# [DSGN|INFR] git forge (Forgejo) server redesign
AI Summary
Purpose:
- Preserve the 2026-07-15 design review for rebuilding two IDC git-forge servers
as independent Forgejo instances (INFR-393), authored by 김현욱.
Key points:
- Fixed premises: hardware RAID (explicitly NOT ZFS), two independent Forgejo
forges sharded by team, no inter-server replication, mixed backup (originals only), internal-only, automatic pull-mirror, built-in local accounts (no SSO).
- Hardware per server: EPYC 9555 64C, 128GB DDR5 ECC, 8× 8TB SATA HDD,
Broadcom 9560-16i (8GB cache), 2× 10GbE LACP.
- Leading storage choice: SSD RAID1 for OS, HDD 8× RAID10 (32TB usable) holding
PostgreSQL + git repos + LFS on one XFS volume. RAID6 becomes viable only if CacheVault (write-back) is present; RAID5 rejected (8TB rebuild URE risk).
- ZFS RAIDZ2 rejected with reasoning: hardware-RAID-under-ZFS is an anti-pattern
(no self-healing, cache coherency), and HBA/IT mode is not the chosen ops model.
- Why two independent forges: Gitea/Forgejo has no native repo sharding; real HA
needs shared RWX storage + external DB cluster + Redis + LB, which contradicts local-RAID servers. Team-unit assignment beats alphabet splitting.
- Backup policy: originals get
forgejo dump+ push-mirror to the sibling server
(live git DR); mirrors are excluded (re-fetchable). Off-site/NAS preferred over A↔B cross backup.
- Open items tracked with a decision table: RAID level, DB kind (PostgreSQL vs
SQLite), deploy method, backup target, CI runner scope, CacheVault presence, outbound path for pull-mirrors.
Relevant when:
- Writing resume/portfolio material about infra design judgment (option scoring,
rejected alternatives with reasons), or continuing the