, PD disaggregation secondary tier + Mooncake MLA/Mamba1 NIXL)
• [x] database-frontier/31: Neon Serverless PostgreSQL: 스토리지-컴퓨트 분리와 CoW 브랜칭이 만드는 새로운 OLTP 운영 모델 — published 2026-07-20 (dynamic fallback: database topic; Neon acquired by Databricks May 2025, ~$1B; storage-compute separation, Pageserver+Safekeeper architecture, O(1) CoW branching via LSN, scale-to-zero 500ms cold start, 80% AI-agent-created DBs)
• [x] ai-frontier/34: AG2 1.0 Beta: AutoGen classic 이후 에이전트 런타임을 ACP·Tool Search·NLIP로 재구성하는 법 — published 2026-07-21 (dynamic fallback: AI engineering topic; AG2 v1.0.0b0 released 2026-07-03, within 90 days; classic surface removal, ACP CLI-agent bridge, Tool Search deferred loading, NLIP remote-agent boundary)
• [x] database-frontier/32: HBase 3.0.0-beta-2: ZooKeeper를 내부로 밀어내고 JDK 17·동기 복제를 전제로 다시 그은 업그레이드 — published 2026-07-21 (dynamic fallback: database topic; Apache HBase 3.0.0-beta-2 released 2026-07-17, within 90 days; ZooKeeper-visible client/control-plane contraction, hbase:replication queue storage, JDK 17 / Hadoop 3.3.5+ baseline, synchronous replication runbook, backup/restore and OpenTelemetry operations)
• [x] ai-frontier/35: Haystack 3.0: 훅·스킬·비동기 파이프라인으로 프로덕션 에이전트 프레임워크를 다시 짠 방식 — published 2026-07-21 (dynamic fallback: AI engineering topic; Haystack v3.0.0 released 2026-07-20, within 90 days; Agent-owned tool loop, progressive skill disclosure, async Pipeline unification, deserialization allowlist)
• [x] ai-frontier/36: LlamaIndex Workflows 1.0: 이벤트 주도 에이전트 워크플로우의 독립화와 llama-deploy 프로덕션 배포 아키텍처 — published 2026-07-21 (dynamic fallback: AI engineering topic; LlamaIndex Workflows v1.0 released 2026-06-22, within 90 days; standalone llama-index-workflows package, typed state, resource injection, llama-deploy Control Plane + MQ distributed runtime)
• [x] database-frontier/33: CockroachDB 26.x: Leader Lease·값 분리(Value Separation)·FIPS 140-3으로 분산 SQL의 운영 경계를 다시 그은 방식 — published 2026-07-21 (dynamic fallback: database/distributed-SQL topic; CockroachDB v26.2 released 2026-05-22 within 90 days, v26.1 released 2026-02-18 within 180 days; Leader Lease Raft+leaseholder unification SIGMOD 2026, Value Separation Pebble blob files 50% WA reduction, FIPS 140-3 native Go 1.24, JWT/OIDC SSO role sync, CDC Protobuf, MAINTAIN privilege)
• [x] ai-frontier/37: GPT-5.6 Sol·Terra·Luna: 세대와 능력 계층을 분리한 모델 패밀리와 Programmatic Tool Calling — published 2026-07-21 (dynamic fallback: AI engineering/systems topic; GPT-5.6 released 2026-07-09; three-tier Sol/Terra/Luna family distilled from same base run, 1.05M context window, 128K output, Feb 2026 cutoff; Programmatic Tool Calling GA in Responses API — JavaScript-driven hosted tool execution runtime)
• [x] database-frontier/34: Apache Doris 4.1: 벡터·전문 검색·분석을 하나의 SQL 엔진으로 통합한 HSAP와 IVF 인덱스 확장 — published 2026-07-21 (dynamic fallback: database/data-platform topic; Apache Doris 4.1.0 released 2026-04-21, 4.1.1 patch 2026-05-24 within 90 days; IVF and IVF_ON_DISK vector indexes for billion/trillion scale, Segment V3 16× faster open / 60× lower memory, DOC mode deferred JSON shredding, Iceberg V3 MERGE INTO, Paimon DDL, ASOF JOIN, Recursive CTE)
• [x] ai-frontier/38: AI SDK 7: WorkflowAgent·Tool Approval·MCP Apps로 에이전트 런타임을 재구성한 방식 — published 2026-07-22 (dynamic fallback: AI engineering topic; AI SDK 7.0.0 published 2026-06-25 within 90 days; WorkflowAgent durable runtime, runtime/tool context split, MCP Apps model-visible/app-visible separation, file/skill provider references, explicit approval/timeout/sandbox surfaces)
• [x] database-frontier/35: YugabyteDB 2026.1: 분산 추적·QPM·Flink CDC·즉시 클론으로 운영 제어면을 넓힌 STS 릴리스 — published 2026-07-22 (dynamic fallback: database/distributed-SQL topic; YugabyteDB v2026.1.0.0 released 2026-06-29 within 90 days; YSQL Distributed Tracing TP, QPM default-on GA, CDC table-rewrite streaming + implicit publication, Flink CDC tech preview, and instant database cloning GA)
• [x] ai-frontier/39: TensorRT-LLM 1.3 RC: 레거시 TensorRT 백엔드 제거와 PyTorch·분리 서빙 전환의 운영 기준 — published 2026-07-22 (dynamic fallback: AI systems topic; TensorRT-LLM v1.3.0rc21 released 2026-07-15 within 90 days; legacy TensorRT backend removal in Python package, native /v1/embeddings batching, AutoDeploy deprecation, and disaggregated serving adoption boundaries)
• [x] database-frontier/36: Materialize v26.33: PostgreSQL 메타데이터 DB를 READ COMMITTED로 바꾸고 혼잡한 클러스터 경합을 줄이는 방식 — published 2026-07-23 (dynamic fallback: database/streaming-SQL topic; Materialize v26.33.0 released to Cloud 2026-07-16 and Self-Managed 2026-07-17, within 90 days; PostgreSQL metadata DB READ COMMITTED consensus queries, session-scoped catalog snapshot cache, timestamp-oracle stall isolation, and replica-targeted EXPLAIN ANALYZE via MCP developer endpoint)
• [x] ai-frontier/40: Google ADK 2.5: 원격 MCP·에이전트→MCP·Cloud Run sandbox로 실행 경계를 다시 잇는 방식 — published 2026-07-23 (dynamic fallback: AI engineering topic; Google ADK v2.5.0 released 2026-07-16, within 90 days; to_mcp_server, server-side remote MCP for ManagedAgent, and CloudRunSandboxCodeExecutor make protocol, auth, and code-execution boundaries explicit)
• [x] ai-frontier/41: Gemini 3.6 Flash·3.5 Flash-Lite: 에이전트 루프에서 토큰 비용을 다시 계산하게 만드는 두 모델 — published 2026-07-23 (dynamic fallback: AI engineering topic; Gemini 3.6 Flash + 3.5 Flash-Lite + 3.5 Flash Cyber released 2026-07-21, within 90 days; 17% fewer output tokens on AI Analysis Index, 65% reduction on DeepSWE, configurable thinking levels for Flash-Lite, 350 tok/s throughput)
• [x] database-frontier/37: Weaviate 1.38: HFresh 디스크 벡터 인덱스·MCP 서버·비동기 복제 재구성으로 달라진 벡터 DB 운영 기준 — published 2026-07-23 (dynamic fallback: database/vector-search topic; Weaviate v1.38.0 released ~June 2026, within 90 days; HFresh disk-based vector index GA (SPFresh-inspired, centroid HNSW + posting LSM), MCP Server GA, async replication single-scheduler rebuild enabled by default)
• [x] ai-frontier/42: Claude Sonnet 5: 새 토크나이저·적응형 사고 기본화·샘플링 제약이 기존 코드에 미치는 영향 — published 2026-07-23 (dynamic fallback: AI engineering topic; Claude Sonnet 5 released 2026-06-30, within 90 days; new tokenizer ~30% more tokens, adaptive thinking on by default, manual extended thinking removed → 400, sampling params to non-default → 400, cybersecurity safeguards, Priority Tier not supported)
• [x] database-frontier/38: Apache Cassandra 5.0.8: CEP-37 Auto Repair가 내장되면서 외부 수리 도구 없이 클러스터를 유지하는 방법 — published 2026-07-23 (dynamic fallback: database topic; Cassandra 5.0.8 released April 2026, CEP-37 Auto Repair backported from 6.0; built-in scheduler with auto_repair_history, RepairTokenRangeSplitter/FixedSplitTokenRangeSplitter, Full/Incremental/Preview repair types, JMX control, table-level priority)
• [x] ai-frontier/43: Claude Fable 5: Mythos급 모델의 API 변화와 Opus 4.8 마이그레이션 기준 — published 2026-07-23 (dynamic fallback: AI engineering topic; Claude Fable 5 released 2026-06-09, within 90 days; Mythos-class above Opus 4.8, 1M context, adaptive thinking always on, safety classifiers refusal handling, 30-day data retention required, effort parameter, $10/$50 pricing)
• [x] database-frontier/39: PostgreSQL 19 Beta 2: SQL/PGQ 그래프 쿼리·내장 REPACK·JIT 기본 비활성으로 확정된 GA 변화 — published 2026-07-23 (dynamic fallback: database topic; PostgreSQL 19 Beta 2 released 2026-07-16, within 90 days; feature freeze confirmed; SQL/PGQ GRAPH_TABLE/CREATE PROPERTY GRAPH, REPACK CONCURRENTLY built-in pg_repack replacement, JIT default off, lz4 TOAST default, parallel autovacuum; GA September/October 2026)
• [x] ai-frontier/44: Mastra 1.51: Durable Agent Crash Recovery로 끊긴 RUNNING 실행을 다시 붙이는 법 — published 2026-07-24 (dynamic fallback: AI engineering topic; Mastra 1.51.0 released 2026-07-15, within 90 days; durable agent crash recovery adds listActiveRuns/recover/recoverActiveRuns, boot-time auto recovery, and HTTP/client recovery surfaces with explicit idempotency and multi-replica caveats)
• [x] database-frontier/40: Apache Pulsar 5.0.0-M1: Scalable Topics·Oxia·V5 Client로 스트리밍 메시징의 운영 경계를 다시 그은 프리뷰 — published 2026-07-24 (dynamic fallback: data-platform topic; Apache Pulsar 5.0.0-M1 released 2026-06-23, within 90 days; preview-only release introducing scalable topics with hash-range segments and push-based lookup sessions, additive pulsar-client-v5 consumer split, Oxia as recommended metadata store with DualMetadataStore migration, structured slog logging, and Gradle / jakarta / connector-split migration boundaries)
• [x] ai-frontier/45: LiteLLM 1.93~1.94: MCP OBO·Responses bridge·Rust Gateway로 LLM 게이트웨이를 제어면으로 바꾸는 방법 — published 2026-07-24 (dynamic fallback: AI engineering topic; LiteLLM v1.93.0 released 2026-07-19, with July 2026 MCP OAuth/OBO and Rust gateway docs; explicit oauth2_flow + OBO token exchange + /responses bridge hardening + opt-in Rust hot path for Anthropic /v1/messages)
• [x] ai-frontier/46: MCP 2026-07-28 스펙 RC: 상태 없는 프로토콜 코어·확장 프레임워크·인증 강화로 에이전트 연결 경계를 다시 그은 방식 — published 2026-07-24 (dynamic fallback: AI engineering topic; MCP 2026-07-28 spec RC locked 2026-05-21, final spec ships 2026-07-28, within 90 days; stateless core removes sessions/initialize handshake, InputRequiredResult replaces SSE, Tasks/MCP Apps as extensions, RFC 9207 iss validation, Roots/Sampling/Logging deprecation, JSON Schema 2020-12)
• [x] database-frontier/41: Apache Gravitino 1.3.0: AWS Glue·논리 뷰 통합 관리·계층 네임스페이스로 멀티엔진 Lakehouse 거버넌스 제어면을 확장한 방식 — published 2026-07-24 (dynamic fallback: data-platform/lakehouse topic; Apache Gravitino 1.3.0 released 2026-06-29, within 90 days; AWS Glue Catalog + Trino/Spark adapters, unified logical view management across Hive/Iceberg/Paimon, multi-level hierarchical namespaces, built-in IDP, federated Iceberg REST Catalog, vended credential auto-refresh; breaking: sensitive catalog properties hidden from load response, Docker path /root/gravitino → /opt/gravitino)
• [x] ai-frontier/47: OpenAI Presence: 정책·가드레일·Codex 개선 루프로 기업 AI 에이전트를 생산에 올리는 방법 — published 2026-07-24 (dynamic fallback: AI engineering topic; OpenAI Presence announced 2026-07-22, OpenAI Deployment Company launched May 2026 with $4B+ backing and Tomoro acquisition; enterprise voice/chat agent platform with policy layers, knowledge isolation, guardrails, simulation/eval, and Codex improvement loop; not self-serve, FDE-led deployment; 75% auto-resolution rate, 15% handoff reduction in 10 days)
• [x] database-frontier/42: Databricks Unity Catalog Managed Iceberg GA: 테이블 포맷 잠금을 해제하고 크로스엔진 쓰기까지 열린 Lakehouse 거버넌스 — published 2026-07-24 (dynamic fallback: data-platform/lakehouse topic; announced at Data + AI Summit June 15-18 2026, within 90 days; Managed Iceberg GA, Iceberg v3 GA (deletion vectors, row tracking, VARIANT), Foreign Iceberg GA with credential vending, UniForm Delta+Iceberg parallel metadata; Predictive Optimization + Liquid Clustering for automatic maintenance; cross-engine read/write via Iceberg REST Catalog API)
• [x] ai-frontier/48: Sleep-time Compute: 에이전트가 쉬는 동안 기억을 정제하고 추론 비용을 5배 줄이는 방법 — published 2026-07-24 (dynamic fallback: AI engineering topic; arXiv:2504.13171 by Letta + UC Berkeley, published April 2026, within 90 days; sleep-time compute two-agent architecture (primary + sleep-time agent), shared memory blocks, rethink_memory() calls; 5× test-time token reduction on Stateful GSM-Symbolic and Stateful AIME, +13%/+18% accuracy improvement; Letta production implementation)
• [x] database-frontier/43: ClickHouse 26.7: 역대 최대 성능 최적화 릴리스 — 운영자가 먼저 확인할 호환성 경계와 변화 — published 2026-07-24 (dynamic fallback: database/analytics topic; ClickHouse v26.7.1.1315-stable released 2026-07-22, within 90 days; record 64 new features, 117 performance optimizations, 329 bug fixes; backward-compat breaks: S3 credential auto-resolution removed, zip/zipx backup rejected on object storage, workload scheduling XML sections removed, Snowflake functions removed; EXPLAIN ANALYZE, JOIN hash table 15% memory / 12% speed, Delta codec 1.5–5× decompression, libdeflate 1.15–1.5×, per-CPU atomics ~30× faster profiling, QBit vector quantization experimental)
• [x] ai-frontier/49: Agno 2.8: scorer·environments·learning zone으로 에이전트 평가를 배포 게이트로 바꾸는 방법 — published 2026-07-25 (dynamic fallback: AI engineering topic; Agno v2.8.0 released 2026-07-20, introducing agno.scorer, isolated agno.environments, execution-matching reliability checks, learning-zone curation, and CI gating examples)
• [x] database-frontier/44: Apache DataFusion Ballista 54: AQE·브로드캐스트 조인·REST 관측성으로 분산 쿼리 엔진을 운영 가능한 수준으로 끌어올린 릴리스 — published 2026-07-25 (dynamic fallback: data-platform topic; Ballista 54.0.0 released 2026-07-12, within 90 days; AQE runtime broadcast join path, static broadcast join defaults even with AQE off, REST/TUI/metrics control plane, and mixed-version upgrade boundary)
• [x] ai-frontier/50: Temporal Python SDK 1.28~1.30: LangGraph·OpenAI Agents·Strands를 durable workflow에 붙이는 방식 — published 2026-07-25 (dynamic fallback: AI engineering topic; Temporal Python SDK 1.28.0/1.29.0/1.30.0 released 2026-06-04, 2026-06-17, 2026-07-02, within 90 days; agent plugin surfaces for Strands, LangGraph, OpenAI Agents, plus Workflow Streams and Nexus signal link propagation)
• [x] ai-frontier/51: DSPy 3.2: 프롬프트 문자열을 버리고 타입 선언으로 LLM을 다루는 방법 — published 2026-07-25 (dynamic fallback: AI engineering topic; DSPy 3.2.1 released 2026-05-05, within 90 days; Signatures/Modules/Optimizers architecture, BetterTogether p→w→p chaining strategy, LiteLLM decoupling, ReActV2 preview in 3.3.0b1)
• [x] database-frontier/45: Apache Paimon 1.3: 스트리밍과 배치를 LSM 테이블 하나로 통합하는 레이크하우스 포맷 — published 2026-07-25 (dynamic fallback: data-platform/lakehouse topic; Apache Paimon 1.3.x latest stable as of May 2026; LSM-based primary key table with Merge Engine/Table Mode/Changelog Producer, Append Table auto small-file merge, Flink native integration and Lookup Join, Deletion Vector MOW mode)
• [x] ai-frontier/52: W&B Weave: 프로덕션 에이전트를 위한 세션·턴·스텝 관측성과 온라인 평가 루프 — published 2026-07-25 (dynamic fallback: AI engineering/observability topic; W&B Weave ground-up rebuild for production agents June 2, 2026; Session/Turn/Step/Tool/Sub-agent hierarchy, OpenTelemetry GenAI semantic conventions, online signals toxicity/PII/hallucination/frustrated-users, CoreWeave ARIA agent integration)
• [x] database-frontier/46: SurrealDB 3.0~3.2: 문서·그래프·벡터를 하나의 ACID 쿼리로 묶는 AI 에이전트 메모리 레이어 — published 2026-07-25 (dynamic fallback: database/multi-model topic; SurrealDB 3.2.1 released 2026-07-10, within 90 days; multi-model relational+document+vector+graph+agent-memory, COMPUTED fields replacing <future>, on-disk redesign, ID-based metadata, synchronized writes default, client-side ACID transactions, HNSW/MTREE vector indexes, kNN operator)
• [x] ai-frontier/53: MiniMax M3: MSA 희소 어텐션으로 1M 토큰 컨텍스트와 네이티브 멀티모달을 동시에 여는 방법 — published 2026-07-25 (dynamic fallback: AI engineering topic; MiniMax M3 released 2026-06-01, within 90 days; 428B MoE open-weight model, MSA sparse attention 28.4× attention compute reduction at 1M context, 14.2× prefill speedup on H800, native multimodal training from step 0, arXiv:2606.13392)
• [x] database-frontier/47: MariaDB 12.3 InnoDB 기반 바이너리 로그: 10년 된 2PC를 없애고 쓰기 커밋 경로를 다시 그은 방식 — published 2026-07-25 (dynamic fallback: database topic; MariaDB 12.3.2 LTS released 2026-05-28, within 90 days; InnoDB-based binary log storing binlog events inside InnoDB tablespaces, 2PC elimination between binlog and redo, single-fsync commit path, crash-safe binlog without sync_binlog)
• [x] ai-frontier/54: LiveKit Agents 1.6: 비동기 도구와 Turn Detector v1으로 실시간 음성 에이전트의 응답 경계를 다시 짜는 법 — published 2026-07-26 (dynamic fallback: AI engineering topic; LiveKit Agents 1.6.0 released 2026-06-11, Turn Detector v1 announced 2026-06-17 and released in 1.6.1 on 2026-06-19, latest 1.6.7 patch released 2026-07-25; async tools, cancellable tool runtime, audio-semantic turn detection, and job/session control-plane boundaries)
• [x] database-frontier/48: TiDB 8.5.7: CPU-aware hotspot scheduling과 partial index로 분산 OLTP의 숨은 낭비를 줄이는 방식 — published 2026-07-26 (dynamic fallback: database topic; TiDB 8.5.7 released 2026-07-09, within 90 days; CPU-aware hot Region scheduling for read hotspots, partial indexes, and per-user max_user_connections in the 8.5 LTS line)
• [x] database-frontier/49: OceanBase 4.6: 하이브리드 검색·벡터 인덱스·HTAP를 한 엔진에 몰아넣은 분산 SQL 업그레이드 — published 2026-07-26 (dynamic fallback: database/data-platform topic; OceanBase V4.6.0_CE released 2026-07-17, within 90 days; native HYBRID_SEARCH SQL, Search Index + phrase-aware full-text retrieval, HNSW segment lifecycle + recall reporting, strong-consistent columnar replica reads, and OBCDC restore-aware incremental sync)
• [x] ai-frontier/55: Mistral Magistral: 증류 없이 GRPO+RLVR만으로 만든 유럽 최초 오픈 소스 추론 모델 — published 2026-07-26 (dynamic fallback: AI engineering topic; Magistral released June 2026, arXiv:2506.10910 accepted ICML 2026; Magistral Small 24B Apache 2.0, GRPO+RLVR no distillation, ~50% AIME-24 improvement, 20+ language CoT, async RL training system)
• [x] database-frontier/50: Apache Spark 4.2: NEAREST BY·Metric Views·Auto CDC로 분석 엔진에 AI 워크로드를 직접 끌어들인 릴리스 — published 2026-07-26 (dynamic fallback: data-platform topic; Apache Spark 4.2.0 GA released 2026-07-14, within 90 days; NEAREST BY top-K vector join, vector distance/similarity/normalization SQL functions, Metric Views governed semantic layer, SQL CHANGES/Auto CDC in Declarative Pipelines, Arrow-first Python default, GEOMETRY/GEOGRAPHY geospatial types)
• [x] ai-frontier/56: Claude Opus 5: Frontier-Bench를 다시 쓴 Opus — effort 사다리·mid-conversation 도구 변경·마이그레이션 체크리스트 — published 2026-07-26 (dynamic fallback: AI engineering topic; Claude Opus 5 released 2026-07-24, within 90 days; model ID claude-opus-5, $5/$25 per MTok same as Opus 4.8, effort ladder low/medium/high/xhigh/max, thinking on by default, thinking-disabled blocked at xhigh/max effort, mid-conversation tool changes beta, default fallbacks mode, 512-token cache minimum, Frontier-Bench 43.3% vs Fable 5 33.7%)
• [x] database-frontier/51: Oracle AI Database 26ai: AI 벡터 검색·MCP 서버·Iceberg를 하나의 SQL로 묶은 AI 네이티브 데이터베이스 — published 2026-07-26 (dynamic fallback: database topic; Oracle AI Database 26ai GA on-premises Jan 2026, AWS RDS support July 2026, within 90 days; renamed from Oracle Database 23ai; AI Vector Search HNSW/IVF, unified hybrid search, JSON Relational Duality Views, Select AI Agent with MCP, Iceberg Lakehouse integration, SQL Firewall, ML-KEM quantum-resistant encryption, RAFT-based globally distributed database)
• [x] ai-frontier/57: Amazon Bedrock AgentCore: 에이전트 세션을 서비스로 분리하고 Agents Classic 종료 전에 확인할 것 — published 2026-07-26 (dynamic fallback: AI engineering topic; Bedrock Agents Classic stops new customers July 30 2026; AgentCore GA with Runtime/Memory/Gateway/Identity/Observability; Gateway Guardrails GA June 2026; InvokeGuardrailChecks API June 2026; Web Search managed tool; A2A and MCP protocol support)
• [x] database-frontier/52: Percona Operator for MySQL 1.2.0: PerconaServerMySQLClusterSet·암호화 백업·자동 스토리지 확장으로 K8s MySQL 운영을 강화하는 방법 — published 2026-07-26 (dynamic fallback: database/K8s-operator topic; Percona Operator for MySQL 1.2.0 released July 3 2026, within 90 days; PerconaServerMySQLClusterSet CRD for InnoDB ClusterSet cross-site DR; XtraBackup AES256 encrypted backups with Vault validation; threshold-based PVC autoscaling; Orchestrator API authentication; PITR binlog in restore objects)
• [x] ai-frontier/58: vLLM v0.26.0: 그룹별 어텐션 백엔드·오브젝트 스토어 KV 계층·Rust 멀티모달로 하이브리드 모델 서빙을 다시 쓰는 방법 — published 2026-07-27 (dynamic fallback: AI systems topic; vLLM v0.26.0 released ~2026-07-25, within 90 days; per-KV-cache-group attention backend selection, tiered KV offloading with object-store secondary tier, TML Inkling 1T model Day-0 support, DeepSeek-V4 routing kernel optimizations, Rust frontend multimodal video/audio)
• [x] database-frontier/53: Patroni 4.1: 쿼럼 기반 동기 복제와 PostgreSQL HA 운영 기준의 재정립 — published 2026-07-27 (dynamic fallback: database/HA topic; Patroni 4.1.4 released 2026-07-07 within 90 days; quorum-based synchronous replication ANY N of M, master→primary terminology breaking change on K8s, patronictl demote-cluster/promote-cluster commands, member_slots_ttl parameter, PostgreSQL 18 support, etcd error handling improvements)
• [x] ai-frontier/59: Nemotron 3 Super 120B: LatentMoE·Mamba2 하이브리드·NVFP4 사전 학습으로 12B 활성 파라미터에서 120B 성능을 내는 방법 — published 2026-07-27 (dynamic fallback: AI systems/engineering topic; NVIDIA Nemotron 3 Super released 2026-04-15, arXiv 2604.12374, within 90 days; hybrid Mamba2-Transformer LatentMoE 120.6B/12.7B active, NVFP4 pre-training from first gradient, MTP native speculative decoding without separate draft model, 1M context, MMLU-Pro 83.73, LiveCodeBench v6 78.69)
• [x] database-frontier/54: Debezium 3.6: RocksDB 오프힙 스키마 이력·MySQL 폴 경로 최적화·Docling SMT로 CDC 운영 경계를 확장한 방법 — published 2026-07-27 (dynamic fallback: database/CDC topic; Debezium 3.6.0.Final released 2026-07-01 within 90 days; RocksDB pluggable off-heap table history/schema storage, MySQL doPoll() 75% alloc reduction 31-42% CPU reduction, Docling SMT for AI/RAG pipelines, Amazon SNS sink with FIFO support, MariaDB signal-based binlog position adjustment, Kafka Connect 4.3.0)
• [x] ai-frontier/60: Hunyuan Hy3: 295B MoE의 21B 활성 파라미터·MTP 레이어·Expert Parallelism으로 오픈소스 추론 에이전트를 배포하는 방법 — published 2026-07-27 (dynamic fallback: AI systems topic; Tencent Hunyuan Hy3 released 2026-07-06, within 90 days; 295B total/21B active MoE with 192 routed experts top-8, 3.8B MTP layer for speculative decoding without separate draft model, GQA 64/8 heads 256K context, Apache 2.0, vLLM multi-node EP deployment on 2×8×H100)
• [x] database-frontier/55: Databricks Lakeflow Connect MySQL CDC: 게이트웨이 없는 통합 파이프라인으로 소스 데이터를 레이크하우스로 싣는 방법 — published 2026-07-27 (dynamic fallback: data-platform topic; Databricks Lakeflow Connect MySQL integrated CDC pipeline Beta July 2026; no ingestion gateway, single pipeline update for extraction+application, AUTO CDC API replaces APPLY CHANGES, Row Filtering GA, Unity Catalog Connection objects and lineage)
• [x] ai-frontier/61: Grok 4.5: Cursor 코파일럿 세션으로 학습된 V9 MoE가 코딩 에이전트 토큰 비용을 4.2배 낮추는 방법 — published 2026-07-28 (dynamic fallback: AI engineering topic; Grok 4.5 released 2026-07-09, within 90 days; 1.5T parameter V9 MoE, Cursor supplemental training on real coding sessions, 500K context window, $2/$6 input/output per MTok, ~80-91 TPS at 8K, SWE-Bench Pro 64.7%, Terminal-Bench 83.3%, 15,954 tokens/task vs 67,020 for Opus 4.8 — 4.2× token efficiency)
• [x] database-frontier/56: Airflow 3.3 멀티언어 태스크 SDK(AIP-108): Java·Go 실행 경계를 Python DAG 안에 들이는 법 — published 2026-07-28 (dynamic fallback: data-platform/orchestration topic; Airflow 3.3.0 released 2026-07-06, within 90 days; AIP-108 Language Task SDK, @task.stub(queue="java") Python DAG declaration, JavaCoordinator JDK 17+ with @Task annotation, ExecutableCoordinator Go binary with airflow.Run(), Execution API XCom proxy, experimental status with possible wire protocol changes in future releases)
• [x] ai-frontier/62: Phi-4-Reasoning-Vision-15B: 훈련 시점에 생각을 심어 200B 토큰으로 경쟁력을 내는 멀티모달 추론 모델 — published 2026-07-28 (dynamic fallback: AI systems topic; Microsoft Phi-4-Reasoning-Vision-15B released 2026-03-04, arXiv:2603.03975, within 180-day fallback window at 145 days; training-time hybrid think/no-think — 20% <think> CoT traces, 80% <no_think> direct answers — not inference-time switching; SigLIP-2 Naflex dynamic-resolution vision encoder, mid-fusion architecture; 200B training tokens quality-first vs 1T+ for competitors; 240 B200 GPUs × 4 days; ScreenSpot_v2 88.2%, MathVista_MINI 75.2%, ChartQA_TEST 83.3%, MMMU_VAL 54.3%; MIT license)
• [x] database-frontier/57: Snowflake Summit 2026: Cortex Sense·Horizon Context·AI Agent Identity로 에이전트 엔터프라이즈의 데이터 제어면을 다시 그은 방식 — published 2026-07-28 (dynamic fallback: data-platform topic; Snowflake Summit 2026 June 15-18 2026, within 90 days; Cortex Sense runtime context enrichment 24%→86% accuracy on structured data questions; Horizon Context semantic/business meaning layer vs Horizon Catalog technical metadata; AI Agent Identity GA with cryptographic per-agent credentials, per-agent RBAC, Access History audit trail; Apache Iceberg v3 GA + Deletion Vectors + Snowflake-managed Iceberg storage; Snowflake Intelligence (CoWork) + Cortex Code (CoCo) GA; Cortex Training Preview; Natoma acquisition for MCP governance; Adaptive Compute GA)
• [x] ai-frontier/63: Mistral Medium 3.5: 세 모델을 하나로 합치고 추론 강도를 요청별로 조정하는 128B 오픈 가중치 모델 — published 2026-07-28 (dynamic fallback: AI engineering topic; Mistral Medium 3.5 released 2026-04-29, within 90 days; 128B dense model replaces Medium 3.1 + Magistral + Devstral 2; 88 layers hidden_size=12288, GQA 96/8 heads head_dim=128, 262144 context; reasoning_effort=none/low/medium/high/max; EAGLE speculative decoding draft head; Pixtral vision encoder from Pixtral-12B; SWE-bench 77.6%, τ³-Telecom 91.4%; $1.50/$7.50 per MTok; Modified MIT license)
• [x] database-frontier/58: Apache Arrow 25.0.0: 인메모리 컬럼형 표준이 Flight SQL과 ADBC로 엔진 간 직렬화 비용을 없애는 방법 — published 2026-07-28 (dynamic fallback: database/data-platform topic; Apache Arrow 25.0.0 released 2026-07-10, within 90 days; ARM64 SVE/SVE2 dynamic SIMD dispatch via xsimd, Flight SQL prepared-statement has_result_set metadata (GH-49497), C Data Interface variadic buffer null clarification, ListView in Parquet read/write, Bloom filter vectorized probe, ComputeLogicalNullCount on ChunkedArray, Table.ToTensor Python, ADBC 23 driver roster; zero-copy IPC eliminates serde between Arrow-native engines)
• [x] ai-frontier/64: GPT-Realtime 2.1: SIP 전화·원격 MCP·추론 강도로 실시간 음성 에이전트의 운영 경계를 다시 그은 릴리스 — published 2026-07-29 (dynamic fallback: AI engineering topic; gpt-realtime-2.1 and gpt-realtime-2.1-mini released 2026-07-06, within 90 days; SIP PSTN phone calling, remote MCP server support, image inputs, configurable reasoning_effort minimal/low/medium/high/xhigh, 25% p95 latency reduction via session caching, session duration extended to 60min, legacy model deprecation effective 2027-01-20)
• [x] database-frontier/59: Milvus 3.0: 데이터를 복사하지 않고 레이크에서 직접 벡터 검색하는 Vector Lakebase 아키텍처 — published 2026-07-29 (dynamic fallback: database/vector-search topic; Milvus 3.0 GA released 2026-07-16, within 90 days; External Collection zero-copy indexing over S3/Iceberg/Parquet/Lance/Vortex, Loon manifest-based storage engine, Vortex Arrow-compatible columnar format, MVCC Snapshot for point-in-time views, native hybrid search full-text+vector in single engine, K8s/Docker deployment, Python/Go/Node.js SDKs)
• [x] ai-frontier/65: FlashInfer 0.6: JIT 컴파일·MLA·MoE EP·Blackwell 지원으로 LLM 서빙의 커널 층을 다시 그은 어텐션 엔진 — published 2026-07-29 (dynamic fallback: AI systems topic; FlashInfer 0.6.15 active patch series July 2026, within 90 days; block-sparse KV PagedAttention format, JIT CUDA kernel compilation per unique (batch, head, seqlen) shape, MLA Matrix Absorption Trick eliminates KV restoration write, MoE Expert Parallelism fused GEMM+dispatch, Blackwell SM100/SM103 CuTe-DSL kernels; default backend for vLLM on Blackwell, SGLang, TRT-LLM, TGI, MLC-LLM)
• [x] database-frontier/60: ProxySQL 4.0 AI/MCP Tier: SQL 프록시가 에이전트 인프라가 되는 방식 — NL2SQL·MCP 서버·RAG·멀티티어 전략 — published 2026-07-29 (dynamic fallback: database/AI-infra topic; ProxySQL 4.0.6 released 2026-03-10, 141 days ago, within 180-day fallback window; latest 4.0.8 released 2026-04-17; MCP server port 6071 30+ JSON-RPC tools, NL2SQL via agentic tool-call architecture NOT internal translation, AutoDiscovery 2-phase schema harvest+LLM enrichment→mcp_catalog.db SQLite+sqlite-vec, genai_variables LLM bridge async OpenAI/Anthropic/local, AI threat detection semantic fingerprint similarity; AI/MCP Tier is EXPERIMENTAL — production stays on 3.0.x/3.1.x; Stable/Innovative/AI-MCP three-tier strategy)

• [x] ai-frontier/66: LMDeploy 0.14: TurboMind·PyTorch 이중 엔진과 FP8 KV 캐시·Responses 엔드포인트로 LLM 서빙 운영 경계를 다시 그은 방법 — published 2026-07-29 (dynamic fallback: AI systems topic; LMDeploy v0.14.0 released 2026-06-24, within 90 days; FP8 KV cache quantization, Qwen3 Omni support, OpenAI Responses-compatible endpoint, Anthropic-compatible endpoint, /get_ppl perplexity endpoint, CUDA 12.8 default wheels)
• [x] database-frontier/61: DragonflyDB 1.38~1.39: Redis 호환 인메모리 DB가 하이브리드 검색·Count-Min Sketch·TTL 최적화로 진화하는 방법 — published 2026-07-29 (dynamic fallback: database topic; DragonflyDB v1.38.0 released 2026-04, v1.39.0 released 2026-06, within 90 days; Top-K/CMS probabilistic data structures, HTTL hash-field TTL, CompactKey embedded TTL 26% memory reduction, vector range search, FT.HYBRID text+vector fusion, Porter stemming, PubSub batching +12% RPS 24-70% fewer syscalls)

• [x] ai-frontier/67: SGLang 0.5.16: Inkling 975B 첫날 지원·DSpark 내장·UnifiedRadixTree 기본화로 이종 어텐션 모델 추론 경계를 다시 그은 방법 — published 2026-07-30 (dynamic fallback: AI systems topic; SGLang v0.5.16 released 2026-07-26, within 90 days; Inkling 975B hybrid MoE SWA+full attention 5:1 day-0 support, DSpark confidence-scheduled variable-length verification speculative decoding built-in 383.7 tok/s ~5 accept length on DeepSeek-V4-Pro TP8 B300, UnifiedRadixTree single KV cache abstraction default for SWA/Mamba2/DSA layer types)
• [x] database-frontier/62: Apache DataFusion 54.0.0: LATERAL 조인·SQL 람다 내장·arrow-avro·Parquet CDC로 Rust 쿼리 엔진의 SQL 경계를 다시 그은 방법 — published 2026-07-30 (dynamic fallback: data-platform topic; Apache DataFusion 54.0.0 released 2026-06-12, within 90 days; 11 weeks 740 commits 139 contributors; LATERAL JOIN CROSS/INNER/LEFT with decorrelation; SQL lambda x->expr with array_transform/array_filter/array_any_match; arrow-avro crate Avro reader replacing avro-rs with Arrow-native type mapping; Parquet content-defined chunking CDC hash-based page boundary; join/scan/planning performance improvements; AggregateFunctionExpr::human_display() Option<&str> breaking change)
• [x] ai-frontier/68: OpenAI Codex Security CLI: 코드베이스를 읽고 취약점을 발견·검증·패치하는 AI 보안 스캐너 — published 2026-07-30 (dynamic fallback: AI engineering topic; open-sourced Apache 2.0 on 2026-07-29; find→validate-in-sandbox→patch pipeline; outputs report.md/findings.json/coverage.json/SARIF; CI/CD integration with --fail-on-severity; Node.js 22+ and Python 3.10+ required; scanner backend gated to approved customers)
• [x] database-frontier/63: Grafana Loki Kafka 수집 레이어: 페타바이트 로그를 10배 빠르게 분석하는 새 아키텍처 — published 2026-07-30 (dynamic fallback: observability/database topic; GrafanaCON 2026 April 21, Grafana 13; Loki 3.x Kafka-backed ingestion replaces 3-way replication; old problem 2.3x storage overhead from ingester time-sync drift; new path RF-1 Kafka durability, volume-based partition routing; redesigned query planner distributed across partitions; 20x less data scanned, 10x faster aggregated queries, 30% cost reduction; distributed mode now requires Kafka — breaks original "object storage only" principle)

• [x] ai-frontier/69: Claude Code 2.1 에이전트 플릿 제어: 배경 실행·서브에이전트 예산·MCP 자동화로 멀티에이전트 운영 기준을 다시 쓴 방법 — published 2026-07-30 (dynamic fallback: AI engineering topic; Claude Code v2.1.198 July 1 background subagent default + Notification hook for background sessions; v2.1.210 MCP auto-background 2-min threshold + Hook timeout misreport fix; v2.1.217 July 21 concurrent subagent cap 20 + nesting blocked + budget cap fix; v2.1.219 July 24 nesting reinstated at depth 3 + session subagent budget 200 CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION)
• [x] database-frontier/64: Redpanda 26.1: R1 엔진 완성 — Cloud Topics·Iceberg Topics·계층 스토리지가 하나의 클러스터에서 동작하는 방식 — published 2026-07-30 (dynamic fallback: data-platform/streaming topic; Redpanda 26.1 released March 2026, 152 days before 2026-07-30, within 180-day fallback window; R1 architecture complete with Cloud Topics GA — S3/ADLS/GCS as primary backing store, no local disk, 90%+ cross-AZ replication cost elimination; Iceberg Topics zero-ETL Parquet materialization; Tiered Storage cost-speed slider; GBAC OIDC group-based permissions; FIPS 140-3 upgrade)
• [x] ai-frontier/70: GPT-5.6 Sol 자기 최적화: AI가 자신의 GPU 커널을 다시 쓰고 추론 비용을 20% 낮춘 방법 — published 2026-07-31 (dynamic fallback: AI systems/engineering topic; OpenAI July 30 2026 cut Luna 80% ($1→$0.20/MTok input) and Terra 20% ($2.50→$2.00); GPT-5.6 Sol autonomously rewrote GPU kernels via Codex using Triton and Gluon, optimized load balancing, speculative decoding draft model, KV cache; FpSan validation; 20% serving cost reduction, 15%+ token generation efficiency; first confirmed case of frontier AI self-optimizing its own serving stack)
• [x] database-frontier/65: Apache Iceberg V4: 대규모 운영이 드러낸 V3의 한계와 다음 포맷 설계 방향 — published 2026-07-31 (dynamic fallback: data-platform/lakehouse topic; Iceberg Summit 2026 June 2026, v4 dev list active July 2026, within 90 days; key v4 proposals: single-file commits O(1) metadata I/O, Parquet metadata replacing Avro for columnar planning, geometry/geography/timestamp_ns new types, write-default/initial-default column values, built-in at-rest encryption; Delta Lake 5.0 Iceberg v4 metadata convergence proposal; primary sources: iceberg dev list, Iceberg Summit 2026 Snowflake recap)

• [x] ai-frontier/71: Claude API July 2026: 캐시를 유지하며 도구를 바꾸고 거절을 자동 재시도하는 방법 — published 2026-07-31 (dynamic fallback: AI engineering topic; mid-conversation-tool-changes-2026-07-01 beta: add/remove tools mid-session via system message tool_addition/tool_removal blocks without invalidating prompt cache; server-side-fallback-2026-07-01 beta: fallbacks:"default" auto-retries refused requests on Anthropic-recommended model in one API call; Managed Agents July 22 updates: initial_events session seeding, environment/memory_store lifecycle webhooks, optional version field; Claude Opus 5 GA July 24)
• [x] database-frontier/66: etcd 3.7.0: RangeStream·Protobuf 완전 교체·v2 제거로 다시 그은 운영 경계 — published 2026-07-31 (dynamic fallback: database/infrastructure topic; etcd v3.7.0 released 2026-07-08, within 90 days; RangeStream RPC beta for chunked large result streaming, SharedBufReadTxMode 2x lease/auth perf, FastLeaseKeepAlive feature gate, keys-only range in-memory optimization; gogo/protobuf → google.golang.org/protobuf complete migration, grpc-middleware v1→v2; v2store/client/discovery completely removed; bbolt v1.5.1, raft v3.7.0, Go 1.26.5; 5 new Watch and request-duration metrics)

• [x] ai-frontier/72: Unsloth 2026: GRPO·MoE 12배 가속·90% VRAM 절감으로 소비자 GPU에서 추론 모델을 훈련하는 방법 — published 2026-07-31 (dynamic fallback: AI engineering topic; Unsloth 2026.7.x; smart gradient checkpointing RAM offload ~1% speed loss; MoE grouped_mm Triton kernels 12× faster 35% less VRAM; Split-LoRA avoids LoRA delta materialization; sequence packing removes shared prompt duplication; 7× longer context RL; AMD GPU and MCP control endpoints added 2026)
• [x] database-frontier/67: Lance 2.x: AI 워크로드를 위한 컬럼형 포맷이 Parquet와 다른 이유 — published 2026-07-31 (dynamic fallback: data-platform/lakehouse topic; Lance format 2.x; Structural Encoding O(1) positional random access without neighbor decoding; Fragment+Deletion File MVCC zero-copy versioning; built-in HNSW/IVF_PQ vector index and FTS; DuckDB lance-duckdb extension GA 2026-05-21; multimodal lakehouse pattern Lance+Iceberg+DuckDB; arXiv:2504.15247)

• [x] ai-frontier/73: Nemotron-Labs-3-Puzzle 75B: Iterative Puzzle 압축으로 MoE 추론 비용을 절반으로 줄이는 방법 — published 2026-08-01 (dynamic fallback: AI systems topic; Nemotron-Labs-3-Puzzle-75B-A9B released 2026-07-09, arXiv:2607.04371, within 90 days; Iterative Puzzle alternating bounded pruning + short KD recovery rounds; 120.7B→75.3B total params, 12.8B→9.3B active, 70GB→44.5GB weights; 2.03× throughput on 8×B200, 1M-token concurrency 1→8 on single H100; 88 blocks: 40 Mamba + 40 MoE + 8 attention; BF16/FP8/NVFP4 on HuggingFace)
• [x] database-frontier/68: DBOS 2.29: PostgreSQL 하나로 내구성 있는 워크플로우를 운영하는 방법 — published 2026-08-01 (dynamic fallback: database/data-platform topic; DBOS Transact v2.29.0 released 2026-07-30, within 90 days; Go package released 2026-07-15, Java 1.0 GA; @DBOS.workflow() + @DBOS.step() exactly-once semantics via Postgres transactions; dynamic queues runtime create/update/delete; workflow metadata JSONB with GIN index; durable streams 20× throughput improvement; LISTEN/NOTIFY integration 60K writes/s ms latency; Transactional Datasources app data + checkpoint in one transaction; no separate orchestration server)

• [x] ai-frontier/74: GLM-5.2: IndexShare 희소 어텐션으로 744B MoE가 1M 토큰 추론 비용을 2.9배 줄이는 방법 — published 2026-08-01 (dynamic fallback: AI engineering topic; GLM-5.2 released 2026-06-13, within 90 days; IndexShare DSA top-k index reuse every 4 layers → 2.9× fewer FLOPs at 1M context; 744–753B total / 40B active MoE, MIT license; MLA + DSA base; 128 experts 8 active per token; HuggingFace glm-community/GLM-5.2 and API via Z.ai)
• [x] database-frontier/69: PostgreSQL 18 DBA 가이드: 비동기 I/O·uuidv7·OAuth로 달라진 운영 기준 — published 2026-08-01 (dynamic fallback: database topic; PostgreSQL 18.4 security patch released 2026-05-14, within 90 days; covers async I/O subsystem io_method=worker/io_uring, uuidv7 B-tree fragmentation fix, virtual generated columns new default, OAuth 2.0 client credentials auth, 11 CVEs including CVE-2025-4207 CVSS 8.8; pg_upgrade path from PG14+)

• [x] ai-frontier/75: 컨텍스트 엔지니어링: 프롬프트를 넘어 LLM 에이전트의 입력 공간을 설계하는 방법 — published 2026-08-01 (dynamic fallback: AI engineering topic; arXiv:2603.09619 "Context Engineering: From Prompts to Corporate Multi-Agent Architecture" March 2026, arXiv:2604.04258 "Context Engineering: A Practitioner Methodology" April 2026; four operations Write/Select/Compress/Isolate from LangChain; five quality criteria relevance/sufficiency/isolation/economy/provenance; multi-agent context isolation patterns; within 180-day fallback window)
• [x] database-frontier/70: OrioleDB beta16: PostgreSQL의 버퍼 매니저 병목을 없애고 새 스토리지 엔진으로 4.5배 성능을 내는 방법 — published 2026-08-01 (dynamic fallback: database topic; OrioleDB beta16 released 2026-06-18, within 90 days; PostgreSQL Table Access Method API, lock-less buffer reading via squizzled pointers, MVCC via Undo Log, Copy-on-Write checkpoints, Row-level WAL, Index-Organized Tables, 64-bit transaction IDs; PostgreSQL 18 support added; SERIALIZABLE support added; 4.5× read-write TPS, 4× read-only, 6× IOPS efficiency vs standard Heap)

• [x] ai-frontier/76: EAGLE 3.1: 투기적 디코딩의 어텐션 드리프트를 FC 정규화로 해결하고 긴 컨텍스트 수용 길이를 2배로 늘린 방법 — published 2026-08-02 (dynamic fallback: AI systems topic; EAGLE 3.1 released 2026-05-26 by EAGLE team + vLLM + TorchSpec; FC normalization fixing attention drift in multi-layer feature fusion; up to 2× acceptance length improvement on long-context workloads; full EAGLE-3 checkpoint compatibility)
• [x] database-frontier/71: Prometheus 3.13 LTS: 운영자가 확인할 보안 패치·PromQL 변경·Native Histogram 성능 개선 — published 2026-08-02 (dynamic fallback: database/observability topic; Prometheus v3.13.0 LTS released 2026-07-01, supported through 2027-07-31; CVE-2026-44990 XSS fix, CVE-2025-4673 HTTP auth header leak fix, min_of/max_of PromQL rename, SHA-256 pagination token, 2× regex speedup, 12-15% chunk population speedup, V2 histogram WAL 50% allocation reduction)

• [x] ai-frontier/77: Claude Managed Agents Dreaming·Outcomes·멀티에이전트: 자기 개선하고 자기 평가하는 에이전트 운영 패턴 — published 2026-08-02 (dynamic fallback: AI engineering topic; Anthropic Managed Agents Dreaming + Outcomes + Multiagent Orchestration announced May 6 2026 at Code with Claude; dreaming research preview, outcomes + multiagent public beta; Dreams API POST /v1/dreams async job, beta headers managed-agents-2026-04-01 + dreaming-2026-04-21, reads memory_store + sessions[], writes new output memory store; Outcomes: user.define_outcome event, grader in separate context window, max_iterations default 3 max 20, evaluation events span.outcome_evaluation_start/ongoing/end; multi-agent orchestration subagent spawning and shared memory store)
• [x] database-frontier/72: Dolt 2.0: Prolly Tree와 Archive 포맷으로 SQL 데이터베이스에 Git-style 버전 관리를 넣은 방법 — published 2026-08-02 (dynamic fallback: database topic; Dolt 2.0 released May 11 2026 within 90 days; Prolly Tree content-addressed probabilistic B-tree with rolling hash chunking, History Independence, structural sharing; Archive format dictionary compression + zStd 30-50% additional storage reduction enabled by default; AutoGC default-on; Adaptive Storage TOAST-like for TEXT/JSON/BLOB; beta MariaDB Vector type version-controlled vectors; MySQL-compatible SQL interface dolt_branches dolt_diff_* AS OF syntax)

• [x] ai-frontier/78: Ling-3.0-Flash: 에이전트 추론을 위한 1/64 전문가 희소성과 KDA+MLA 하이브리드 설계 — published 2026-08-02 (dynamic fallback: AI engineering topic; Ling-3.0-Flash released 2026-07-26 by InclusionAI/Ant Group, within 90 days; 124B total / 5.1B active MoE with 1/64 expert sparsity; KDA+MLA 5:1 hybrid attention: KDA linear-complexity long-context + MLA for precise retrieval; 262K token context; cluster-level hierarchical KV caching reducing long-input TTFT 60-80%; designed for production-scale agents at Alipay; no formal technical report published as of 2026-08-02)
• [x] database-frontier/73: ScyllaDB 2026.2: Trie SSTable 인덱스 기본화·Raft 강한 일관성·DynamoDB 스트림 GA — published 2026-08-02 (dynamic fallback: database topic; ScyllaDB 2026.2 released 2026-06-29, within 90 days; Trie-based SSTable index promoted to default: 20-230% throughput / 31-63% latency improvement; Strongly Consistent Tables via per-Tablet Raft groups experimental — 1 round-trip vs Paxos LWT 3 round-trips; Alternator DynamoDB Streams GA for CDC pipelines; vNode-to-Tablet zero-downtime online migration experimental; connection storm mitigation P99 5s→3.5ms)

• [x] ai-frontier/79: 에이전트 메모리 설계 2026: Mem0 단일 패스 추출·다중 신호 검색으로 LLM 에이전트의 기억을 토큰 효율적으로 관리하는 방법 — published 2026-08-03 (dynamic fallback: AI engineering topic; mem0 token-efficient memory algorithm published April 2026, within 90 days; ADD/UPDATE/DELETE/NOOP four-operation memory management; single-pass extraction treats user+agent utterances equally; multi-signal retrieval: semantic similarity + BM25 keyword + entity graph matching; benchmarks: LoCoMo 91.6 at 6956 tok/query, LongMemEval 93.4 at 6787 tok/query, BEAM 1M 64.1, BEAM 10M 48.6; key gains: temporal reasoning +29.6, multi-hop +23.1; three scopes: user_id, run_id, agent_id; 25% degradation from 1M to 10M tokens is open problem)
• [x] database-frontier/74: DuckDB 1.5 Variegata: VARIANT 타입·Quack 프로토콜·DuckLake로 임베디드 OLAP의 경계를 다시 그은 방법 — published 2026-08-03 (dynamic fallback: database/data-platform topic; DuckDB 1.5.0 released 2026-03-09 codenamed Variegata, 1.5.3 released 2026-05-20, 1.5.5 released 2026-07-22, all within 90 days; VARIANT type: binary-encoded self-contained type info per row, JSON shredding to Parquet typed columns, 100x JSON analysis speedup, Snowflake VARIANT Parquet compatibility; Quack protocol: HTTP client-server, serializes DuckDB internal vector blocks directly no transcoding, shared token auth, multi-user concurrent writes, core_nightly until DuckDB 2.0; DuckLake 1.0 GA 2026-04-13: SQL catalog backed by SQLite/PostgreSQL/DuckDB, Iceberg REST catalog compat in 1.5.3; GEOMETRY built-in type, Parquet first-class spatial column; DuckDB 2.0 planned fall 2026 with Quack stable)

• [x] ai-frontier/80: XGrammar-2: 에이전트 LLM의 동적 구조화 생성을 6배 빠르게 만드는 TagDispatch와 Cross-Grammar Cache — published 2026-08-03 (dynamic fallback: AI systems topic; arXiv:2601.04426, ACM CAIS 2026 May 26-29 2026, within 90 days; TagDispatch tag-triggered grammar switching within a single request — <|tool_call|>→tool grammar, <|tool_call_end|>→free-form, zero-copy mask handoff; Cross-Grammar Cache substructure reuse across different grammars sharing terminals/productions — identifies common subautomata and shares precomputed token masks; JIT compilation 8.1× preprocessing speedup vs eager; Earley-based adaptive token mask cache eliminates redundant mask recomputation; repetition state compression 99.6× fewer states for repetition structures; overall 6× tool-calling compilation speedup, near-zero end-to-end overhead; integrated as default structured generation engine in vLLM and SGLang; key limitation: tag-grammar pairs must be defined at request initialization time, not dynamically added mid-stream)
• [x] database-frontier/75: Databricks Lakebase: Lakehouse에 PostgreSQL을 통합하여 OLTP와 분석을 하나의 플랫폼으로 운영하는 방법 — published 2026-08-03 (dynamic fallback: data-platform/database topic; Databricks Lakebase autoscaling default March 12 2026, autoscaling Beta with expanded CU range June 2026, both within 90 days; built on Neon acquired May 2025 ~$1B; storage-compute separation Neon WAL service + S3 page storage, scale-to-zero, CoW branching; autoscaling 0.5-32 CU, HA 1 primary + up to 3 read replicas; Synced Tables: Unity Catalog Delta Lake → Lakebase PostgreSQL via Lakeflow pipelines, read-only, supports native SQL apps; Lakehouse Sync: Lakebase PostgreSQL → Unity Catalog Delta Lake via CDC, SCD Type 2 history, Beta status; Unity Catalog required; connection pooling via PgBouncer/PgCat recommended; sync delay 30s-5min for Synced Tables; Lakebase PostgreSQL 17; compared to standalone Neon: Unity Catalog governance, bidirectional sync, Databricks billing, no external endpoint)

• [x] ai-frontier/81: ObjectCache: KV 캐시를 오브젝트 스토리지에 저장해 LLM 서빙 DRAM 병목을 해소하는 방법 — published 2026-08-03 (dynamic fallback: AI systems topic; arXiv:2605.22850, ETH Zurich + HPE Labs + CUHK Shenzhen, May 2026, within 90 days; layerwise object-storage retrieval for KV cache reuse — stores KV cache in S3-compatible object storage instead of DRAM pools; semantic gap: object storage moves objects but GPU needs layer-ordered streaming; layerwise delivery protocol matches GPU consumption order to overlap I/O with compute; bandwidth scheduler tracks per-request layer progress for priority-based allocation; NIXL + Ceph RGW + DAOS prototype on 100 Gbps RoCE; 64K context: 5.6% TTFT overhead vs local DRAM; 4K context: 56-75ms overhead; bandwidth scheduler 1.2-1.8x TTFT improvement vs equal sharing)
• [x] database-frontier/76: MySQL 26.7: 캘린더 버전 관리 전환과 Change Stream Applier·Thread Pool 오픈소스화로 달라진 운영 기준 — published 2026-08-03 (dynamic fallback: database topic; MySQL 26.7.0 GA released July 2026, within 90 days; first Innovation release post-LTS 9.7 using CalVer YY.M.P format; Change Stream Applier APPLIER_VERSION=2 opt-in per-channel replacement for MTA with 1-1024 worker threads per channel; Thread Pool Plugin moved from Enterprise to Community Edition; InnoDB undo log truncation robustness improvements; upgrade performance improvement for general_log/slow_log ALTER TABLE; next Innovation MySQL 26.10.0 October 2026; CSA is experimental in 26.7.0)

• [x] ai-frontier/82: SMetric: 에이전트 LLM 서빙에서 세션 기준 스케줄링이 클러스터 처리량을 높이는 방법 — published 2026-08-04 (dynamic fallback: AI systems topic; arXiv:2607.08565, KDD 2026 Research Track, July 2026, within 90 days; session-centric scheduling for agentic LLM workloads where agents not humans issue requests; agentic KV cache reuse >80% vs 54-62% chat; SMetric = session turn number as stateless scheduling metric; first request routes by load balance, follow-up requests by cache-aware routing to same instance; two-tier KV cache: local + global tier; 10-16% higher cluster TPS under PD colocation, 2-34% higher prefill TPS under disaggregation; tested on Qwen3 models; authors: SJTU team)
• [x] database-frontier/77: Redis 8.10: Compact Hashes·HIMPORT·BACKUP으로 달라진 인메모리 DB 운영 기준 — published 2026-08-04 (dynamic fallback: database topic; Redis 8.10.0 GA released 2026-07-29, within 90 days; Compact Hashes template encoding stores field names once for schema-sharing keys, config hash-rdb-load-* params, metrics hash_templates/hash_template_keys/used_memory_hash_templates; HIMPORT high-throughput compact hash bulk insertion; BACKUP node-side MP-AOF-based backup and restore; TLS peer certificate-based server-to-server authentication; new commands LMOVEM/BLMOVEM move multiple list elements, SUNIONCARD/SDIFFCARD set cardinality operations; XREAD/XREADGROUP MAXCOUNT/MAXSIZE args; performance: lpSeek validation, wide HSET/HMSET batching, HNSW one-byte locks, rax leaf-inlining)

• [x] ai-frontier/83: 이종 하드웨어 LLM 추론 설계: 4개 설계 축과 3가지 경계 결정 — published 2026-08-04 (dynamic fallback: AI systems topic; arXiv:2606.29708 published 2026-06-30, within 90 days; heterogeneous PD inference design space, 4 axes: accelerator/precision/interconnect/KV-residency, 3 boundary decisions: compute-placement/KV-representation/KV-ownership, precision policy per-role not global; A100 wastes 97% HBM bandwidth during prefill compute-bound phase; heterogeneous configs 61% cheaper/hour 48% cheaper/token vs homogeneous; 18 institutions including Shanghai Innovation Institute SJTU MetaX)
• [x] database-frontier/78: Debezium 3.7.0.Alpha1: CockroachDB 직접 CDC·Ingres·YashanDB 커넥터·PyDebeziumAI 벡터 DB로 CDC 경계를 확장한 알파 프리뷰 — published 2026-08-04 (dynamic fallback: database/CDC topic; Debezium 3.7.0.Alpha1 released 2026-07-30, within 90 days; CockroachDB sinkless CDC via core changefeed CREATE CHANGEFEED FOR TABLE without INTO clause streams over SQL connection without Kafka; Ingres/YashanDB incubating source connectors; JDBC sink expanded to CockroachDB/SingleStore/StarRocks; PyDebeziumAI incubating INSERT/UPDATE/DELETE auto-translated to vector ops for Chroma/PGVector/Milvus; ExtendedTimestamp handles high-precision timestamps exceeding INT64; schema history memory optimization; Debezium Engine shutdown strategies IMMEDIATE/SAFE/GRACEFUL; 97 issues resolved; Alpha status not production-ready)
• [x] ai-frontier/84: LLM 서빙 KV 캐시 최적화 체계화: 실행·배치·표현의 3차원 분류체계 — published 2026-08-04 (dynamic fallback: AI systems topic; arXiv:2607.08057, ACL 2026 Findings, July 9 2026, University of Melbourne + HUST; systematic survey of KV cache optimization across temporal/spatial/structural dimensions for LLM serving; temporal: execution-level reuse (semantic cache, prefix cache, RadixAttention) and scheduling-level optimization (disaggregation, SMetric-style routing); spatial: memory hierarchy placement (DRAM→SSD→object storage: PagedAttention, vLLM, ObjectCache) and migration policies; structural: representation compression (quantization: FP8/INT8/binary, sparsity, pruning, low-rank factorization); cross-dimension interactions; systems mapped: vLLM, SGLang, ObjectCache, SMetric, Mooncake, NVIDIA Dynamo)
• [x] database-frontier/79: Neo4j 2026: CalVer 전환·바이너리 양자화 GA·GenAI 함수·DISJOINT BY로 그래프 DB 운영 기준 변화 — published 2026-08-04 (dynamic fallback: database/graph-DB topic; Neo4j 2026.05.0 released May 28 2026, 2026.06.0 released July 2 2026, within 90 days; first graph database covered in this series; CalVer YY.MM.P replacing semantic versioning from 2026; Binary Quantization GA in 2026.05 (fp32→1-bit, 32x DRAM reduction, 97-99% recall with re-ranking, configurable quantizationFallback threshold); Hi-Fidelity Quantization preview in 2026.06 (2-stage: BQ candidate list + full-vector re-rank, >99% recall); GenAI Cypher functions: chunkByTokenLimit/chunkBySentence for chunking, aggregateStructuredCompletion for property extraction with JSON schema; DISJOINT BY clause for MERGE/SET preventing deadlocks via lock-free parallel writes; ABAC native user tags ga; ACYCLIC path constraint; new string functions: substrings/charCodeAt/codePointAt; CDC txCommitTime property; operator upgrade path: neo4j.conf rolling update)

• [x] ai-frontier/85: ExpertPlex: MoE LLM 서빙에서 전문가 가중치를 공유하고 타일 단위로 프리필·디코드를 조율하는 방법 — published 2026-08-05 (dynamic fallback: AI systems topic; arXiv:2607.18002 submitted July 20 2026, Peking University, within 90 days; module-level P/D disaggregation: shares Expert FFN weights (>95% duplication eliminated) across prefill and decode pools while disaggregating only lightweight Attention modules; Adaptive Persistent Kernels (APK): single resident CUDA kernel per MoE GPU eliminates kernel relaunch overhead, schedules at CTA cluster (tile) granularity, bounded preemption at tile boundaries for urgent decode without CPU involvement; attention-initiated MoE communication: overlaps network comm with computation, reduces network contention vs expert-initiated; tile-to-cluster model: joint cross-stack optimizer for placement, parallelism, comm-compute overlap, and tile scheduling; results: up to 2.01× goodput improvement on MiniMax-M2.7 and GLM-5.1-FP8)
• [x] database-frontier/80: CrateDB 6.4: LTTB 시계열 다운샘플링·GROUP BY 최적화·PostgreSQL 호환성 확장 — published 2026-08-05 (dynamic fallback: database topic; CrateDB 6.4.0 released July 20 2026, within 90 days; distributed SQL on shared-nothing architecture + Lucene shards + PostgreSQL wire protocol; largest_triangle_three_buckets(x,y,n) SQL aggregate: LTTB algorithm selects n representative points preserving visual shape via max-area triangle selection per bucket, returns (x,y) pair array; GROUP BY optimization: SortedDocValues ordinal-based processing for string-keyed columns without WHERE clause, up to 140% faster; to_char() improvements: ordinal suffix TH/th, double-quote literal quoting, ~25% speed improvement; pg_catalog.pg_user view added; DATE + interval arithmetic now returns DATE (was TIMESTAMP); ALTER REPOSITORY updates backup repo settings without drop+recreate; blake3(str) scalar function for BLAKE3 checksums; varchar(n) in aggregations now identical to text)
• [x] ai-frontier/86: GoodServe: 이종 GPU 클러스터에서 에이전트 LLM 추론을 굿풋 최적화하는 Predict-and-Rectify 라우팅 — published 2026-08-05 (dynamic fallback: AI engineering topic; arXiv:2605.16867 published May 16 2026, SJTU + CUHK Shenzhen, within 90 days; GoodServe predict-and-rectify routing for agentic LLM inferences over heterogeneous GPU clusters; MoE-style MLP ensemble output length predictor: 1.4× better than LLM-based, 3.8× better than history-based; EMA-smoothed black-box instance efficiency estimation; just-enough instance selection heuristic combining demand and resource info; runtime migration on SLO violation risk detection; up to 27.4% goodput improvement vs existing routing; 512-instance cluster simulation at 10K req/s with 5ms routing overhead; tested on H800/A800/A40/V100 heterogeneous cluster)
• [x] database-frontier/81: Confluent Platform 8.3: Kafka 4.3 기반 Flink SQL 통합·KRaft 전환 단순화·Kafka Streams 관측성 개선 — published 2026-08-05 (dynamic fallback: data-platform/streaming topic; Confluent Platform 8.3.0 released late July 2026, within 90 days; built on Apache Kafka 4.3.0 (KRaft-only); CMF 2.4 consolidates Flink SQL management: CC Flink UI deprecated, CMF UI becomes single-pane with 16 read-only + 20 write tools; CFK 6 new CRDs for Flink SQL objects as Kubernetes CustomResources preview; KRaft migration pre-check utility, 2.5DC support, dynamic quorum Controller scale-up without rolling restart; Kafka Streams metrics in USM: throughput/consumer-lag/end-to-end latency/stream thread state, KRaft mode 8.3.0+ only; Confluent Health+ EOL begins, migrate to USM)

• [x] ai-frontier/87: AdaServe: 요청마다 SLO가 다를 때 투기적 디코딩을 맞춰 LLM 서빙 효율을 높이는 방법 — published 2026-08-06 (dynamic fallback: AI systems topic; arXiv:2501.12162, EuroSys 2026 Edinburgh April 27-30 2026, within 90 days; first LLM serving system for multi-SLO via SLO-customized speculative decoding; hardware-aware speculation tree construction per request SLO budget; speculate-select-verify pipeline decouples draft generation from verification token selection; dynamic parameter adaptation to workload variation; SLO violation reduced up to 4.3×, goodput improved up to 1.9× vs best baseline; 73% higher SLO attainment, 74% higher goodput vs state-of-the-art; authors: Zikun Li, Zhuofu Chen et al., CMU Zhihao Jia lab)
• [x] database-frontier/82: Databricks Unity AI Gateway GA: 에이전트·모델·MCP를 Unity Catalog 안에서 거버넌스하는 방법 — published 2026-08-06 (dynamic fallback: data-platform topic; Unity AI Gateway GA announced August 4 2026, within 90 days; extends Unity Catalog to govern AI agents, LLM models, MCP servers, and skills; three governance dimensions: Asset Governance (ABAC registration), Traffic Governance (routing, hard spend caps, token cost attribution), Behavior Governance (service policies, guardrails, unified tracing); MCP server registration as securable objects; Smart Routing beta for quality/cost/availability-aware model selection; service policies and LLM guardrails remain in Beta; Data + AI Summit 2026 June announcements; GA builds on Unity Catalog existing GRANT/REVOKE and Access History infrastructure)
• [x] ai-frontier/88: HYPIC: 하이브리드 어텐션 LLM 서빙에서 포지션 독립적 캐싱으로 TTFT를 2.45배 낮추는 방법 — published 2026-08-06 (dynamic fallback: AI systems topic; arXiv:2607.01299 submitted July 1 2026, revised July 12 2026, within 90 days; first serving system for position-independent caching on hybrid-attention models; problem: linear-attention layers use per-request recurrent state not per-token KV cache so existing PIC breaks; solution: segment-cumulative transition operator (T_S, e_S) as new caching primitive for linear attention, enabling O(1) state composition of independently cached segments; boundary seam window repairs full-attention layer continuity across segments; cold prefill parallelized across workers via segment self-containment; results: 2.45× TTFT reduction, 2.0× peak throughput, within 3.3 points full-recompute accuracy, 6.1× cold-prefill speedup at 8 workers; tested on 4 hybrid-attention models and 5 workloads)
• [x] database-frontier/83: PostgreSQL 19 Beta 2: SQL/PGQ 그래프 쿼리·REPACK CONCURRENTLY·병렬 오토베이큠으로 달라진 운영 기준 — published 2026-08-06 (dynamic fallback: database topic; PostgreSQL 19 Beta 2 released July 16 2026, feature freeze, GA expected September-October 2026, within 90 days; SQL/PGQ: SQL:2023 standard property graph queries on relational tables via CREATE PROPERTY GRAPH + GRAPH_TABLE + MATCH syntax, no separate graph engine, fixed-depth patterns only in first version; REPACK CONCURRENTLY: pg_squeeze promoted to core, SHARE UPDATE EXCLUSIVE for initial copy + internal replication slot + logical decoding to capture changes + brief ACCESS EXCLUSIVE only for file swap, consolidates VACUUM FULL / CLUSTER / pg_repack / pg_squeeze; parallel autovacuum: autovacuum_max_parallel_workers GUC, granular scoring with autovacuum_freeze_score_weight and autovacuum_vacuum_score_weight; pg_plan_advice extension bundled for planner hint stabilization via pg_stash_advice; sequence logical replication: sequences now synchronized in publications; GROUP BY ALL: auto-groups non-aggregate expressions; new monitoring: pg_stat_lock per-lock-type stats, pg_stat_autovacuum_scores; default changes: lz4 TOAST default, JIT off by default)

• [x] ai-frontier/89: SpectrumKV: Prefill-Decode 분리 서빙에서 토큰마다 다른 정밀도로 KV 캐시 전송 대역폭을 절반으로 줄이는 방법 — published 2026-08-06 (dynamic fallback: AI systems topic; arXiv:2606.08635 submitted June 7 2026, within 90 days; SpectrumKV addresses KV cache as network payload in PD disaggregation; replaces binary token selection (PDTrim) with 3-tier mixed-precision per-token assignment: FP16 for attention sinks and high-importance tokens, INT8 for mid-importance, INT4 for low-importance when model tolerates; importance scoring and quantization at prefill worker before transfer, dequantization at decode worker via table lookup after receipt; deployment-time NIAH probe (3 trials) determines INT4 compatibility: Qwen2.5-7B fails INT4 falls back to FP16+INT8, Mistral-7B and Gemma-2-9B stable with all 3 tiers; at b=0.5 budget: TTFT 50-62% reduction across all model/length pairs; WikiText-2 perplexity change +1.97% to -0.44% vs PDTrim +25.85% to +35.63%; decode throughput stable or improved at longer contexts due to reduced memory bandwidth pressure)
• [x] database-frontier/84: Apache Iceberg 1.11.0: 서버사이드 스캔 계획·DynamicIcebergSink·메타데이터 암호화로 REST 카탈로그가 프로덕션 수준이 된 방법 — published 2026-08-06 (dynamic fallback: data-platform/lakehouse topic; Apache Iceberg 1.11.0 released May 19 2026, within 90 days; server-side scan planning in REST catalog: POST .../plan request returns FileScanTasks, three response modes: immediate/plan-id polling/parallel /tasks for large scans, planning moves off query engine driver reducing memory consumption and startup time; partition statistics scan API for optimizer use; DynamicIcebergSink for Flink (experimental): single sink routes records to runtime-determined table, on-demand table creation, dynamic schema evolution and partition spec adaptation, breaks one-sink-per-table constraint; end-to-end metadata encryption first production-ready: AES-256-GCM, envelope encryption with DEK+KEK, external KMS support (AWS KMS, Google KMS, HashiCorp Vault), key rotation built-in, manifest list and manifest file encryption with integrity checks; Variant type support expanded for Parquet/Arrow/filtering; Spark 4.1 + 3.5.8 added (MERGE INTO schema evolution, LIMIT pushdown); Flink 2.1 with deletion vector support; Java 11 dropped requires Java 17+; deprecated positional delete APIs removed)

• [x] ai-frontier/90: MuxWise: GPU 하나 안에서 Prefill과 Decode를 동시에 실행해 처리량을 2.2배 높이는 방법 — published 2026-08-07 (dynamic fallback: AI systems topic; arXiv:2504.14489 ASPLOS '26, April 2026, SJTU/HKU/NUS, within 90 days; intra-GPU prefill-decode multiplexing via SM partitioning; Prefill and Decode run concurrently on disjoint SM sets within a single GPU using CUDA Persistent Kernels + SM masks; three components: bubble-less multiplex engine eliminates inter-node KV transfer and P/D wait bubbles; contention-tolerant estimator predicts HBM+L2 bandwidth contention for dynamic SM ratio tuning; SLO-aware dispatcher tracks TTFT/TBT SLO budgets and adjusts admission; 2.20× average throughput improvement vs disaggregated baselines, 3.06× peak; not based on CUDA MPS/MIG — single-process SM-level partitioning; requires model to fit on a single GPU; GPU idle bubble reduced from 15-30% to near-zero; fixed 50:50 SM split without estimator loses up to 40% throughput)
• [x] database-frontier/85: OpenSearch 3.8: MCP 통합·4.16배 빠른 벡터 수집·gRPC 스트리밍으로 AI 검색 플랫폼 운영 기준을 새로 쓴 릴리스 — published 2026-08-07 (dynamic fallback: database/search-platform topic; OpenSearch 3.8 released August 2026 after 3.7 June 9 2026, within 90 days; Base64 Float32 vector encoding: 768-dim JSON array 16KB→4KB, parsing cost reduced from O(dims) to O(1) via ByteBuffer direct read, 4.16× index throughput improvement, backward-compatible with JSON array; MCP server plugin (opensearch-mcp): exposes search/index_info/knn_search/aggregation tools to AI agents via Model Context Protocol, Claude/Cursor compatible, requires read-only role separation in production; gRPC streaming ML inference: gRPC bidirectional streaming for Neural Search plugin to external model servers (Triton/TGI), Protobuf binary, HTTP/2 multiplexing, enables streaming reranking; radial search improvements: HNSW engine radial mode added, max_distance+min_score dual threshold, ef_search dynamic control per query; developer tooling: visual PPL query builder in Dashboards, Grok Debugger in Dev Tools, automated search relevance evaluation with NDCG/MAP/Recall@K; rolling upgrade 3.7→3.8 supported; MCP plugin separate install required)

• [x] ai-frontier/91: VeriCache: 손실 KV 캐시 압축을 무손실 추론으로 바꾸는 검증-스왑 프레임워크 — published 2026-08-07 (dynamic fallback: AI systems topic; arXiv:2605.17613 May 2026, within 90 days; VeriCache makes lossy KV cache compression lossless via draft-verify pattern; compressed KV cache (SnapKV/H2O/ScissorHands/KVQuant) runs as drafter on GPU HBM while full-precision KV transfers from DRAM/SSD via PCIe simultaneously; verification sync at block boundaries: accepted tokens output directly, rejected tokens resampled from full KV to restore exact distribution; HBM-bound decode and PCIe-bound full-KV swap run in parallel overlapping both latencies; 25-40 accepted tokens per verification round vs 5-10 for small-model speculative decoding; 3.82× throughput on Llama-2-70B 128K context, 2× for remote prefix caching; KL divergence < 0.01 nats vs uncompressed baseline; uniform compressor interface supports any stateless KV compression method; requires DRAM or SSD capacity for full KV cache storage; unsuitable for short context or interactive workloads where PCIe transfer overhead dominates)
• [x] ai-frontier/92: Claude Enterprise Inference Hooks: 모든 프롬프트 앞에 DLP 검사점을 놓는 방법 — published 2026-08-07 (dynamic fallback: AI systems/enterprise topic; Anthropic announcement August 5 2026, within 90 days; Enterprise Inference Hooks add a webhook checkpoint between every incoming prompt and the Claude model; DLP server (Netskope/Palo Alto XSIAM/Zscaler/Proofpoint) receives the full prompt text before it reaches Claude and returns allow/deny/transform; shadow mode for observe-only deployment; hooks run server-side on the Anthropic platform, not client-side; supported hook types: before_prompt (intercept and optionally transform), after_response (inspect and optionally block); authentication via HMAC-SHA256 request signing; 300ms p99 SLA for hook server response; existing Claude system-prompt guardrails continue to layer underneath hooks)
• [x] database-frontier/86: Databricks Lakehouse//RT: Reyden 엔진이 전용 실시간 DB를 대체하는 방법 — published 2026-08-07 (dynamic fallback: database topic; announced DAIS June 16 2026, Beta July 15 2026, within 90 days; Reyden engine ground-up rewrite not based on Photon, named after Reynold Xin = Reynold's Dream Engine, Apache Spark co-creator; sub-100ms P99 at 12,000 QPS, 10ms on smaller datasets; 16× better than existing real-time stacks StarRocks/ClickHouse/Druid/Redis-based; reads Delta Lake and Apache Iceberg tables directly from cloud object storage with no data copying; Delta Log in-memory cache eliminates file discovery overhead; column statistics + file skipping for point queries; read-only ANSI SQL only (no DML); Unity Catalog required for all table access, governance policies apply automatically to real-time queries; eliminates separate real-time serving stack and data copy pipelines; Beta only, GA SLA not guaranteed; large full-scan queries still better served by Photon)
• [x] database-frontier/87: Apache Flink CDC 3.6: Oracle·Hudi·PostgreSQL 스키마 진화를 단일 파이프라인으로 묶는 방법 — published 2026-08-07 (dynamic fallback: data-platform/CDC topic; Flink CDC 3.6.0 released March 30 2026, 130 days before write date, outside 90-day window but within 180-day window, stated in chapter; Oracle Source Pipeline connector promoted from Flink SQL to declarative YAML pipeline; Oracle LogMiner-based CDC requires supplemental logging enabled; Hudi Sink Pipeline connector new in 3.6: CoW/MoR support, compaction.trigger.strategy num_commits/time_elapsed, Hive Metastore sync; PostgreSQL Schema Evolution: DDL propagation to sinks now supported, metadata column access from WAL; Transform framework: VARIANT type + JSON_VALUE/JSON_ARRAY functions for semi-structured column extraction; Fluss Sink lenient mode stabilized for streaming lakehouse; schema.change.behavior modes: EVOLVE all DDL throw on fail / TRY_EVOLVE tolerate failure / LENIENT convert+drop incompatible / IGNORE swallow all DDL / EXCEPTION error on any DDL; LENIENT recommended for most CDC pipelines; Flink 1.20.x + 2.2.x supported; JDK 11 minimum required, Java 8 dropped)
• [x] ai-frontier/93: Diffusion LLM 서빙: 양방향 어텐션이 KV 캐시 전제를 깨는 방식과 Fast-dLLM·HERALD·DiLaServe의 해법 — published 2026-08-08 (dynamic fallback: AI systems topic; Fast-dLLM arXiv:2505.22618 NVlabs ICLR 2026 May 2026, HERALD arXiv:2606.21633 SNU+UC Berkeley June 19 2026, DiLaServe arXiv:2606.29094 June 27 2026, Sangam arXiv:2607.04206 July 5 2026, all within 90 days; dLLM bidirectional attention invalidates KV cache on every token commit; Fast-dLLM: block-wise approx KV + confidence-aware parallel decoding 2-3.6× throughput; HERALD: CPU-GPU cooperative KV offloading 5-10% KV budget 2.47× throughput; DiLaServe: dynamic confidence threshold for SLO attainment +56.6pp with 0.9% quality loss; Sangam: AR serving stack retrofit taxonomy for dLLMs)
• [x] database-frontier/88: SedonaDB 0.4: GPU 레이트레이싱 코어로 공간 조인을 5.93배 빠르게 만드는 방법 (RayBooster · VLDB 2026) — published 2026-08-08 (dynamic fallback: database/geospatial topic; SedonaDB 0.4.0 released June 26 2026, within 90 days; first production database to use GPU ray-tracing RT cores for spatial joins; RayBooster accepted VLDB 2026 Industry Track Ohio State University collaboration; Z-stacking encodes row IDs into unused Z-axis enabling single global BVH for entire geometry batch; SoA layout gives O(1) geometry random access; RelateEngine computes DE-9IM matrix on RT cores via single code path for all geometry type/predicate combinations; 5.93× faster than CPU R-tree, 59% lower cost; consumer RTX GPU outperforms H100 because RT cores absent on data-center compute GPUs; also: Python DataFrame API, R dplyr bindings, Geography type via s2geometry, 26 new spatial SQL functions, GeoParquet improvements, conda-forge packaging, 187 issues resolved)
• [x] ai-frontier/94: DualMap: 분산 LLM 서빙에서 캐시 어피니티와 부하 분산을 동시에 해결하는 이중 해시링 라우팅 (ICLR 2026) — published 2026-08-08 (dynamic fallback: AI systems topic; arXiv:2602.06502 ICLR 2026 May 2026, within 90 days; DualMap distributed LLM serving router achieving both KV cache affinity and load balancing via "power of two choices"; two independent hash functions H1/H2 map each request prefix to two candidate instances, current KV hit rate + queue length + TTFT prediction selects the better one; same prefix always maps to same candidate pair so KV reuse concentrated in 2 instances; different prefixes scatter across different pairs so cluster-wide load evens out; three auxiliary techniques: SLO-aware routing switches to less-busy instance when predicted TTFT exceeds SLO; hotspot-aware rebalancing migrates requests from overloaded instances at configurable threshold (default 1.5×); dual-hash-ring scaling independently updates two rings for minimal KV invalidation during scale-out; 2.25× effective request capacity vs prior SOTA; vLLM engine integration validated)
• [x] database-frontier/89: Valkey 9.1: DB별 ACL·Lua 모듈화·JSON 로그로 Redis 포크의 보안과 운영성을 강화한 방식 — published 2026-08-08 (dynamic fallback: database/in-memory-DB topic; Valkey 9.1.0 released July 2026, 9.1.1 patch late July 2026, within 90 days; Valkey = Redis 7.x fork under Linux Foundation BSD-3 license by AWS/Google/Oracle after Redis 7.4 BSL/RSALv2 license change; db N qualifier in ACL SETUSER: ACL SETUSER tenant_a on >pw ~app: +@all db 0 restricts user to logical DB 0 only, NOPERM on other DBs, db = all DBs (default); Lua module decoupling: lua-enabled no config or CONFIG SET lua-enabled no disables EVAL/EVALSHA/FUNCTION without restart; log-format json: structured JSON output with timestamp/pid/role/level/message fields; INFO tls tls_cert_not_after_seconds field for cert expiry monitoring; tls-reloadcert yes for hot cert reload; SAN URI mTLS via tls-client-cert-uri-san for SPIFFE/SPIRE integration; memory reduction: up to 20% small strings via embstr encoding boundary raise + header compression, ~10% sorted set via listpack serialization improvement; I/O threading throughput +17% pipeline batching improvement; CVE-2026-56684 heap overflow in LMPOP CVSS 8.8 critical, CVE-2026-63639 AUTH timing side-channel CVSS 5.3 medium, both in 9.1.1; Amazon ElastiCache for Valkey 9.1 GA)
• [x] ai-frontier/95: PEEK: 대기 큐를 미리 보고 KV 캐시를 예약하는 LLM 서빙 최적화 — published 2026-08-08 (dynamic fallback: AI systems topic; arXiv:2607.02525 July 2025, within 90 days; PEEK = Predictive Queue-Informed KV Cache Management; incremental Radix Tree over pending queue exposes prefix-sharing clusters before scheduling; dual-walk matching against cache tree for LPM; cluster-aware admission: pioneer first then siblings reuse; eviction hook protects blocks still referenced in queue; multi-lane stride scheduler for fairness; Rust core + ~800 LOC Python shims per engine; SGLang results: 3.0× cache hit, 7.9× TTFT reduction, 6.7× E2E latency, 3.6× throughput; vLLM results: 2.6× cache hit, 7.1× TTFT, 5.5× E2E, 4.5× throughput)
• [x] database-frontier/90: Apache Fluss: 실시간 레이크하우스를 위한 스트리밍 스토리지 — published 2026-08-08 (dynamic fallback: database/streaming-storage topic; Apache Fluss graduated ASF TLP August 6 2026, within 90 days; columnar Arrow-based streaming storage for lakehouse; Log Tables append-only offset-ordered Arrow columnar; PK Tables with dual KV Store (point lookup) + Log Store (CDC changelog); Tiering Service auto-offloads to Iceberg/Paimon/Hudi; 10× faster streaming reads via column pruning vs row-oriented; Flink native source/sink/lookup-join; Spark Structured Streaming source; Doris KV Store direct scan; developed at Alibaba Cloud by Jark Wu (Flink PMC); ASF TLP graduation ensures multi-org governance and long-term stability)

• [x] ai-frontier/96: Bullet: 웨이브 양자화와 Attention 병목을 진단하고 동적 SM 분할로 LLM 서빙 처리량을 1.55배 높이는 방법 (ASPLOS 2026) — published 2026-08-09 (dynamic fallback: AI systems topic; arXiv:2504.19516, ASPLOS 2026, Sun Yat-sen University; two root causes of GPU underutilization: wave quantization in prefill and attention bottleneck; Performance Estimator profile-augmented analytical model; SM partitioning via libsmctrl with dynamic ratio; temporal orchestration recycles idle SMs to decode during tail waves; 1.26× average throughput up to 1.55×; 86.2% SM active cycles vs SGLang 75% (+11.2pp); 13.5× prefill latency reduction vs SGLang-1024; 1.86× end-to-end speedup; evaluated on A800/H100/H20 with Llama3.1-70B, Qwen3-235B-A22B, Splitwise dataset; built on SGLang; CUDA ≤ 12.6, Python ≥ 3.12.9)
• [x] database-frontier/91: ClickHouse 26.7: QBit 벡터 검색 정밀도 제어·조인 최적화·EXPLAIN ANALYZE로 OLAP 운영 기준을 갱신한 방법 — published 2026-08-09 (dynamic fallback: database topic; ClickHouse 26.7.0 released July 23 2026, within 90 days; 61 new features, 112 performance optimizations, 329 bug fixes; QBit 4 improvements: Int8 quantization direct storage, strided storage 10/16 bit planes 37.5% read reduction, randomized Hadamard rotations for better quantization quality, quantization codecs for 2-stage search with CODEC_AUTO; JOIN 3 improvements: probe-side granule skipping via primary-key/skip-index analysis, hash table compaction 8-byte index-based value ALL joins as compact as ANY median 12% faster 15% less memory INNER UInt64 21% faster 38% less memory, automatic join ordering 313× speedup 339× fewer rows 592× less peak memory; EXPLAIN ANALYZE runs query and shows per-operator rows_read/rows_output/elapsed_us/blocks_read; position-aware phrase search experimental allow_experimental_text_index_phrase_search stores token positions in text indexes; GROUP BY ORDER BY LIMIT heap-based top-N avoids full sort; unified URL access)
• [x] ai-frontier/97: 소비자 GPU로 클라우드 SLO 달성하기: 로컬 MoE 추론의 스트림-로딩 CPU–GPU 하이브리드 설계 (OSDI '26) — published 2026-08-09 (dynamic fallback: AI systems topic; arXiv:2606.10493, OSDI 2026 accepted, Tsinghua University + Xingyun Integrated Circuits; Stream-Loading Prefill (SLP) 3-thread pipeline overlaps DRAM→GPU expert weight loading with GPU compute at sub-layer granularity; DSLP = Distributed SLP across 2× RTX 5090 with SmallEP expert sharding; AVX-512 FP8 GEMV CPU-side decode on AMD EPYC dual-socket; node-level PD disaggregation for concurrency; RTX 5090 single GPU: 1200 tok/s prefill, 32K prompt <30s TTFT; 2× RTX 5090 DSLP+SmallEP: 1800 tok/s, 45K prompt <30s; CPU GEMV: 4–5× decode latency reduction; DeepSeek-V3/R1 671B model with 400GB+ weights on consumer hardware)
• [x] database-frontier/92: Elasticsearch 9.5: columnar 인덱스 모드와 logsdb_columnar로 로그 스토리지를 줄이는 방법 — published 2026-08-09 (dynamic fallback: database/search-engine topic; Elasticsearch 9.5.0 released 2026-08-04, within 90 days; columnar index mode Tech Preview: doc values only, no inverted index, no BKD tree, append-only analytics workloads; logsdb_columnar mode: columnar storage + logsdb defaults @timestamp/host.name sort + inverted index on message field only; VectorDB index mode auto-calibration: DiskBBQ quantization depth 4–8 bit auto-select, preconditioning auto-enable, oversampling auto-adjust based on vector distribution; Anthropic Claude API observability integration added alongside OpenAI/Bedrock/Vertex; rolling upgrade from 9.4; columnar modes GA planned for 9.6)
• [x] ai-frontier/98: 램프를 평탄하게, 피크가 아니라: Chunked Prefill이 LLM 추론 전력 역학과 그리드 예비 용량에 미치는 영향 — published 2026-08-09 (dynamic fallback: AI systems topic; arXiv:2608.01250, published 2026-08-02, within 90 days; Chunked Prefill (already default in vLLM/SGLang) reduces power ramp rate 7.0–34.6% without changing peak power; up to 42.6% ramp reduction at high whale fraction/size; 20.3–22.7% reduction in fast-ramping grid reserve capacity at representative operating point; studied across 95%–99.9% reliability levels; joint optimization framework for chunk size vs latency SLO)
• [x] database-frontier/93: DBLifeBench: 데이터베이스 전체 생애주기로 LLM을 평가한다 — Text2SQL 특화 훈련의 저주와 DBA 역할의 미래 — published 2026-08-09 (dynamic fallback: database/AI intersection topic; arXiv:2608.03794, published 2026-08-04, within 90 days; first benchmark across 5 DB lifecycle phases: Design, Implementation, Operation, Debugging, Maintenance; Progressive-Text2SQL with Dynamic Reasoning Graph: 67.8% complex query accuracy vs 41.3% direct prompting; "Curse of Specialization": Text2SQL fine-tuned models score 48–50% average across 5 phases vs 63–66% for general-purpose models; open-source vs proprietary gap largest in Operation and Debugging phases)
• [x] ai-frontier/99: InferenceBench: AI 에이전트에게 LLM 서빙 최적화를 맡기면 무슨 일이 일어나는가 — published 2026-08-10 (dynamic fallback: AI engineering topic; arXiv:2607.20468, published July 2026; 4-scenario benchmark TTFT/TPOT/throughput/combined; Claude Sonnet 4.6 best agent at 8.08×; SMAC3 best non-agent optimizer at 10.20–11.53×; 93.9% chose vLLM; median 1 non-default config tried; 65% passed quality+integrity gates; 58.3% passed with ≤1 config change)
• [x] database-frontier/94: Apache Parquet ALP 인코딩: 부동소수점 컬럼을 ZSTD 수준으로 압축하면서 압축 해제를 100배 빠르게 만드는 방법 — published 2026-08-10 (dynamic fallback: data-platform topic; ALP encoding merged to parquet-format spec July 29 2026; encoding value 10 ADAPTIVE_LOSSLESS_FLOATING_POINT; SIGMOD 2024 paper Afroozeh & Boncz CWI; two-path algorithm PseudoDecimal + Front-bit vectorized; FLOAT/DOUBLE support; 7-byte header per page; default vector size 1024; ZSTD-comparable ratio with 1–2 orders of magnitude faster decompression; DuckDB early implementation available)
• [x] ai-frontier/100: AFlex: Attention과 FFN을 분리해 GPU 주파수를 독립 제어하고 LLM 서빙 에너지를 49% 줄이는 방법 — published 2026-08-10 (dynamic fallback: AI systems topic; arXiv:2608.01891 submitted 2026-08-03, within 90 days; AFlex disaggregates LLM serving into 4 pools PA/PF/DA/DF; Global Scheduler ILP solves for frequency assignments per epoch; Local DVFS Controller converts frequency decision to hardware P-state; 49% energy reduction vs disaggregated serving; 48% vs frequency-scaling alone; TTFT+TPOT SLOs maintained; SGLang implementation; A800 GPUs; Qwen3-32B + Mixtral-8×7B validated)
• [x] database-frontier/95: Apache Flink Agents 0.3: 스트리밍 이벤트 드리이브 AI 에이전트에 Exactly-Once 보장·Agent Skills·Mem0 기억을 더한 운영 아키텍처 — published 2026-08-10 (dynamic fallback: data-platform topic; Flink Agents 0.3.0 released 2026-06-19 and 0.3.1 released 2026-07-25, within 90 days; Exactly-Once via Kafka external action state store + Flink checkpointing; Agent Skills catalog system in 0.3.0; Durable Execution Reconciler in 0.3.0; Mem0-based Long-Term Memory in 0.3.0; Flink 2.3 support in 0.3.1; preview status)
• [x] ai-frontier/101: TaiChi: PD 집약과 분리를 통합해 균형 SLO에서 처리량을 77% 높이는 LLM 서빙 아키텍처 — published 2026-08-10 (dynamic fallback: AI systems topic; arXiv:2508.01989 submitted August 2026, within 90 days; TaiChi unifies PD aggregation and disaggregation by differentiating GPU instances into Prefill-Heavy (PH) and Decode-Heavy (DH) pools with three-slider ratio control; Latency Shifting reallocates decode slots to at-risk TTFT requests; Flowing Decode Scheduling controls TPOT at instance level; Length-Aware Prefill Scheduling controls TTFT; 77% goodput improvement under balanced TTFT/TPOT SLOs vs pure aggregation; 43% improvement over pure disaggregation; 19.3% improvement over Splitwise mixed strategy)
• [x] database-frontier/96: 브랜칭 데이터베이스는 아직 에이전트 워크로드를 감당할 수 없다: BranchBench와 Agentic Data Environments — published 2026-08-10 (dynamic fallback: database topic; arXiv:2607.07397 July 8 2026 within 90 days; arXiv:2604.17180 April 19 2026 within 180-day window; BranchBench 5 agentic workflows: agentic SWE, failure reproduction, data curation, MCTS, simulation; Neon CoW O(1) snapshots 25× slower branch creation than Dolt; Dolt Prolly Tree near-instant branch ops 5–4000× slower queries as branches accumulate; Agentic Data Environments broader framework: agents operate over files, processes, APIs, system state simultaneously; neither Neon nor Dolt fully satisfies agentic workload requirements)
• [x] ai-frontier/102: Multi-Stream LLM: 병렬 스트림 계산으로 에이전트의 읽기·생각·행동을 동시에 실행하는 방법 — published 2026-08-11 (dynamic fallback: AI systems topic; arXiv:2605.12460 May 2026 within 90 days; Multi-Stream LLMs enable parallel I/O in LLMs via multiple independent streams with cross-stream attention; each stream has its own KV cache and positional encoding; cross-stream attention allows streams to read each other without merging into single sequence; enables parallel read+think+act in one forward pass; agents no longer blocked waiting for previous step completion; authors: Guinan Su, Yanwu Yang, Xueyan Li, Jonas Geiping; Max Planck Institute / ETH Zurich / Tübingen AI Center / ELLIS; GitHub: seal-rg/streaming)
• [x] database-frontier/97: PostHog의 DuckDB 데이터 웨어하우스: DuckLake·DuckGres·Firecracker 단일 테넌트로 ClickHouse 다음을 구성한 방법 — published 2026-08-11 (dynamic fallback: database/data-platform topic; PostHog blog June 29 2026 within 90 days; single-tenant DuckDB per org in Firecracker MicroVMs; DuckLake SQL-backed Iceberg-compatible catalog on S3; DuckGres Postgres wire protocol server wrapping DuckDB; DuckHog local DuckDB extension for self-hosted orgs; ClickHouse still used for event analytics; Lifecycle Manager handles MicroVM spin-up ~300ms, 60s idle shutdown; separates hot event analytics (ClickHouse) from warehouse queries (DuckDB per tenant))
• [x] ai-frontier/103: Bole: 하이브리드-어텐션 LLM에서 트리 투기 디코딩을 효율적으로 만드는 방법 — published 2026-08-11 (dynamic fallback: AI systems topic; arXiv:2608.01651 submitted 2026-08-03, within 90 days; Bole solves tree speculation for hybrid-attention LLMs combining full Transformer attention and linear recurrent layers (Mamba/SSM); three innovations: Tree-Structured Closed Form enables parallel verification of all draft tree nodes via closed-form linear-attention recurrence product 3.4–7.7× computation reduction; Token-Level Factor Encoding stores per-token factor matrices instead of full recurrent state 82–99% transient memory reduction; Batch-Wide Verification Budget adaptive tree width/depth in SGLang based on real-time acceptance rate; integrated in SGLang; supports Jamba, Falcon Mamba, Zamba2, Mamba-2 architectures)
• [x] database-frontier/98: AlloyDB 2026: BM25 전문 검색·ScaNN 벡터·AI SQL·레이크하우스 페더레이션으로 PostgreSQL에서 하이브리드 데이터 플랫폼을 구성하는 방법 — published 2026-08-11 (dynamic fallback: database/data-platform topic; Google Cloud Next '26 April 22-24 2026 ~110 days within 180-day fallback window; BM25 via pg_textsearch Preview — IDF-weighted ranking over GIN index, Korean tokenizer limited in Preview; ScaNN index Preview — 10B vector scale 6× faster than HNSW per Google claim; AI SQL functions Preview — analyze_sentiment, summarize, if, rank, generate, forecast via Vertex AI Gemini; Lakehouse Federation BigQuery GA + Apache Iceberg Preview — push-down query no data movement; Columnar Engine HNSW+Columnar 4× faster vector queries GA; PostgreSQL 18 GA; Database Insights MCP server GA)
• [x] ai-frontier/104: SPORK: 에이전트 LLM의 도구 대기 시간을 투기적 포킹으로 18% 줄이는 방법 — published 2026-08-11 (dynamic fallback: AI engineering topic; arXiv:2607.03333 July 3 2026 within 90 days; training-free controller; prefix-cache fork predicts tool name 74.6-99.6% accuracy; confidence gate; partial-token accept; 18% GAIA P95 reduction 131.9s→108.1s Qwen3-32B; 4B-32B Dense+MoE; Tsinghua University + Meituan)
• [x] database-frontier/99: Turso v0.7: SQLite를 Rust로 다시 써서 동시 쓰기·전문 검색·벡터를 하나의 엔진에 통합한 방법 — published 2026-08-11 (dynamic fallback: database topic; Turso v0.7.0 SQLite-compatible Rust rewrite; MVCC BEGIN CONCURRENT Hekaton-style; concurrent writes early preview August 3 2026 within 90 days; async io_uring completion-based I/O; Tantivy BM25 FTS; built-in FLOAT32[N] vector search; PostgreSQL sequences with sequence_watermark(); ICU collations; beta warning dropped)
• [x] ai-frontier/105: FASER: 연속 배치 내 투기 디코딩의 단계 직렬화를 공간 분할 중첩으로 해결하는 방법 — published 2026-08-12 (dynamic fallback: AI systems topic; FASER arXiv:2604.20503 April 2026 within 90 days; per-request γ adaptation: each request tracks own acceptance rate α̂ and adjusts γ per chunk boundary independently; early token pruning: rejected suffix tokens evicted at chunk boundary before verify phase CUDA kernel; spatial SM multiplexing: draft and verify kernels partitioned to disjoint SM groups via libsmctrl eliminating serial wait; 53% throughput improvement; 1.92× latency improvement; vLLM prototype; Wenyan Chen, Chengzhi Lu, Yanying Lin, Dmitrii Ustiugov)
• [x] database-frontier/100: SQL Server 2025: DiskANN 벡터 인덱스와 VECTOR 타입으로 관계형 DB가 RAG 스토어가 되는 방법 — published 2026-08-12 (dynamic fallback: database topic; SQL Server 2025 announced May 2025 preview RTM November 2025 within 180-day fallback window; VECTOR(N) type GA float32 1-1998 dims optimized binary storage; VECTOR_DISTANCE() GA cosine/euclidean/dot full table scan exact search; CREATE VECTOR INDEX DiskANN Preview PREVIEW_FEATURES=1 required; VECTOR_SEARCH() Preview TVF ANN search; float16 Preview; DiskANN PQ compressed vectors in DRAM full graph on SSD 10-30× less DRAM vs HNSW; TOP_N 3-5× oversampling recommended for recall; PREVIEW_FEATURES database-scoped)
• [x] ai-frontier/106: Databricks Unity AI Gateway GA: 에이전트·모델·MCP를 하나의 거버넌스 레이어로 통제하는 방법 — published 2026-08-12 (dynamic fallback: AI engineering topic; Unity AI Gateway GA 2026-08-04; announced at Data + AI Summit 2026 2026-06-16; built on Unity Catalog; Agent Registry for all agentic assets; Smart Routing by complexity/cost/quality; Hard Spend Caps real-time budget enforcement; PII guardrails and prompt injection detection; MCP governance registry with managed connectors GitHub/Jira/Slack/Google Drive; governed audit logs for all AI interactions; within 90 days)
• [x] database-frontier/101: Apache Spark 4.2: Auto CDC·Metric Views·지리 타입으로 데이터 엔지니어링의 수작업을 엔진으로 옮긴 방법 — published 2026-08-12 (dynamic fallback: data-platform topic; Apache Spark 4.2.0 released July 2026 within 90 days; powers Databricks Runtime 19; CHANGES clause for batch and streaming CDC from Data Source V2; apply_changes() Python API for Auto CDC SCD Type 1 upsert; Metric Views native semantic layer with MEASURES+DIMENSIONS in Spark SQL; GEOMETRY/GEOGRAPHY native geospatial types with ST_* OGC functions; Arrow-first PySpark UDFs and IPC default; Real-Time Mode PySpark streaming; Spark Connect gRPC+Arrow remote client maturity; INSERT INTO REPLACE ON upsert syntax; is_valid_variant() for VARIANT validation; Spark Web UI Bootstrap 5 dark mode)
• [x] ai-frontier/107: SwiftCache: 멀티턴 대화 LLM 서빙에서 NVLink로 이종 GPU 메모리를 공유해 컨텍스트 길이를 4배 늘리는 방법 — published 2026-08-12 (dynamic fallback: AI systems topic; arXiv:2606.16135 June 15 2026 within 90 days; collaborative inference system enabling heterogeneous co-located models to share idle GPU memory via NVLink instead of PCIe offload; Layer Stream Cache keeps only active layer KV in local HBM; Elastic Cache dynamically adjusts donated pool; P99 TTFT reduced up to 69%; max context length extended up to 3.98×; single multi-GPU server scope limited by NVLink range)
• [x] database-frontier/102: ClickHouse 26.7: QBit 스트라이드·위치 인식 구문 검색·JOIN 순서 최적화로 OLAP 엔진이 AI 분석에 더 가까워진 방법 — published 2026-08-12 (dynamic fallback: database topic; ClickHouse 26.7.0 released 2026-08-06 within 90 days; 61 new features, 112 performance optimizations, 329 bug fixes; QBit: Int8 quantization, stride parameter for per-group dimension I/O, bit plane control 37.5% I/O reduction at 10/16 planes with 99%+ recall, randomized Hadamard rotation for uniform quantization, quantization codecs for two-stage search; phrase search: positional text indexes experimental allow_experimental_text_index_phrase_search; JOIN: new join-ordering algorithm, compact hash tables, complex condition pruning; fused GROUP BY ORDER BY LIMIT pipeline; /play Web UI tabs with persisted state URL integration schema-aware autocompletion column visualizations; EXPLAIN ANALYZE improvements; JIT regex; Apple Silicon improvements)
• [x] ai-frontier/108: MORI: 에이전트 LLM의 도구 호출 대기 시간을 이용한 KV 캐시 오프로딩 — published 2026-08-13 (dynamic fallback: AI systems topic; arXiv:2606.00866 May 30 2026 within 90 days; MORI exploits tool-call idle windows in agentic LLM inference to proactively offload KV cache from GPU HBM to CPU DRAM; idleness spectrum predictor classifies tool calls by expected idle duration; dynamic boundary shifting adapts offload threshold to real-time HBM pressure; admission control compares expected completion time with PCIe reload latency; 20-71% throughput improvement; 18-43% TTFT reduction; 2.1× active batch size on same HBM; validated on Claude Code/Cursor/GitHub Copilot workflow traces; vLLM v0.4 implementation; UC Berkeley+Stanford+Renmin+Georgia Tech)
• [x] database-frontier/103: PLOP: SQL 쿼리 플랜에서 LLM 시맨틱 연산자 배치를 최적화하는 방법 — published 2026-08-13 (dynamic fallback: database topic; PLOP arXiv:2604.09944 April 2026 within 90 days; SemCEB arXiv:2606.23081 June 2026 within 90 days; PLOP uses bitmask DP to find optimal placement of LLM-based semantic operators (sem_filter, sem_join, sem_agg) relative to traditional SQL operators in hybrid query plans; cost function = Σ rows_in × unit_cost per operator; cardinality estimator samples 5% rows to predict semantic selectivity; up to 1.5× speed improvement 4.3× cost reduction vs naive placement; SemCEB 44 queries 5 schemas benchmark for semantic operator cardinality estimation; LOTUS framework integration; DP exploration O(2^n × n) completes in milliseconds for n≤8)
• [x] ai-frontier/109: LiveMem: 컨텍스트가 교체되어도 기억을 유지하는 장기 실행 LLM 추론 설계 — published 2026-08-13 (dynamic fallback: AI systems topic; arXiv:2608.02515 August 2026 within 90 days; LiveMem adds fixed-capacity persistent memory state module to pretrained full-attention LLMs; dual-path: main attention path for current active KV cache, memory state path M for cross-context continuity; state update rule integrates expiring KV at context turnover; memory state persists across arbitrary context replacements at constant size; top performance on LongMemEval benchmark; maintains accuracy when supporting evidence removed from current context; targets long coding sessions, multi-turn customer service, multi-cycle research agents; preprint stage as of 2026-08-13; UC Santa Barbara / 6 authors)
• [x] database-frontier/104: PostgreSQL 19 Beta 3: 28개 보안 취약점 패치와 GA 직전 운영자 점검 체크리스트 — published 2026-08-13 (dynamic fallback: database topic; PostgreSQL 19 Beta 3 released 2026-08-13 alongside 18.6/17.11/16.15/15.19/14.24; 28 CVEs across all supported versions; critical: CVE-2026-6471 logical decoding arbitrary dlopen; high: CVE-2026-14663 pgcrypto silent plaintext, CVE-2026-14664 regexp heap buffer overflow, CVE-2026-14666 RLS cache bypass; medium: CVE-2026-6464 psql COPY FROM STDIN injection, CVE-2026-14662 tsvector undersize, CVE-2026-14668 ctid type confusion, CVE-2026-6469 extended stats ownership loss, CVE-2026-6470 type USAGE privilege missing; 110+ bug fixes; confirmed PG19 features: SQL/PGQ GRAPH_TABLE, REPACK CONCURRENTLY, parallel autovacuum, GROUP BY ALL, lz4 TOAST default, JIT off by default, sequence logical replication, pg_plan_advice bundled, virtual generated columns; GA expected September/October 2026)

Open Questions

• 각 시리즈 최종 편 수(현재 1~10 가정).
• [x] ai-frontier/110: Gimbal: MoE LLM 서빙에서 DP 스케줄링과 전문가 배치를 동시에 협조시키는 방법 — published 2026-08-13 (dynamic fallback: AI systems topic; arXiv:2606.15177 June 2026 within 90 days; Gimbal coordinated cross-level scheduling for MoE LLM serving; DP-engine scheduler uses 4 fine-grained backend pressure signals: KV cache usage, remaining prefill work, queue pressure, MoE expert pressure; SJF queue ordering with aging reduces head-of-line blocking without output-length prediction; source-DP-to-expert activation statistics with MINLP-guided heuristic for expert placement balancing load, communication cost, migration stability; TTFT -42.9%, TPOT -33.3%, throughput +3.0% vs vLLM at high load)
• [x] database-frontier/105: Apache DataFusion 54: LATERAL 조인·스칼라 서브쿼리 물리 실행·CDC Parquet로 Rust 쿼리 엔진이 넓어진 방법 — published 2026-08-13 (dynamic fallback: data-platform topic; DataFusion 54.0.0 released 2026-06-12 within 90 days; 11 weeks development 740 commits 139 contributors record; physical uncorrelated scalar subquery operator replaces JOIN rewrite; Sort-Merge Join semi/anti/mark joins use per-row bitset not materialized pairs 20-50x faster near-unique joins; DynComparator join-key comparison 12% microbenchmark improvement 5% TPC-H improvement; LATERAL join support; SQL lambda functions array_transform; arrow-avro crate Avro reader replacement; Content-Defined Chunking CDC Parquet writer for better deduplication and incremental storage)
• [x] ai-frontier/111: KAIROS: 에이전트 AI 서빙에서 컨텍스트 인식 GPU 전력 최적화 — published 2026-08-14 (dynamic fallback: AI systems topic; arXiv:2604.16682 April 2026 within 90 days; KAIROS uses agent context as first-class control signal to jointly manage GPU frequency via DVFS, per-instance concurrency, and multi-instance request placement; three context signals: context length growth rate, GPU memory pressure, agent progress stage; key insight: lowering GPU frequency without context awareness causes memory thrashing when KV cache is full; 27% average GPU power reduction single instance up to 39.8% maximum, 46.3% multi-instance; SLO maintained at P5 throughput 20 tokens/s; evaluated on NVIDIA H100 + vLLM with mini-swe-agent and Terminus-2 on SWE-bench Verified, DABStep, Terminal-Bench 2.0; authors: Yichao Yuan, Mosharaf Chowdhury, Nishil Talati)
• [x] database-frontier/106: SingleStore 9.0: BM25 전문 검색 강화, 쓰기 가능 뷰, Projections으로 HTAP 플랫폼이 넓어진 방법 — published 2026-08-14 (dynamic fallback: database topic; SingleStore 9.0 production-recommended release 2026; BM25 GLOBAL function cross-partition scoring for accurate full-text ranking in distributed tables; custom analyzers tokenizer+token-filter+char-filter+custom stop words; nori Korean analyzer customization; boolean/boost/proximity/phrase search; multi-column FTS queries; full-text pushdown optimization; writable views UPDATE/INSERT/DELETE on views for Oracle/MySQL migration compatibility; vector index merger cross-segment consolidation; vector index cache memory limit; fallback scan mechanism for vector+filter queries; INDEX information schema view; Projections secondary sort+shard keys for 2-3x range filter GROUP BY performance; docs.singlestore.com/db/v9.1/release-notes/)
• [x] ai-frontier/112: vLLM 0.27.0: NVIDIA Rubin·FlashAttention4·Model Runner V2로 LLM 서빙 생태계를 넓힌 방법 — published 2026-08-14 (dynamic fallback: AI systems topic; vLLM 0.27.0 released 2026-08-09 within 90 days; 561 commits 242 contributors; NVIDIA Rubin sm_107 NVLink support; AMD ROCm gfx1250 Radeon RX 9000 series support; DeepSeek-V4 kernel ~2× speedup over V3 kernel; FlashAttention4 FP8 precision + headdim-256 on SM100 H100/H200/Blackwell; Model Runner V2 V1 extends beyond decoder-only to encoder models, embeddings, classification; Rust gRPC frontend V1 experimental; PyTorch 2.13.0; Kimi K3 and Qwen3.5 MoE model support added)
• [x] database-frontier/107: Milvus 3.0: 레이크 네이티브 벡터 데이터베이스로의 전환 — External Collections와 Woodpecker WAL — published 2026-08-14 (dynamic fallback: data-platform topic; Milvus 3.0.0 released 2026-07-29 within 90 days; LF AI & Data graduated project; External Collections over object storage Lance/Iceberg/Parquet/Vortex without data copy; Woodpecker built-in WAL deployable as standalone service replacing mandatory Pulsar/Kafka; SINDI-based sparse vector index redesign for hybrid dense+sparse search; StructArray nested struct field type support; faceted search aggregated results alongside vector search; online schema evolution add/backfill/drop without collection rebuild; Spark DataSource V2 native connector for Spark/Databricks/EMR; FAISS passthrough and TEXT index extensions; Apache 2.0 license)
• [x] ai-frontier/113: IntentKV: 에이전트 LLM의 멀티턴 KV 캐시 병목을 교차 턴 의도로 정밀하게 줄이는 방법 — published 2026-08-14 (dynamic fallback: AI systems topic; arXiv:2606.09916 June 2026 within 90 days; IntentKV cross-turn intent-aware KV cache pruning for multi-turn agent inference; frozen base LLM + lightweight QueryMemory module; session-level intent vector updated per turn; memory-attention scoring identifies relevant tokens for current query; zero-initialized residual cross-attention head; slot-map eviction compatible with prefix cache; 8k KV budget: Qwen3-8B 23.9% average token reduction, Qwen2.5-14B 30.7%; BCP top-100 worst case: 77.8% token reduction, 92.6% KV read reduction; 20.5k vs 92.3k max request tokens; Shanghai Jiao Tong University Junjie Li, Jiong Lou, Jie Li)
• [x] database-frontier/108: Apache Polaris 1.7.0: 멱등 쓰기·자격증명 벤딩 강화·고아 파일 정리로 Lakehouse 카탈로그 신뢰성을 높이는 방법 — published 2026-08-14 (dynamic fallback: database topic; Apache Polaris 1.7.0 released 2026-08-02 within 90 days; opt-in idempotent writes for createTable/updateTable: polaris.idempotency.enabled=true, Idempotency-Key header, entity-property approach, TTL PT5M; PR #5086 removed IdempotencyStore and idempotency_records table; SigV4 sessionPolicy field for IAM session policy JSON in STS AssumeRole; GCS principal attribution via Workload Identity Federation GCS_PRINCIPAL_ATTRIBUTION_ENABLED; UTF-8 percent-encoding for table storage locations replaces JVM default charset; ManifestFileCleanupTaskHandler handles Iceberg v2 delete manifests; async task retry on transient false returns; semantic model API beta; breaking: JDBC schema v5 events.catalog_id nullable NULL replaces __realm__; bootstrap --schema-version option removed; MaintenanceService.performMaintenance requires OptionalLong overrideRunId; tagged by JB Onofré at commit 4ac2f05)
• [x] ai-frontier/114: RLVR이 LLM 추론 경계를 실제로 넓히는 방법: GRPO의 내재적 보상 메커니즘과 CoT-Pass@K — published 2026-08-15 (dynamic fallback: AI systems topic; arXiv:2506.14245 June 14 2026 within 90 days; RLVR with verifiable rewards implicitly incentivizes correct reasoning in base LLMs; GRPO Group Relative Policy Optimization advantage A_i=(r_i-mean)/std from G samples without critic network; CoT-Pass@K new metric measuring both final answer and intermediate reasoning correctness; binary reward still improves reasoning paths not just answer probability; effect appears in early training when base LLM has sufficient capability; spurious guess paths suppressed; DeepSeekMath GSM8K 82.9→88.2 MATH 46.8→51.7; DeepSeek-R1 Mistral Magistral Qwen3 training recipe; Zhu et al. Shanghai AI Lab)
• [x] database-frontier/109: Apache Arrow ADBC 24: 컬럼형 DB 연결 표준의 성숙과 드라이버 생태계 재편 — published 2026-08-15 (dynamic fallback: database topic; ADBC 24 released 2026-07-28 within 90 days; Arrow Database Connectivity columnar-native DB connection API replacing JDBC/ODBC row-wise transfer; execute_query returns Arrow RecordBatch stream; execute_partitions for parallel reads; ingest for bulk Arrow batch insert; C Data Interface zero-copy cross-language Arrow buffer sharing; Arrow Flight SQL wire protocol client-server; Driver Foundry split: DataFusion BigQuery Databricks Snowflake moved out of Apache Arrow repo; Arrow-maintained drivers: PostgreSQL SQLite FlightSQL; Java close exception type narrowed from Exception; PostgreSQL lazy transaction start; Python package restructuring; 57 issues 142 PRs in this release)
• [x] ai-frontier/115: JouleShare: 배치 LLM 서빙의 요청별 에너지 귀속 문제를 Shapley 값으로 해결하는 방법 — published 2026-08-15 (dynamic fallback: AI systems topic; arXiv:2608.00026 July 2026 within 90 days; JouleShare framework for request-level energy attribution in batched LLM serving using Shapley values; offline harness replays 2^N request subsets under vLLM with GPU power telemetry to compute exact coalition energy function v(S); Shapley axioms efficiency/symmetry/null-player/linearity ensure mathematically fair attribution; token-proportional baseline L1 error 0.440 static 0.458 continuous batching; JCalib lightweight regression model trained on (request features, Shapley energy) pairs; JCalib features: input token count, output token count, concurrent request count, total batch tokens; JCalib L1 error 0.116 static 0.177 continuous; 73.6% L1 reduction static 61.4% continuous; authors Qi Luo Kunlin Li Ziwen Wang Dongsheng Wang Yun Chen; applications: sustainability reporting Scope 1-3 emissions, multitenant chargeback, workload energy profiling; distinct from AFlex/KAIROS optimization approaches)
• [x] database-frontier/110: PostgreSQL 19 Beta 3 프리뷰: REPACK CONCURRENTLY·SQL/PGQ 그래프 쿼리·논리 복제 개선이 바꾸는 것들 — published 2026-08-15 (dynamic fallback: database topic; PostgreSQL 19 Beta 3 released 2026-08-13 within 90 days; REPACK CONCURRENTLY: combines VACUUM FULL + CLUSTER non-blocking brings pg_repack functionality into core; SQL/PGQ ISO SQL:2023 Property Graph Queries: GRAPH_TABLE function CREATE PROPERTY GRAPH graph queries on existing relational tables no data migration; extended AIO framework automatic I/O worker scaling from PG18 base; pg_plan_advice + pg_stash_advice bundled extensions for persistent query planner hint storage; NOT IN → ANTI JOIN automatic conversion when NOT NULL provable; LZ4 default TOAST compression replacing pglz; logical replication: sequence values now replicated preventing post-failover key conflicts; wal_level change no restart when already at replica+; PG14 EOL notice issued; GA expected September 2026)
• [x] ai-frontier/116: DSpark: 반자동회귀 초안 생성과 신뢰도 스케줄링으로 LLM 추론을 60–85% 가속하는 방법 — published 2026-08-15 (dynamic fallback: AI engineering topic; arXiv:2607.05147 July 6 2026 within 90 days; DSpark confidence-scheduled speculative decoding with semi-autoregressive generation; semi-autoregressive Markov head rank-256 low-rank factorization adds per-position bias conditioned on previous accepted token; fixes suffix decay in parallel draft models; confidence head estimates prefix survival probability per position; hardware-aware scheduler routes verification budget to high-expected-return tokens; 60-85% per-user generation speed on DeepSeek-V4-Flash, 57-78% on V4-Pro; +26.7-30.9% accepted length vs EAGLE-3, +16.3-18.4% vs DFlash on Qwen3; 0.2-1.3% latency overhead; lossless: preserves target distribution; DeepSeek AI; attaches to existing V4 checkpoints without model weight changes)
• [x] database-frontier/111: MySQL 26.7: 캘린더 버전 출시·Thread Pool 오픈소스화·포스트 양자 암호화·Change Stream Applier — published 2026-08-15 (dynamic fallback: database topic; MySQL 26.7.0 GA released 2026-07-28 within 90 days; first calendar-versioned MySQL Innovation release YY.M.P format replacing SemVer for Innovation track; next release 26.10.0 October 2026; Thread Pool plugin Community Edition GA formerly Enterprise-only; thread_pool_max_unused_threads default 2→32; Change Stream Applier new per-channel opt-in multi-threaded replication SQL applier modular scheduling; Group Replication default communication stack XCOM→MYSQL uses standard MySQL connection security; post-quantum cryptography ML-DSA-44/65/87 with OpenSSL 3.5 TLS 1.3 hybrid groups all connection channels; InnoDB undo tablespace truncation progress stored in undo header replacing local log files; MySQL 9.7.2 LTS and 8.4.11 LTS also released same day)
• [x] ai-frontier/117: SpecBox: LLM 생성 스트림으로 MCP 샌드박스 콜드스타트를 예측해 없애는 방법 — published 2026-08-16 (dynamic fallback: AI systems topic; arXiv:2607.23933 July 27 2026 within 90 days; SpecBox speculative sandbox scheduling for MCP-connected LLM agent serving; intent-aware prewarming analyzes streaming token output to detect upcoming tool call before use_tool is fully emitted, overlaps sandbox startup with inference; stochastic prefetching builds sandbox dependency graph from historical traces to probabilistically pre-warm next tools; semantic result caching by embedding similarity not exact match; zero-copy shared-memory out-of-band transport; P99 end-to-end latency 2.9x reduction; peak memory 45.9% reduction vs warm-standby; evaluated on MCPBench 200 multi-turn trajectories 32 MCP tool servers; AgentScope prototype integration; Beihang University / University of Leeds / University of Sydney)
• [x] database-frontier/112: OpenSearch 3.8: Base64 벡터 수집·Radial Search 재설계·MCP Flow Agent로 k-NN 운영 비용을 줄이는 방법 — published 2026-08-16 (dynamic fallback: database topic; OpenSearch 3.8.0 released 2026-08-05 within 90 days; Base64 knn_vector ingestion: 74% network payload reduction, 4.16x bulk ingest throughput, 83% median latency reduction vs JSON array; Radial Search HNSW graph traversal redesign: distance-based bounds per layer, 2.1x throughput, 45% lower latency, recall 0.85→0.97; MCP support added to Flow Agent and Conversational Flow Agent completing all 4 agent architectures; list-tools REST API and connector-level tool description overrides; gRPC-over-HTTP/2 for ML prediction streaming; Amazon Linux 2 deprecation announced, EOL 2026-06-30, actual removal in future version)
• [x] ai-frontier/118: vLLM Decode Context Parallelism: 텐서 병렬성의 한계를 넘어 KV 캐시를 시퀀스 차원으로 샤딩하는 방법 — published 2026-08-16 (dynamic fallback: AI systems topic; vLLM blog 2026-08-07 within 90 days; DCP shards KV cache along sequence dimension so each GPU stores 1/N of every request's KV; addresses TP limitation with GQA/MLA models where KV heads << GPUs; decode flow AllGather Q → local attention on KV slice → AllGather+ReduceScatter; 8×B200 Kimi K2.6 MLA benchmark: TP-only ~1,863 tok/s/GPU ceiling at concurrency 64 OOM vs DCP 6,091 tok/s/GPU at concurrency 512 82% KV usage; 3.3× throughput improvement; bimodal long-context agentic workload: 53% requests 64K+ tokens; CLI: --dcp <size>; supports GQA and MLA; MLA load-time Q projection replication skips AllGather Q at decode; cp_kv_cache_interleave_size default 1)
• [x] database-frontier/113: Redpanda 26.1: 하나의 클러스터로 네 가지 스트리밍 모드를 선택하는 Adaptable Streaming Engine — published 2026-08-16 (dynamic fallback: data-platform topic; announced 2026-03-31, 138 days ago, within 180-day fallback window; no equivalent architectural departure in streaming platforms within 90 days; Adaptable Streaming Engine 4 modes: Write Caching ultra-low latency in-memory acks no disk write; Tiered Storage local NVMe first then async S3/GCS offload balanced; Iceberg Topics Tiered Storage + Parquet Iceberg table materialization zero-ETL analytics; Cloud Topics pass-through write payload direct to S3/GCS metadata-only on local disk 90% cross-AZ network cost savings; Kafka protocol compatible topic-level config switches modes; write.caching / redpanda.remote.write / redpanda.iceberg.mode / redpanda.cloud.storage.enabled; Cloud Topics suited for observability AI training feeds; Iceberg Topics suited for ML feature stores real-time analytics)
• [x] ai-frontier/119: FlashAttention-4: Blackwell를 위한 알고리즘-커널 공동 설계 — published 2026-08-16 (dynamic fallback: AI systems topic; arXiv:2603.05451 March 5 2026, 164 days ago, within 180-day fallback window; vLLM 0.27.0 August 9 2026 integration within 90 days provides fresh operational relevance; FA4 co-designs algorithm and kernel for NVIDIA Blackwell asymmetric scaling where tensor cores 2× but softmax/shared-memory unchanged; TMEM 256KB per SM on-chip tensor-core-wired scratchpad stores partial softmax statistics avoiding shared-memory bottleneck; UMMA 128×256×16 tile 2-CTA 256×256×16 cooperative MMA spans two CTAs TMEM; TMA async data movement global HBM to shared/tensor memory; software exp() emulation via exp2 + minimax polynomial to bypass SFU bottleneck; conditional softmax rescaling skips rescale when max-change below threshold saving ~30% softmax ops; 4-stage async pipeline TMA-load UMMA-1 SW-exp UMMA-2 overlapping all stages; B200 SM100 target; B300 SM103 Blackwell Ultra planned; ~940 TFLOPS on B200 32K seq vs 310 TFLOPS on H100 FA3; vLLM 0.27.0 auto-selects FA4 on Blackwell; code at Dao-AILab/flash-attention flash_attn/cute/)
• [x] database-frontier/114: Apache Cassandra 6.0 Accord: 리더 없는 합의 프로토콜로 ACID 트랜잭션을 구현하는 방법 — published 2026-08-16 (dynamic fallback: database topic; Cassandra 6.0-alpha1 released 2026-05-01, 107 days ago, within 180-day fallback window; CEP-15 Accord leaderless consensus any-node coordinator no leader election; fast path one round-trip PreAccept+AcceptOK+Commit using HLC timestamps; slow path Paxos-shaped fallback under contention or partial failure; reorder buffer per-node buffers timestamps for global ordering across geo-distributed partitions; fast-path electorates pre-configured sub-quorums 3f+1 for fast path 2f+1 for slow path; CEP-21 CMS replaces Gossip metadata with linearized Accord transaction log for consistent cluster topology; Accord off by default operators upgrade without enabling; no mixed-version cluster during Accord activation; Strict Serializable isolation; BEGIN TRANSACTION...COMMIT TRANSACTION CQL block syntax with LET bindings and IF conditional writes; counter columns excluded from Accord transactions; alpha1 GA timeline not confirmed)
• [x] ai-frontier/120: KV 캐시를 살 수 있을까? Prefill CDN으로 LLM 추론 비용을 50배 줄이는 방법 — published 2026-08-17 (dynamic fallback: AI systems topic; arXiv:2606.13361 "Can I Buy Your KV Cache?" June 11 2026 within 90 days; Prefill CDN concept publishers precompute KV caches consumers download and skip prefill; KV cache is deterministic function of text+model+precision enabling safe precomputation and sharing; Qwen3-4B: cache reuse 9-50x cheaper compute than prefill; 3774-token document 80M agents: re-prefill ~$1.5M vs reuse ~$0.03M 49.7x savings; provider economics 0.1x cache-read tariff 10x discount to users ~50x actual compute savings; challenges: large size (hundreds MB to 1+ GB nearly incompressible) model-bound (one model/precision's cache useless to another); related: TokenDance arXiv:2604.03143 April 2026 multi-agent collective KV cache sharing; connection to existing prompt caching APIs (Anthropic 0.1x rate) and vLLM KV offload; open questions: break-even length per model storage+network cost analysis not yet published for large models 70B+)
• [x] database-frontier/115: DuckDB v2.0 Cyanoptera 프리뷰: 서버·트리거·비동기 I/O로 임베디드 OLAP의 경계를 다시 그은 방법 — published 2026-08-17 (dynamic fallback: database topic; DuckDB v2.0 "Cyanoptera" preview published 2026-08-17 today within 90 days; fall 2026 GA planned; Quack protocol graduates stable HTTP/2 internal vector block serialization 1 RTT per query CONNECT ATTACH syntax; triggers: BEFORE/AFTER INSERT/UPDATE/DELETE FOR EACH ROW/STATEMENT transition tables REFERENCING OLD/NEW TABLE multiple triggers per event RETURNING DROP TRIGGER; async I/O GA Parquet+CSV async reads I/O layer independent from query processing layer more parallelism for remote reads; new SQL parser replaces PostgreSQL-derived parser PEG-based extensible; new default storage format v1 .db files incompatible EXPORT DATABASE/IMPORT migration path required; C API reworked language bindings need recompile; ICU library removed icu extension embeds IANA timezone data ~45KB; SQL additions FETCH FIRST ROWS ONLY OVERLAY UNNEST in GROUP BY MERGE/UPDATE FROM semantics; named Cyanoptera after cinnamon teal duck)
• [x] ai-frontier/121: XGrammar-2: TagDispatch·Cross-Grammar Cache로 에이전트 LLM의 동적 구조적 생성을 실현하는 방법 — published 2026-08-17 (dynamic fallback: AI engineering topic; arXiv:2601.04426 January 2026 within 90 days; ACM CAIS '26 May 2026; XGrammar-2 extends original xGrammar PDA+adaptive token mask cache with TagDispatch for runtime grammar switching mid-generation and Cross-Grammar Cache for substructure reuse across grammars; Earley parser replaces original recursive descent for better incremental parsing; JIT compilation; 6x faster grammar compilation vs xGrammar-1; default constrained decoding backend in vLLM 0.6.3+ SGLang 0.4+ TRT-LLM 0.12+ MLC-LLM 0.18+; agentic tool-call schema switching without recompiling per turn)
• [x] database-frontier/116: CloudNativePG 1.29: Image Catalog·ServiceAccount 통합으로 Kubernetes PostgreSQL 운영의 경계를 다시 그은 방법 — published 2026-08-17 (dynamic fallback: database/K8s-operator topic; CloudNativePG 1.29.0 released 2026-03-31 within 90 days; Image Catalog: ImageCatalog ClusterImageCatalog CRDs abstract extension container images per PostgreSQL major version; postgres-extensions-containers project provides pgvector pg_cron pg_partman official images; spec.serviceAccountTemplate enables IRSA Azure Workload Identity EKS GKE annotation injection without Helm post-hooks; podSelectorRefs generates pg_hba.conf host rules from Kubernetes pod selector + namespace selector; PgBouncer Advanced TLS separate client-server TLS settings; CVE-2026-44477 cert bypass fix; PostgreSQL 18 Beta experimental support; CNPG-I Plugin Interface early preview; lease-based failover via Kubernetes Lease no external DCS; Barman Cloud WAL archive + base backup to S3/Azure Blob/GCS)
• [x] ai-frontier/122: Speculate with Memory: 메모리로 LLM 에이전트 추론을 가속하는 방법 — published 2026-08-17 (dynamic fallback: AI engineering topic; arXiv:2607.12236 July 14 2026 within 90 days; memory-augmented speculative agent execution equips drafter with three online memory systems: contrastive transition table tracks action-sequence statistics and applies contrastive learning to downweight mispredicted actions, episodic memory retrieves contextually similar past trajectory segments for in-context augmentation, confusion tracker suppresses repeatedly-wrong actions via exponential moving average confusion score; all three memories update online after each trajectory completion at <3% overhead; 19-39% relative accuracy improvement on action prediction vs stateless baseline; up to 2.5x improvement on observation prediction for repetitive-action-space tasks; lossless: large model retains final authority; orthogonal to token-level speculative decoding; authors Yu Li Qinyuan Ye Prafulla Kumar Choubey Jiaxin Zhang Chien-Sheng Wu)
• [x] database-frontier/117: Databricks Lakehouse//RT와 Omnigent: 에이전트 시대의 실시간 레이크하우스 아키텍처 — published 2026-08-17 (dynamic fallback: data-platform topic; Databricks DAIS June 2026 and August 2026 updates within 90 days; Lakehouse//RT Beta powered by Reyden engine queries Delta Lake and Apache Iceberg tables directly at sub-100ms latency and 12,000 QPS; Reyden fully asynchronous execution model non-blocking high-concurrency design native Delta/Iceberg Parquet parsing; 10ms smaller datasets 100ms larger; up to 16x faster than separate real-time serving stack; no data movement no format conversion no CDC pipelines; Unity Catalog governance applies natively to all queries; Omnigent Apache 2.0 open-source meta-harness github.com/omnigent-ai/omnigent; Runner wraps any agent harness (Claude Code Codex Cursor Pi custom) in sandboxed session with uniform API; Server provides centralized policy management session sharing URL-based real-time collaboration audit log; contextual policies Python stateful tracking full session history allow/deny/transform; YAML custom agents compose multi-step harness-agnostic pipelines; both products Beta in 2H 2026)
• [x] ai-frontier/123: Sarathi-Serve와 청크형 프리필: LLM 서빙에서 TTFT-처리량 균형을 찾는 방법 — published 2026-08-18 (dynamic fallback: AI systems topic; Sarathi-Serve arXiv:2308.16369 OSDI 2024; chunked prefill splits long-context prefill into fixed-size chunks interleaved with decode steps to eliminate HoL blocking caused by compute-heavy prefill iterations stalling decode tokens; stall-free batching: each iteration bounded by chunk_size preventing any single prefill from monopolizing GPU; TTFT-throughput tradeoff: smaller chunks → lower TTFT, larger chunks → higher throughput; optimal chunk size 512-2048 tokens for typical mixed workloads; vLLM 0.25.0+ default enable_chunked_prefill=true; SGLang --chunked-prefill-size; TRT-LLM context_chunking_policy=greedy_longest; complementary to PD disaggregation: chunked prefill reduces HoL per node while PD separation routes prefill/decode to dedicated nodes)
• [x] database-frontier/118: Kafka MirrorMaker 2: 멀티 리전 클러스터 복제와 재해 복구 운영 패턴 — published 2026-08-18 (dynamic fallback: data-platform topic; Kafka 4.3 KRaft environment; MM2 MirrorSourceConnector replicates records with remote topic prefix source.topic; MirrorCheckpointConnector translates consumer group offsets across clusters via offset mapping stored in mm2-offsets.source.internal; MirrorHeartbeatConnector measures replication lag; offset translation API RemoteClusterUtils.translateOffsets() for DR failover; Active-Passive DR: one-directional replication source→target, failover requires re-pointing consumers to target and translating committed offsets; Active-Active bidirectional: both clusters replicate to each other with cycle prevention via topics.exclude pattern source\\..+ replication.factor separation; KRaft context: no ZooKeeper dependency; mm2.properties configuration with sync.group.offsets.enabled=true sync.topic.acls.enabled=true replication.factor=3; common issues: offset translation lag during DR, topic-not-found after failover, cycle protection regex misconfiguration)
• [x] ai-frontier/124: TaiChi: PD 집계와 분리를 통합해 LLM 서빙 굿풋을 77% 높이는 방법 — published 2026-08-18 (dynamic fallback: AI systems topic; arXiv:2508.01989 August 2026; TaiChi unifies PD aggregation and PD disaggregation under a single architecture with differentiated GPU instances: Prefill-Heavy Instances use large chunk sizes and tolerate decode interference for low TTFT, Decode-Heavy Instances minimize phase interference for stable TPOT; three tunable parameters: P-heavy/D-heavy instance ratio, chunk size on each type; collapsing to 100% P-heavy = Sarathi-Serve aggregation, 100% D-heavy = DistServe disaggregation; latency shifting: GPU resources freed from SLO-slack requests are reallocated to at-risk requests, maximising goodput not average latency; two scheduling algorithms: Flow Decode Scheduling regulates per-request decode token throughput to prevent TPOT violations, Length-Aware Prefill Scheduling assigns prefill to instances considering prompt length to prevent short prompts blocking behind long ones; 77% goodput improvement over state-of-the-art baselines under balanced TTFT+TPOT SLOs; authors: Chao Wang, Pengfei Zuo, Zhangyu Chen, Yunkai Liang, Zhou Yu, Ming-Chang Yang; institutions: CUHK + Huawei Cloud + Sun Yat-sen University)
• [x] database-frontier/119: Apache Spark 4.2: CDC 통합·메트릭 뷰·실시간 스트리밍·Arrow 기본값으로 달라진 운영 지형 — published 2026-08-18 (dynamic fallback: data-platform topic; Apache Spark 4.2.0 released 2026-07-14 within 90 days; four major changes: (1) CHANGES clause via DSv2 Changelog mixin interface unifying Delta/Iceberg/Hudi CDC APIs, engine handles dedup/UPDATE conversion/net-change computation; (2) Auto CDC in Spark Declarative Pipelines for SCD Type 1 upserts, APPLY CHANGES INTO deprecated; (3) Metric View with WITH METRICS clause embeds business KPI definitions in engine with path-based name resolution via SET PATH/CURRENT_PATH(); (4) Structured Streaming Real-Time Mode RTM promoted from experimental to officially supported, single-digit ms latency for stateless pipelines, at-least-once delivery, epoch checkpoints, Python Data Sources support; (5) Arrow UDF default-on: spark.sql.execution.pythonUDF.arrow.enabled and spark.sql.execution.arrow.pyspark.enabled both default true, regular @udf now uses Arrow IPC path without signature change; upgrade note: check Arrow type compatibility for existing UDFs; JIRA: SPARK-55668 CDC umbrella, SPARK-55948 DSv2 CDC API)
• [x] ai-frontier/125: Qwen3.8-Max: 2.4조 파라미터 오픈 MoE 플래그십이 에이전트 벤치마크 정점에 선 방법 — published 2026-08-18 (dynamic fallback: AI systems topic; Alibaba Qwen3.8-Max open weights released 2026-08-12 within 90 days; 2.4T total parameters 22B active per token 512 experts 10+1 active MoE; 92 layers hybrid attention: 69 Gated DeltaNet linear-attention O(n) + 23 Full Attention O(n²); 1M token context window; OSWorld-Verified 86.1 vs Claude Fable 5 85.0 vs GPT-5.6 Sol Max 83.2; Apache 2.0 license; vLLM expert parallelism --tensor-parallel-size 8 for 2-node A100 serving; SGLang 0.4.3+ native support; BF16 requires 2×8×A100 for full model; quantized AWQ/GPTQ fits single A100×8 node; Gated DeltaNet uses delta rule memory update ΔW=v⊗k gating vector β; full attention layers placed every 4 linear layers; API available 2026-08-03 open weights 2026-08-12)
• [x] database-frontier/120: Spotify Random Access Parquet: AI 에이전트 시대에 데이터 레이크에서 포인트 쿼리를 제공하는 외부 인덱스 아키텍처 — published 2026-08-18 (dynamic fallback: data-platform topic; Spotify Engineering Blog July 2026 within 90 days; RAP external index stores (lookup_key)→(parquet_file_uri, row_group_id, row_offset) pointer mappings without duplicating data; same Iceberg table serves analytics ML AI-agent and online-service point queries; three indexing strategies: hash bucketing (deterministic file-to-bucket mapping one-hop lookup), co-grouping (sort all rows for same key into same Parquet file), sort-based (binary search on sorted column + index skips the search); append-only index fragments built incrementally on each Iceberg commit; fragments stored in Parquet format attached to Iceberg snapshot metadata; S3 Range GET on (file, row_group_id, offset) to fetch only needed bytes; target latency <100ms cold; index size independent of data size; no data replication vs Redis/DynamoDB copy approach; AI agent agentic reasoning loops issue chains of point lookups creating new demand not served well by full-scan Parquet)
• [x] ai-frontier/126: LLM KV 캐시 다계층 오프로딩: GPU HBM에서 CPU·NVMe·오브젝트 스토리지까지의 설계 패턴 — published 2026-08-19 (dynamic fallback: AI systems topic; KV cache tiering for LLM serving extending GPU HBM via CPU DRAM, NVMe SSD, and object storage; KV cache size formula 2×L×H_kv×D_head×S×B; LLaMA-3 70B 128K context requires ~40GB KV alone; vLLM PagedAttention CPU swap; NIXL-based tiered KV in vLLM 0.26+ with Mooncake connector for P/D disaggregation; FlexGen NVMe offloading for single-GPU batch inference; FP8 KV cache halves memory footprint; operational metrics: gpu/cpu_kv_cache_usage_perc, num_preempted_reqs, cache_hit_rate)
• [x] database-frontier/121: Apache Parquet 내부 구조: 열 통계·인코딩·블룸 필터로 데이터 레이크 읽기 비용을 줄이는 방법 — published 2026-08-19 (dynamic fallback: data-platform/database topic; Apache Parquet format internals: row groups (128MB default), column chunks, pages (1MB default), file footer with Thrift metadata; encoding schemes: PLAIN_DICTIONARY/RLE_DICTIONARY for low-cardinality columns, DELTA_BINARY_PACKED for monotonic integers/timestamps, BYTE_STREAM_SPLIT for floats; row-group statistics min/max/null_count for predicate pushdown; Column Index (Parquet 2.10+) for page-level skipping within row groups; Split Block Bloom Filter (Parquet 2.9+) for equality predicate acceleration with configurable FPP; sorting data by filter columns maximizes statistics selectivity; Z-order/Hilbert curve for multi-dimensional clustering; DuckDB/Spark/Trino/Arrow support for all three skip mechanisms)
• [x] ai-frontier/127: MoE 모델 추론 최적화: Expert 병렬성·WideEP·로드밸런싱 실전 패턴 — published 2026-08-19 (dynamic fallback: AI systems topic; vLLM Elastic EP blog 2026-05-14 within 90 days; WideEP large-scale MoE serving with DP Attention + EP Expert FFN split; Elastic Expert Parallelism dynamic scale up/down via Ray Serve; All-to-All communication overlap with CUDA stream pipelining; UltraEP arXiv:2606.04101 June 2026; ReaLB arXiv:2604.19503; DeepSeek V3 WideEP GB200 26.2K prefill TPGS 10.1K decode TPGS; expert load imbalance metrics and auxiliary loss balancing)
• [x] database-frontier/122: Apache Arrow Flight SQL: 컬럼형 고성능 DB 연결의 원리와 운영 패턴 — published 2026-08-19 (dynamic fallback: data-platform/database topic; ADBC 24 released 2026-07-28 within 90 days; Arrow Flight SQL protocol for high-performance columnar DB connectivity over gRPC; eliminates row-serialization and type-conversion overhead; parallel endpoint streaming for partition-parallel data extraction; StarRocks native support; DuckDB via sqlflite/porter; PostgreSQL pg_arrow_flight_sql 0.1.0 adapter; ADBC unified client API for Python/Java; 2-5x throughput vs JDBC on bulk extraction benchmarks; zero-copy RecordBatch to Pandas/Polars/Spark)
• [x] ai-frontier/128: 추론 시간에 계산을 더 쓰면 LLM이 더 잘 추론하는가: 세 가지 레짐과 운영 함의 — published 2026-08-19 (dynamic fallback: AI engineering topic; arXiv:2608.04001 submitted 2026-08-04; three TTS regimes: single-trajectory sequential scaling, leaf-level scaling with terminal reduction, prefix-level scaling/search; prefix tree formalism; overthinking phenomenon; operational implications for serving engineers)
• [x] database-frontier/123: Apache DataFusion Comet 1.0: Spark 쿼리를 Arrow 네이티브 Rust 엔진으로 가속하는 방법 — published 2026-08-19 (dynamic fallback: data-platform topic; Apache DataFusion Comet 1.0.0 released 2026-08-07; Spark physical plan interception executing via DataFusion Rust engine; Arrow C Data Interface zero-copy JNI; Codegen Dispatch for unsupported expressions; CachedParquetFileReaderFactory; ~2× speedup on TPC-DS SF1000; Iceberg support)

• [x] ai-frontier/129: SmoothAgent: Lookahead Context Engineering으로 장기 실행 에이전트의 컨텍스트 변환 TTFT를 62% 줄이는 방법 — published 2026-08-20 (dynamic fallback: AI systems topic; arXiv:2607.00151 July 2026 within 90 days, PVLDB; segment-decomposable context transformations proven prefix-local; lookahead programming model schedules transforms as async ops with KV cache preload; direct KV swap on turn transition avoids re-prefill spike; four context engineering strategies covered — offloading, keep-recent-K, summarization, sub-agent isolation; tail TTFT reduced 62.0% average (model A) and 61.5% (model B); summarization up to 11.9× TTFT improvement; offloading gains smaller under high concurrency; admission control prevents lookahead from harming foreground on overloaded GPUs; RoPE-family models are safe target; GitHub PanZaifeng/SmoothAgent)
• [x] database-frontier/124: Databricks + Electric: PGlite WASM Postgres와 Shape 동기화로 AI 에이전트마다 자체 데이터베이스를 주는 방법 — published 2026-08-20 (dynamic fallback: database/data-platform topic; Databricks acquisition of Electric announced 2026-08-11 within 90 days; PGlite = full PostgreSQL compiled to WebAssembly single-user mode, <3MB gzipped, no Linux VM, embeddable in browsers/Node/Bun/Deno/agent sandboxes; PGlite v0.4 released 2026-03-25 decoupled initdb as separate WASM process via Emscripten PROXYFS + PostGIS + connection multiplexing; supports pgvector for local RAG; PGlite weekly downloads grew 1M→13M from Aug 2025 to Aug 2026; Electric Shape sync engine tails Postgres logical replication WAL, matches changes to registered Shapes (partial replicas with WHERE clauses), fans out via HTTP long-poll (CDN-cacheable, firewall-friendly, stateless); per-Shape files: shape log + offset index; three-tier Databricks stack Lakehouse (Delta/Iceberg) + Lakebase (Neon acquired May 2025, serverless Postgres, CoW branching) + PGlite (per-agent local); write-through vs offline-first tradeoff; operational boundaries: data volume ~GB, replication slot WAL retention, Shape filter is security boundary)
• [x] ai-frontier/130: SafeKV: 프리픽스 캐시 공유가 뚫린 타이밍 사이드채널을 선택적 공유로 막는 방법 — published 2026-08-20 (dynamic fallback: AI systems/security topic; arXiv:2508.08438 August 2026 within 90 days; global prefix cache sharing (vLLM/SGLang/TRT-LLM) creates API-observable timing side channel; unprivileged remote attacker measures TTFT to detect cached prefixes and reconstruct sensitive prompts token-by-token; SafeKV three-component design: hybrid multi-tier detection pipeline (rule-based + NER + context validator) classifies KV cache entries as public/private at prefill time, unified radix-tree with owner-scoped private leaves across HBM/DRAM/SSD tiers, entropy-based access monitoring detects narrow probe patterns and temporarily flips subtrees to private; 94-97% timing attack mitigation; vs per-user isolation 40.58% TTFT improvement, 2.66× throughput; Qwen3-235B multi-turn chat cache-induced TTFT overhead 50.41%→11.74%; related: Shadow in the Cache arXiv:2508.09442 KV-Cloak reversible obfuscation, CachePrune arXiv:2605.23640, Agent-Assisted Side-Channel on RAG arXiv:2606.21842)
• [x] database-frontier/125: QuestDB 10.0: QWP 단일 이진 프로토콜로 ILP·PG Wire를 대체하고 Arrow 스트리밍 읽기까지 한 커넥션으로 처리하는 방법 — published 2026-08-20 (dynamic fallback: database/time-series topic; QuestDB 10.0.0 released 2026-08-06 within 90 days; QWP binary columnar WebSocket protocol over existing HTTP port 9000 unifies writes and Arrow reads on single connection; 5.3M→19M rows/s ingestion vs ILP (~3.6×); 220M rows/s Arrow streaming reads with sub-millisecond first batch latency; 3–4× payload compression on typical rows; client-side store-and-forward with sf_durability=periodic disk checkpoint + request_durable_ack=on server confirmation for RPO≈0 in Enterprise; ILP and PG Wire retained for backward compat; Parquet partitions as first-class citizens with schema evolution; Enterprise 4.0 object-storage tiering; Iceberg/DuckLake catalog registration without data copy; Live Views Beta for millisecond-latency window function results; breaking: SHOW PARTITIONS adds seqTxn+isRemotelyServed columns, pg_class.relkind per-object kind, HTTP EXPLAIN plain text, range window overflow rejected at compile time, universal query cancellation)

• [x] ai-frontier/131: XQuant: KV 캐시 대신 활성화 X를 저장하고 K·V를 재물질화해 LLM 서빙의 메모리 벽을 넘는 방법 — published 2026-08-20 (dynamic fallback: AI systems topic; arXiv:2508.10395 within 90 days; caches quantized post-LayerNorm input activation X instead of K/V and rematerializes K=X·W_K, V=X·W_V on-the-fly at decode; 1 tensor instead of 2 gives immediate 2× reduction on MHA; XQuant-CL shared-base + per-layer delta compression reaches sub-1.4-bit equivalent; results 7.7× memory savings with <0.1 PPL on Llama-2-7B/13B, 10× at 0.01 PPL and 12.5× at 0.1 PPL for XQuant-CL, Mistral-7B 3-bit 5× at <0.1 PPL; GQA/MLA extension via SVD on W_K/W_V into shared latent space; roofline analysis shows re-projection GEMMs hidden by HBM load savings on H100/H200/B200; llama.cpp GitHub discussion #15400 tracks reference implementation)
• [x] database-frontier/126: Databricks OpenSharing SecureConnect GA: 방화벽 뒤 스토리지를 리시피언트마다 열지 않고 공유하는 방법 — published 2026-08-20 (dynamic fallback: data-platform topic; SecureConnect GA in Databricks AWS August 2026 release notes, Azure Premium accounts default mid-September 2026 with the rest by early 2027, within 90 days; Databricks-managed proxy routes storage access on behalf of recipients so providers allowlist stable Databricks IP ranges ONCE instead of per-recipient; provider-side materialize + filter with pre-signed short-lived URLs delivered to recipients; serverless recipients zero-config, classic/open recipients need one-time outbound allow to Databricks IPs; one-way proxy — no inbound to recipient networks; announced at Data + AI Summit 2026 June 10 as part of OpenSharing under Linux Foundation with OpenAI/SAP/Stripe/Atlassian/LSEG/Amadeus backing, Iceberg REST client compatibility, on-prem/private-cloud support via Everpure/MinIO/Qumulo; provider billed for egress today, recipient charges and expanded charge types planned)
• [x] ai-frontier/132: Lodestar: 온라인 학습으로 LLM 추론 요청을 라우팅하는 방법 — published 2026-08-21 (dynamic fallback: AI systems topic; arXiv:2606.00946 submitted 2026-05-31 within 90 days; online-learning LLM inference router for distributed GPU clusters; continuously collects cluster-state snapshot (KV cache utilization, queue length, request characteristics, hardware utilization) per request and trains online reward predictor to maximize reward e.g. minimize TTFT; no offline training phase — learns efficient routing within ~5 minutes of live traffic; 1.41× lower avg TTFT and 1.47× lower P99 TTFT vs prefix-cache + load-aware heuristic baseline; up to 2.15×/1.86× on homogeneous and 4.38×/4.42× on heterogeneous clusters; cloud-native, integrates with vLLM; adapts automatically to cluster topology changes and new instances; authors: Gangmuk Lim, Wanyu Zhao et al., UIUC + ByteDance + University of Edinburgh)
• [x] database-frontier/127: dbt Core 2.0 Fusion 엔진: Rust 런타임 오픈소스화와 Parquet 아티팩트로 데이터 모델 빌드 기반을 바꾼 방법 — published 2026-08-21 (dynamic fallback: data-platform topic; dbt Core 2.0 alpha released 2026-06-01 within 90 days, coinciding with Fivetran + dbt Labs merger completion; Fusion engine's Rust core open-sourced under Apache 2.0 for first time; dbt Core and Fusion now share a single Rust runtime — open-source users get same high-performance foundation; significant parse-time improvements especially on large projects via parallel Rust parser; new Parquet artifacts (manifest.parquet) replace manifest.json — directly queryable via DuckDB SQL; tightly-defined language spec — type mismatches caught at parse time not runtime; v2.0.0-alpha.2 available on GitHub; adapter compatibility: BigQuery/Snowflake/Databricks early support; migration check: existing JSON-parsing CI scripts need Parquet update or JSON-output option; sources: docs.getdbt.com/blog/dbt-core-v2-is-here, github.com/dbt-labs/dbt-core/blob/main/docs/roadmap/2026-06-announcing-v2.md)

• [x] ai-frontier/133: vLLM 0.27: Kimi K3 네이티브 지원·Rubin SM107 초기 활성화·Rust gRPC 제어면이 만드는 서빙 아키텍처 변화 — published 2026-08-21 (dynamic fallback: AI systems topic; vLLM v0.27.0 released 2026-08-10 within 90 days; Kimi K3 full support stack — AttnRes kernels, DeepGEMM, DSpark AR fusion speculative decoding, shared-expert sharding, 370 tok/s on 16× GB300 NVL72; Qwen3.5 dense+MoE; Rust frontend gRPC control plane for engine health, abort, server/model discovery, KV event source discovery; NVIDIA Rubin SM107 NVLink all-reduce early enablement; ROCm gfx1250; FlashAttention 4 FP8 KV cache + headdim-256; Model Runner V2 extended to encoder-only attention, sequence pooling, token classification; PyTorch 2.13 breaking upgrade; 561 commits 242 contributors)
• [x] database-frontier/128: Redis 8.10: Compact Hashes로 스키마 공유 메모리를 줄이고 gRPC 제어면·AOF 백업을 더한 방법 — published 2026-08-21 (dynamic fallback: database topic; Redis 8.10 released July 2026 within 90 days; Compact Hashes new encoding stores field names once per shared-schema group reducing per-key overhead for uniform-structure hashes; HIMPORT high-throughput bulk compact-hash insertion; LMOVEM/BLMOVEM multi-element list move; SUNIONCARD/SDIFFCARD set cardinality without materializing; XREAD/XREADGROUP MAXCOUNT+MAXSIZE limits; TLS peer certificate server-to-server authentication replacing shared-password; BACKUP node-side multi-part-AOF backup command; 8.10.1 security fixes buffer-overflow UAF in RDB load/Vector Sets/TLS)

• [x] ai-frontier/134: KAIROS: 에이전트 AI 추론에서 GPU 전력을 컨텍스트 인식 주파수 제어로 27% 줄이는 방법 — published 2026-08-21 (dynamic fallback: AI systems topic; arXiv:2604.16682 submitted 2026-04-17 within 90 days; Yichao Yuan, Mosharaf Chowdhury, Nishil Talati, University of Michigan; KAIROS first work to optimize power for agentic inference serving using agent context as first-class control signal; jointly manages GPU DVFS frequency, per-instance concurrency, and multi-instance request placement; tracks requests at agent session granularity to detect context growth and tool-call idle gaps; agentic workloads are decode-dominated due to prefix caching reuse — giving larger DVFS headroom than stateless serving; thrashing risk when KV cache memory fills: naive DVFS lowers bandwidth and worsens decode bottleneck; average 27% power reduction up to 39.8% while maintaining target performance SLO)
• [x] database-frontier/129: PostgreSQL 19 Beta 3: pg_plan_advice 플랜 힌트·병렬 Autovacuum·보안 패치 28개로 GA 직전에 확인해야 할 변화 — published 2026-08-21 (dynamic fallback: database topic; PostgreSQL 19 Beta 3 released 2026-08-13 within 90 days; pg_plan_advice first official plan hint module; parallel autovacuum index workers for large tables; pg_stat_autovacuum_scores new view for vacuum prioritization monitoring; NOT IN to ANTI JOIN automatic optimizer conversion; SIMD-accelerated COPY FROM; LZ4 default TOAST compression replacing pglz; logical replication sequence sync stabilization race-condition fixes enabling wal_level=replica; 28 CVE security patches across PG14-18 including heap buffer overflows and RCE; PG14 EOL notice issued — 14.24 likely last patch; GA expected September/October 2026)
• [x] ai-frontier/135: RAD·SLAI: LLM 추론 스케줄링의 처리량 최적성과 TBT SLO 인식 설계 — published 2026-08-22 (dynamic fallback: AI systems topic; arXiv:2508.01002 preprint August 2025, ACM SIGMETRICS 2026 presentation June 2026 within 90 days; Bari, Hegde, de Veciana, UT Austin; RAD Resource-Aware Dynamic scheduler: optimal tiling + dynamic resource allocation proved throughput-optimal under mild conditions; SLAI SLO-Aware LLM Inference scheduler: real-time TBT deadline prioritization for decode + LPL shortest-prompt-first ordering for prefill; evaluation on openchat_shareGPT4 Mistral-7B RTX ADA 6000 vs Sarathi-Serve: −53% median TTFT, +26% max serving capacity at TTFT ≤ 0.5s; two-phase prefill/decode tiling formalism; throughput optimality proof; practical LPL ordering applicable without scheduler changes)
• [x] database-frontier/130: Databricks Lakebase: Lakehouse에 Postgres를 내장하고 브랜치로 운영하는 방법 — published 2026-08-22 (dynamic fallback: database/data-platform topic; Databricks Lakebase programmatic management GA August 2026 within 90 days; built on Neon acquired May 2025; storage-compute separation: Pageserver (Postgres 8KB pages on S3/GCS with LSN versioning) + Safekeeper (WAL quorum durability) + compute node (standard Postgres, cold-start <500ms, zero idle cost); database branching O(1) CoW by LSN pointer — terabyte production DB branches in seconds with near-zero additional storage; synced tables: Unity Catalog Delta/Iceberg → read-only Lakebase Postgres schema with seconds latency; programmatic management GA: REST API, Databricks CLI, Python/Java/Go SDKs; operations: projects, branches, endpoints, roles, credentials, synced tables, catalog bindings; compliance profile workspaces default-enabled; write-intensive OLTP limitation due to Safekeeper quorum in commit path)

• [x] ai-frontier/136: D2F: 확산 언어 모델이 KV 캐시를 얻고 자기회귀 모델보다 2.5배 빨라지는 방법 (arXiv:2508.09192) — published 2026-08-22 (dynamic fallback: AI systems topic; arXiv:2508.09192 August 2026 within 90 days; SJTU Deng Lab; D2F Discrete Diffusion Forcing applies block-wise causal attention to dLLMs — within-block bidirectional attention, between-block causal mask — enabling standard KV cache for completed blocks; dLLMs (LLaDA-8B, Dream-7B) normally use bidirectional attention preventing KV cache reuse — every denoising step recomputes full N² attention; inter-block parallel decoding predicts subsequent blocks without waiting for prior block completion; asymmetric distillation: bidirectional dLLM teacher trains block-wise causal D2F student on Bespoke-Stratos-17k 12h 8×A100-SXM4-40GB; 50× faster than vanilla dLLMs, 2.5× faster than LLaMA3-8B/Qwen2.5 on GSM8K; D2F-Dream-Base-7B-LoRA HuggingFace; GitHub: SJTU-DENG-Lab/Discrete-Diffusion-Forcing)
• [x] database-frontier/131: Elastic 9.5: Columnar 인덱스 모드·VectorDB 자동 설정·배치 쿼리 실행으로 달라진 운영 기준 — published 2026-08-22 (dynamic fallback: database topic; Elastic 9.5 released 2026-07-09 within 90 days; three structural changes: (1) Columnar Mode — Columnar Metrics 3 bytes/sample ~20% further savings, Columnar Logs inverted index only on message field rest columnar, index.mode=columnar_logs; (2) VectorDB Index Mode index.mode=vectordb + Auto-calibration automatic quantization depth/preconditioning/oversampling selection for target recall; (3) Batched Query Execution search.batched_query_phase=true default on groups shards on same data node into single round-trip; TSDB up to 5× range query speedup via dense sequential access pattern optimization; PromQL endpoint _prometheus/api/v1/query Tech Preview; skew caution: disable batching when shard processing time highly unequal)
• [x] ai-frontier/137: FP8 혼합 정밀도 LLM 학습: Transformer Engine과 Delayed Scaling으로 H100·B200 학습 비용을 절반으로 줄이는 방법 — published 2026-08-22 (dynamic fallback: AI systems/engineering topic; NVIDIA Transformer Engine FP8 training; E4M3 for forward pass weights/activations, E5M2 for backward gradients; Delayed Scaling per-tensor amax history 1024 iterations avoids per-step all-reduce; GEMM output accumulated in FP32 via Tensor Core hardware; DeepSeek-V3 671B FP8 pretraining case study 2048×H800 achieved BF16-parity perplexity; PyTorch 2.13 torchao.float8 stable API; 1.5–2× training throughput improvement on H100/B200; overhead: amax history memory, loss spike monitoring required)
• [x] database-frontier/132: Polars 1.x: Lazy 실행·스트리밍·Arrow 네이티브로 Pandas 병목을 없애는 Rust DataFrame 엔진 — published 2026-08-22 (dynamic fallback: data-platform topic; Polars v1.0 GA July 2024; Rust + Apache Arrow columnar memory layout; Lazy API predicate pushdown, projection pushdown, CSE for query optimization; Streaming mode chunked execution for out-of-core datasets exceeding RAM; Expression API SIMD-vectorized parallel column operations; SQL interface via SQLContext for familiarity; DuckDB zero-copy via adbc; PyIceberg Polars scan integration; Spark Arrow-optimized writes from Polars DataFrames; 10–30× faster than Pandas on typical aggregation benchmarks; memory usage 2–5× lower than Pandas for same workload)
• [x] ai-frontier/138: Gemini 3.7 Flash: 튜너블 사고 수준과 DeepSWE 65% 로 에이전트 작업의 기준을 바꾼 모델 — published 2026-08-23 (dynamic fallback: AI engineering topic; Gemini 3.7 Flash released 2026-08-13 within 90 days; model ID gemini-3.7-flash; 1,048,576 input tokens / 65,536 output tokens; thinking_budget removed, thinking_level=low/medium/high replaces it; temperature/top_p/top_k/candidate_count must be default or 400 error; DeepSWE v1.1 65.3% vs 48.6% for 3.6 Flash (+16.7pp); FrontierCode 1.1 Main 43.6% vs 34.4%; Terminal-bench 2.1 85.8% vs 78.0%; function calling, code execution, computer use preview, search grounding, structured output; multimodal text/image/audio/video/PDF; knowledge cutoff March 2026; introductory pricing $0.75/$3.75 per 1M tokens until 2026-12-31 then $1.50/$7.50; powers Google Spark personal agent)
• [x] database-frontier/133: DuckDB 2.0 Cyanoptera 프리뷰: 비동기 I/O·클라이언트-서버·VARIANT 샤딩·트리거로 달라지는 운영 경계 — published 2026-08-23 (dynamic fallback: database/data-platform topic; DuckDB v2.0 preview blog 2026-08-17 within 90 days; codename Cyanoptera; async I/O: I/O layer decoupled from query processing, S3 CSV 887s→45s ~20× improvement, up to 40× on some workloads, async Parquet and CSV reads; client-server mode via Quack extension + CONNECT statement; VARIANT type: automatic shredding of repeated structures, Parquet read/write, extraction pushdown, variant_* functions; triggers: BEFORE/AFTER, FOR EACH ROW/STATEMENT, REFERENCING OLD/NEW TABLE transition tables, multiple triggers per event, DROP TRIGGER; new SQL parser rewrite; new default storage format faster for large indexes and wide tables; C API rework with breaking changes; 10,000+ commits since v1.5 March 2026; GA planned fall 2026; nightly builds available)
• [x] ai-frontier/139: DSPy 3.3.0: GEPA와 dspy.Flex로 프롬프트와 코드 흐름을 동시에 최적화하는 방법 — published 2026-08-23 (dynamic fallback: AI engineering topic; DSPy 3.3.0 released 2026-08-03 within 90 days; GEPA Generalized Error-driven Prompt Augmentation ICLR 2026 integrated as first-class optimizer — reflection-mutation loop using failure traces; 35× fewer rollouts than GRPO, up to 20% accuracy improvement; dspy.Flex first DSPy primitive where GEPA can rewrite entire module source code including Python control flow, Predictor splits, and helper functions; dspy.ReActV2 rebuilds ReAct pattern with provider-native function calling API, parallel tool calls, dspy.History typed conversation history; new typed provider-neutral LM interface abstracting OpenAI/Anthropic/Google at type-adapter layer; pip install dspy==3.3.0)
• [x] database-frontier/134: Apache Airflow 3.3: 상태 저장 태스크(AIP-103)·다중 언어 SDK(AIP-108)·플러거블 재시도로 달라진 운영 기준 — published 2026-08-23 (dynamic fallback: data-platform topic; Airflow 3.3.0 released 2026-07-06, 3.3.1 bugfix 2026-08-12 within 90 days; AIP-103 Stateful Tasks: task_state_store key-value accessor for watermarks/progress across retries, per-key retention, clear_on_success, SparkSubmitOperator ResumableJobMixin adoption; AIP-108 Multi-Language Task SDK experimental: @task.stub queue-based routing to JavaCoordinator/ExecutableCoordinator for Java and Go tasks, Variables/Connections/XCom proxied via Execution API; AIP-105 Pluggable Retry Policies: custom retry strategy classes per task; FanOutMapper/FixedKeyMapper/SegmentWindow asset partitioning additions; breaking: OpenTelemetry metrics changed from Gauges to Histograms)
• [x] ai-frontier/140: LLM 0.32: 추론 트레이스·서버사이드 도구·Responses API로 LLM CLI가 에이전트 루프를 지원하는 방법 — published 2026-08-23 (dynamic fallback: AI engineering topic; LLM 0.32 released 2026-08-04 within 90 days; Simon Willison's LLM CLI Python library; reasoning traces streamed to stderr via -R/--reasoning flag, JSON logged to SQLite; OpenAI /v1/responses endpoint replaces /v1/chat/completions for GPT-5.6 models — background tools remain open between turns; server-side tools: WebSearch/CodeInterpreter for OpenAI, WebSearch/WebFetch/CodeExecution/AnthropicMCP for Claude; Message/Part structured output redesign replacing previous dicts; PauseChain exception for human-in-the-loop tool loop control; content-addressed SQLite logging overhaul deduplicating attachments; sqlite-utils>=4.0 required; pip install llm==0.32)
• [x] database-frontier/135: pgvector 0.8.x: 반복 인덱스 스캔·sparsevec·필터 검색으로 PostgreSQL 벡터 DB를 운영하는 법 — published 2026-08-23 (dynamic fallback: database topic; pgvector 0.8.0 released 2024-10-30, 0.8.6 released 2026-07-29 within 90 days; four vector types: vector/halfvec/bit/sparsevec; six distance operators: L2 <->, cosine <=>, inner product <#>, L1 <+>, Hamming <~>, Jaccard <%>; iterative index scan: hnsw.iterative_scan=strict_order/relaxed_order, max_scan_tuples=20000 default, ivfflat.max_probes; prevents filter-induced candidate exhaustion by continuing HNSW graph traversal in batches until LIMIT filled; sparsevec stores only non-zero elements for BM25/SPLADE with cosine/inner-product index support; improved planner cost estimation for filter selectivity; DBA checklist: ef_search, maintenance_work_mem, per-column statistics, REINDEX CONCURRENTLY, autovacuum_vacuum_scale_factor tuning; hybrid dense+sparse RRF pattern; partition-per-tenant strategy)
• [x] ai-frontier/141: Jalapeño: OpenAI와 Broadcom이 설계한 LLM 전용 추론 ASIC — GPU에서 전용 실리콘으로 가는 이유 — published 2026-08-24 (dynamic fallback: AI systems topic; Jalapeño OpenAI+Broadcom custom LLM inference ASIC announced 2026-06-24 within 90 days; reticle-sized single die on TSMC 3nm; 2.5D silicon interposer with 8 HBM3E stacks targeting 20TB/s aggregate memory bandwidth; systolic array architecture optimized for low-precision matrix multiply (FP8/INT4) for transformer decode; 9-month design cycle from architecture to tape-out; deployment target late 2026 at OpenAI datacenters; designed to serve 10T+ parameter models in multi-chip module; comparison to Trainium2: OpenAI chooses custom silicon over third-party for power efficiency and supply independence; three operational implications: inference latency determinism, GB/s/W efficiency vs GPU, operational risk of custom silicon at scale)
• [x] database-frontier/136: Apache Flink 2.3: FROM_CHANGELOG·TO_CHANGELOG 연산자와 Materialized Table DDL 완성으로 스트리밍-배치 경계를 좁힌 방법 — published 2026-08-24 (dynamic fallback: data-platform/streaming topic; Apache Flink 2.3.0 released 2026-06-25 within 90 days; 15 FLIPs implemented; FROM_CHANGELOG table-valued function converts changelog stream (CDC RowKind +I/-U/+U/-D) to append stream with retract or upsert strategy; TO_CHANGELOG wraps append result as changelog for downstream sinks; Materialized Table CREATE accepts explicit column definitions including WATERMARK and PRIMARY KEY; ALTER MATERIALIZED TABLE ADD/MODIFY/DROP column and RENAME TO DDL; fine-grained REFRESH with PARTITION BY for incremental-only recompute; native S3 filesystem flink-s3-fs-aws on AWS SDK v2 eliminating hadoop-aws dependency (experimental); Adaptive Partition Selection monitors downstream partition load at runtime and routes data to less busy partitions to reduce backpressure in shuffle-heavy and window aggregation pipelines; Flink 2.3 base engine not previously covered in series — Flink CDC 3.6 and Flink Agents 0.3 cover separate sub-projects)
• [x] ai-frontier/142: SwiftSpec: 드래프터와 타깃을 서로 다른 GPU에 분리하고 비동기로 겹쳐 실행해 LLM 저지연 추론을 실현하는 방법 (ASPLOS 2026) — published 2026-08-24 (dynamic fallback: AI systems topic; arXiv:2506.11309 June 2026 within 90 days; ByteDance Seed + University of Chicago; disaggregated speculative decoding with disjoint draft GPU group and target GPU group; async pipeline overlaps draft iteration N with target verification of iteration N-1; parallel tree generation (k>1 draft tokens per position), tree-aware KV cache management for target, fused gather/scatter kernels for tree attention; 1.75× end-to-end speedup vs baseline; 348 tokens/s on Llama-3-70B with 8×H800; ASPLOS 2026 accepted)
• [x] database-frontier/137: pg_duckdb 1.0: PostgreSQL 안에 DuckDB 컬럼 엔진을 내장해 분석 쿼리를 2–7배 빠르게 만드는 방법 — published 2026-08-24 (dynamic fallback: database/data-platform topic; pg_duckdb v1.0.0 released July 2026 within 90 days by MotherDuck + DuckDB team; GitHub duckdb/pg_duckdb; embeds DuckDB directly inside PostgreSQL process — no sidecar; SET duckdb.execution=true session GUC routes SELECT/COPY/CTAS to DuckDB vectorized engine; 1.0 adds parallel table scanning (duckdb.max_threads), new types DOMAIN/VARINT/TIME/TIMETZ/BIT/VARBIT/UNION/MAP/STRUCT, community extension loading (httpfs, spatial, iceberg), COPY auto-detection of .parquet/.json/.ndjson with Azure/HTTP locations; MotherDuck serverless offload via MDTOKEN; TPC-H SF10 benchmark 2–7× faster on 32-core server; memory caveat: duckdb.memory_limit pool separate from shared_buffers; MVCC boundary: MotherDuck-replicated data not in PostgreSQL transaction scope; PostgreSQL 14+ required; 3000+ GitHub stars, 1M+ downloads as of May 2026)
• [x] ai-frontier/143: BaseRT: Apple Silicon에서 Native Metal로 LLM 추론 처리량의 새 기준을 세운 방법 — published 2026-08-24 (dynamic fallback: AI systems topic; arXiv:2607.00501 July 2026 within 90 days; native Metal inference runtime for Apple Silicon M-series; layer-shape-specific kernel fusion + Unified Memory direct mapping + batched MTLCommandBuffer dispatch; Q2-FP16 quantization support across Qwen3/Llama-3.2/Gemma-4 families; decode throughput 1.15-1.56× faster than llama.cpp, 1.35× over MLX on M4 Pro Q4 models; prefill throughput up to 1.78× over MLX for Qwen3-30B-A3B Q4 on long prompts; follow-up arXiv:2607.19438 extends to Apple M5 Neural Accelerator)
• [x] database-frontier/138: Litestream 쓰기 가능 VFS: SQLite 페이지를 S3에 두고 관리형 DB 없이 단일 프로세스 애플리케이션을 운영하는 방법 — published 2026-08-24 (dynamic fallback: database topic; Litestream writable VFS released February 2026 within 180-day fallback window; bex.co practical guide published 2026-08-07; LTX format: page-number-sorted transaction changesets with trailer index for O(1) page lookup; LSM-tree compaction: L0 immediate upload → 30s → 5min → hourly snapshots; VFS read: per-page S3 Range HTTP requests from in-memory page index; VFS write: LITESTREAM_WRITE_ENABLED=true single-writer constraint; conflict detection via epoch+TXID comparison before upload; ~1s durability window; suitable for per-tenant DBs, agent state stores, restart-tolerant single-writer tools; not suitable for concurrent multi-writer OLTP or latency-sensitive reads)
• [x] ai-frontier/144: AOSpec: 에이전트 LLM 서빙에서 행동과 관측을 동시에 투기하여 TAO 루프 지연을 줄이는 방법 — published 2026-08-25 (dynamic fallback: AI systems topic; arXiv:2608.00881 August 2026 within 90 days; authors Hao Mark Chen, Jinnan Guo, Wayne Luk, Hongxiang Fan; AOSpec co-speculates both Action (what tool call will be generated) and Observation (what the tool will return) simultaneously while the LLM is generating its tool call tokens; enables pre-loading of speculated KV cache for the Observe step before actual tool execution completes; lossless: incorrect speculations are discarded and replaced with actual tool result; orthogonal to token-level speculative decoding; prior work SPORK speculates action only, SpecBox speculates sandbox startup, IdleSpec speculates planning — AOSpec is first to speculate observation content itself; applicable to coding agents, research agents, any multi-step agentic workload)
• [x] database-frontier/139: Apache DataFusion 55: MERGE INTO·Sort Pushdown·Range Partitioning으로 Rust 쿼리 엔진이 Lakehouse 표준에 가까워진 방법 — published 2026-08-25 (dynamic fallback: data-platform topic; DataFusion 55.0.0 released July/August 2026 within 90 days; blog post issue #24216 opened 2026-08-10; three major additions: (1) MERGE INTO — MergeInto logical node in datafusion-expr, SQL planner with WHEN MATCHED/NOT MATCHED branch handling, TableProvider merge_into hook for delta-rs/Iceberg-rs integration; (2) Sort Pushdown — WITH ORDER table declaration propagates preferred_ordering to TableScan, SortExec removed from physical plan when file ordering matches query ORDER BY, DynamicFilter threshold skips trailing Parquet files in TopK queries; (3) Range Partitioning — RangePartitionExec stabilized for global ORDER BY shuffle, sampling-based partition boundary calculation; downstream impact: delta-rs native MERGE INTO, DataFusion Comet MERGE path, Ballista global sort, LanceDB/IOx ORDER BY optimization; sort pushdown requires files to actually be sorted — incorrect WITH ORDER declaration causes wrong results)
• [x] ai-frontier/145: Sorted-F: 비균질 프리필·디코드 길이에서 표준 스케줄러의 근사비가 무한대가 되는 이유와 F-메트릭 해법 — published 2026-08-25 (dynamic fallback: AI systems topic; arXiv:2508.06133 v4 updated June 2026 within 90 days; authors Meixuan Wang (Tsinghua), Yinyu Ye (Stanford/HKUST), Zijie Zhou (HKUST); offline LLM serving scheduling under fixed KV-cache memory budget with heterogeneous prefill and decode lengths; NP-hard problem; FCFS/shortest-output-first/total-size-priority all have unbounded approximation ratios; Sorted-F constant-factor approximation via F-metric balancing batch-size against downstream decode cost; exact DP for small instances, local-search and greedy heuristics for larger, LP-guided and receding-horizon online variants; evaluated on public workloads mixing short conversations and long-document summarization; consistent latency reduction vs baselines; near-optimal on tractable LP-relaxation lower bound)
• [x] database-frontier/140: ClickHouse 26.7: DPsub 조인 재정렬·QBit Int8·Delta 코덱 벡터화로 분석 쿼리를 최대 8배 빠르게 만드는 방법 — published 2026-08-25 (dynamic fallback: database topic; ClickHouse 26.7 released 2026-08-06 within 90 days; DPsub join-reordering considers all valid join orders via dynamic programming, supports non-inner joins, some queries up to 8× faster; compact 8-byte index-based hash-join row references for all key types — median 12% speedup 15% lower peak memory on 100–300M row joins; Delta codec SIMD vectorization 1.5–5× faster decompression for 8/16/32-bit types, column scan up to 20% faster; QBit Int8 support 75% memory reduction, strided storage reads fewer dimensions for two-stage search, randomized Hadamard rotation for better quantization, quantization codecs for coarse+fine search; EXPLAIN ANALYZE per-stage timing/parallelism/rows-read; filtered projections store only rows matching a condition; wide-integer comparisons up to 7×; statistical/bitwise aggregates up to 4×)
• [x] ai-frontier/146: 출력 길이 불확실성이 있는 LLM 추론 스케줄링: A_min과 A_max로 예측 오류를 흡수하는 방법 (arXiv:2508.14544) — published 2026-08-25 (dynamic fallback: AI systems topic; arXiv:2508.14544 August 2026 within 90 days; adaptively robust LLM inference scheduling under prediction uncertainty in output length; A_max algorithm uses predicted upper bound d_max as job size for conservative preemption, preventing memory overflow; A_min algorithm uses predicted lower bound d_min as job size, achieves O(log(d_max/d_min)) competitive ratio — optimal up to constant; interval prediction [d_min, d_max] from ML model bridges naive SJF and conservative upper-bound approaches; A_min dominates A_max for throughput by exploiting tighter lower bounds; real-world simulation shows 30-40% latency reduction vs SJF-upper-bound at 80% GPU utilization; online scheduling framework with preemption-on-overflow mechanism; applicable to vLLM/SGLang/TensorRT-LLM FCFS queues where token budget prediction is available)
• [x] database-frontier/141: Databricks 2026년 8월: RBAC GA·관리형 Iceberg 공유 GA·DML 최소권한으로 달라진 데이터 거버넌스 경계 — published 2026-08-25 (dynamic fallback: database/data-platform topic; Databricks August 2026 release notes; RBAC GA: ASSUME ROLE SQL command swaps session permissions to role's grants only (not additive), audit log captures identity_metadata.run_by + identity_metadata.run_as, ABAC masks/filters via current_role(), two patterns — warehouse-service-account + role OR human-user + assume role; Managed Iceberg Sharing GA: share managed Iceberg tables via open Iceberg REST Catalog API to external engines without data copy, fine-grained row/column sharing with Unity Catalog governance, Python SDK create_share + share_iceberg_tables; INSERT/UPDATE/DELETE DML privileges Beta: replace MODIFY with minimum necessary privilege per operation, prevents unintended writes; Lakeflow Connect SharePoint GA + PagerDuty Beta)

• [x] ai-frontier/147: AnchorKV: 앵커-잔차 표현으로 KV 캐시를 20배 줄이면서 토큰을 버리지 않는 방법 — published 2026-08-26 (dynamic fallback: AI systems/engineering topic; arXiv:2608.02901 submitted August 3 2026 within 90 days; anchor-residual KV cache compression scheme: k anchors stored at full precision, all other tokens represented as residuals relative to nearest anchor, refinement pass corrects tokens with highest output impact; 20× compression ratio without discarding any token; 99% full-cache score retention at 70B scale; compression runs once post-prefill on frozen model, attention arithmetic unchanged; orthogonal to quantization and eviction approaches; advantages over eviction: no token loss; advantages over quantization: residual distribution is tighter yielding higher compression per bit)
• [x] database-frontier/142: pg_clickhouse v0.10: PostgreSQL에서 ClickHouse로 분석 쿼리를 투명하게 위임하는 FDW의 작동 방식 — published 2026-08-26 (dynamic fallback: database/data-platform topic; pg_clickhouse v0.10.0 released July 2026 within 90 days; open-source Apache 2.0 PostgreSQL Foreign Data Wrapper; transparent query pushdown of WHERE/GROUP BY/ORDER BY/HAVING/SEMI JOINs/window functions to ClickHouse columnar engine; TPC-H complete pushdown coverage 12→16 of 22 queries; C driver replaces binary driver — doubles function/aggregate pushdown surface, fixes two concurrency race conditions in v0.9; partial subquery pushdown GA; ClickHouse -If suffix aggregates and statistical aggregates newly pushed down; TPC-H benchmark Q1 4693ms→268ms 17.5×, Q3 742ms→111ms 6.7×, Q6 764ms→53ms 14.4×; fastest PostgreSQL analytics extension on ClickBench; remaining gap: subquery JOIN tree deparser, DML support roadmapped)
• [x] ai-frontier/148: Sangam: 확산 언어 모델(dLLM)을 AR 서빙 스택으로 효율적으로 처리하는 방법 — published 2026-08-26 (dynamic fallback: AI systems topic; arXiv:2607.04206 submitted July 5 2026 within 90 days; Sangam serving system for cached dLLM inference; problem: dLLMs use bidirectional attention preventing exact KV reuse unlike AR models; Fast-dLLM/dKV-Cache approximate caching creates recurring prefill/decode structure; dLLM decodes are block-sized not token-sized; chunked prefill unavailable due to bidirectional attention; solution: deficit token-budget scheduler admits in-flight decodes first, admits whole indivisible prefills when accumulated budget allows, carries unused budget forward to prevent starvation; three deployment modes: colocated (each worker runs prefill+decode), disaggregated 5P3D (dedicated prefill/decode workers with KV transfer), hybrid 5P3C (dedicated prefill + colocated overflow); evaluated on LLaDA-8B and MDLM-large with ShareGPT and synthetic traces on 8 GPU cluster; colocated lowest prefill queuing delay, disaggregated best when P:D ratio matches workload, hybrid graceful degradation)
• [x] database-frontier/143: Apache DataFusion Comet 1.0.0: Spark 물리 계획을 Rust로 실행하는 네이티브 가속기의 GA와 운영 기준 — published 2026-08-26 (dynamic fallback: data-platform topic; Apache DataFusion Comet 1.0.0 released August 7 2026 within 90 days; Spark plugin replacing physical execution with Rust DataFusion; donated to Apache March 2024, first release 0.1.0 five months later, 20 releases, 120+ contributors; 1.0.0: 400+ Spark expressions supported, all 4 join operators accelerated, window functions/explode/sampling, fully native shuffle with FFI removed on read side, ANSI mode full support for Spark 4.0+, Iceberg tables: planning stays in Iceberg Java on driver, reading moves to iceberg-rust on executors; 24000+ unit tests from Spark's own test suite run in Comet CI; activate with spark.comet.enabled=true; unsupported expressions fall back to JVM; shuffle output row ordering may differ from JVM shuffle)
• [x] ai-frontier/149: Kimi K3: KDA 하이브리드 선형 어텐션과 Stable LatentMoE로 2.8조 파라미터 오픈 프론티어를 세운 방법 — published 2026-08-26 (dynamic fallback: AI systems topic; arXiv:2607.24653 Moonshot AI July 2026 within 90 days; 2.8T total / 104B active MoE; Stable LatentMoE 896 experts top-16 with latent-space routing; SiTU-GLU scaled-tanh soft-cap on SwiGLU branches prevents activation explosion at extreme sparsity; Quantile Balancing QB sets router bias to exact load-target point without auxiliary routing loss; Kimi Delta Attention KDA hybrid 3:1 — 3 KDA linear-attention O(n) layers + 1 Gated MLA full-attention layer per block; KDA delta-rule recurrence with channel-wise forget gate; 1M token context; 75% KV-cache reduction over full-attention at 1M context; 6× decode throughput at 1M context; MXFP4 QAT throughout SFT+RL stages MXFP8 activations; 2.5× scaling efficiency over Kimi K2; Modified MIT license; vLLM 0.27.0 native AttnRes+DeepGEMM+DSpark stack 370 tok/s on 16×GB300 NVL72; requires 64+ GPU supernode)
• [x] database-frontier/144: Apache Paimon 2.0: Variant 타입·REST 카탈로그·벡터 인덱스로 스트리밍 레이크하우스를 AI 시대로 확장한 방법 — published 2026-08-26 (dynamic fallback: data-platform/lakehouse topic; Apache Paimon 2.0.0 + PyPaimon 2.0.0 released August 7 2026 within 90 days; Variant type Plain/Shredded layouts for semi-structured JSON with path extraction CAST and predicate filtering; Iceberg REST catalog integration via metadata.iceberg.storage=rest-catalog enabling dual-engine serving; HNSW vector index + BM25 full-text search in single table for RAG pipelines; PyPaimon 2.0.0 GA with PyTorch/Ray direct integration; Apache Fluss integration as sub-second stream buffer tier pairing with Paimon lake tier; Paimon 1.3 already covered database-frontier/45 — 2.0 focus is AI workload extensions)
• [x] ai-frontier/150: vLLM 0.27.0: Kimi K3 전 스택 착지·FA4 FP8·gRPC 제어면·MRv2 비생성 확장으로 추론 인프라를 다시 쌓는 방법 — published 2026-08-27 (dynamic fallback: AI systems topic; vLLM v0.27.0 released 2026-08-10 within 90 days; 561 commits 242 contributors 64 new; Kimi K3 full stack: AttnRes kernels + DeepGEMM + DSpark AR fusion + compressed-tensors + shared-expert sharding option; Model Runner V2 expands to encoder-only attention, sequence pooling, token classification, embedding generation; FlashAttention 4 FP8 KV cache + headdim-256 on SM100 + JIT warmup infra; Rust frontend gRPC control plane: engine-aware health reporting, abort control, server/model discovery, KV event source discovery, vllm-bench CLI integration; NVIDIA Rubin sm_107 NVLink all-reduce + ROCm gfx1250 early support; DP+EP external LB fault tolerance framework; Elastic EP async preparation; PyTorch 2.13.0 + torchvision 0.28.0 + Triton 3.7.1; new models: Qwen3.5 dense/MoE, K-EXAONE-2.0-750B-A37B, VaultGemma, jina-embeddings-v5-text-nano)
• [x] database-frontier/145: Apache Arrow 25.0: SVE 동적 디스패치·Feather V1 정리·Flight SQL 결과셋 시그널·IPC 단일화로 컬럼형 생태계를 정비하는 방법 — published 2026-08-27 (dynamic fallback: database/data-platform topic; Apache Arrow 25.0.0 released 2026-07-10 within 90 days; 25.0.1 patch 2026-08-10; 66 contributors 268 commits 222 issues; C++ ARM64 SVE dynamic dispatch via runtime CPU detection replacing NEON-only; xsimd replaces home-grown CPU feature detection for x86/ARM/WASM unified vector abstraction; Feather V1 DeprecationWarning — V1 read/write deprecated, V2 unaffected; Flight SQL ActionCreatePreparedStatementResult.has_result_set boolean field lets server signal whether prepared statement returns result set; Python hypot compute kernel added; pa.OSFile accepts int file descriptor in addition to str path; ChunkedArray.ComputeLogicalNullCount mirrors Array/ArrayData; Table.ToTensor complements RecordBatch.ToTensor; 25.0.1 narrowed Feather deprecation to V1-only, removing FutureWarning from V2 operations; Polars ecosystem compatibility restored; ADBC spec 1.1.0 current, 1.2 milestone richer metadata/catalog in progress)
• [x] ai-frontier/151: FlashInfer 0.6.17: MoE 전문가 병렬성·Blackwell MLA·MXFP4 통합 API로 LLM 서빙 커널 스택을 재조립하는 방법 — published 2026-08-27 (dynamic fallback: AI systems/engineering topic; FlashInfer v0.6.17 released 2026-08-11 within 90 days; MoE-EP production-ready: flashinfer.moe_ep module with CUDA-graph capture+replay, fused single-launch quantize-and-stage hot path, prequantized weight packs, symmetric-buffer workspaces pooled across layers, persistent knob cache, fault-tolerance rank masking for NCCL-EP and NIXL-EP; Blackwell SM12x: FP4 accuracy fixes for NVFP4 bugs on DGX Spark/RTX PRO, W4A16 cooperative persistent launches; MXFP4 unified MoE API — W4A8/W4A16 alongside per-tensor routed FP8 with shared-expert fusion; MLA Decode for Kimi K3 on Blackwell: 96 global query heads, TP-local down to 6, speculative query lengths up to 8, long-context parallelism, CuTe-DSL tile packing, TRTLLM-gen dense/sparse MLA extended to non-power-of-two head counts; MiniMax-M3 sparse attention integration; Ulysses Sequence Parallelism public API: head-scatter/sequence-gather all-to-all with fused NVLink-P2P kernel; integrated into vLLM, SGLang, MLC-Engine, TensorRT-LLM; arXiv:2501.01005)
• [x] database-frontier/146: Snowflake Horizon Catalog과 Iceberg V3: 삭제 벡터·VARIANT·행 수준 계보로 멀티엔진 Lakehouse 거버넌스를 완성하는 방법 — published 2026-08-27 (dynamic fallback: database/data-platform topic; widened to 180-day window — no equivalent multi-engine governance lakehouse topic within 90 days; Snowflake Iceberg V3 GA 2026-05-07 (112 days) + Horizon Catalog REST API GA 2026-02 within 180-day window; deletion vectors: Roaring Bitmap in Puffin files O(1) per-row vs V2 positional delete merge-join O(log n); VARIANT type native columnar semi-structured storage with path predicate pushdown; row lineage _row_commit_snapshot_id for CDC and Snowflake Streams integration; geometry/geography types and nanosecond timestamps; Horizon Catalog: Polaris-powered REST endpoint https://<account>.snowflakecomputing.com/polaris/api/catalog; Spark/Trino/Flink/DuckDB/Dremio/StarRocks/Doris/PyIceberg compatibility; Snowflake RBAC row+column access propagated to external engines (row-level policy not propagated — internal-only); Snowflake Open Catalog: managed Apache Polaris free now pay-per-request planned; V3 format upgrade ALTER ICEBERG TABLE ... SET FORMAT VERSION = 3; V3-aware engine minimums: iceberg-spark-runtime 1.7.0+, Trino 482+, iceberg-flink 1.7.0+, DuckDB 1.5.0+)
• [x] ai-frontier/152: KV 캐시 벡터 양자화와 주의 보존 변환: 쿼리가 보는 곳에 비트를 집중하는 방법 — published 2026-08-27 (dynamic fallback: AI engineering topic; arXiv:2608.04074 August 4 2026 within 90 days; USC ECE team; formulates KV cache quantization as rate-distortion transform coding problem; distortion D_K = E[||Q·K̂ᵀ - Q·Kᵀ||²]; optimal linear transforms T_K, T_V derived from Q/attention-weight covariance eigenvectors; VQ outperforms scalar quantization at 2 bits/element = 8× memory reduction vs BF16)
• [x] database-frontier/147: Apache DataFusion 55: Sort Pushdown·TopK 동적 임계값·MERGE INTO로 Rust 쿼리 엔진이 넓어진 방법 — published 2026-08-27 (dynamic fallback: data-platform/query-engine topic; DataFusion 55.0.0 released July/August 2026 within 90 days; Sort Pushdown uses Parquet row-group min/max statistics for ORDER BY LIMIT pruning; TopK dynamic threshold min-heap updated at every row-group boundary; MERGE INTO SQL support with TableProvider hook; Parquet virtual columns row_number; GroupValuesColumn nested types; scalar UDF NullPropagation metadata; topk_tpch 5/11 queries 3-4× faster total 44% runtime reduction)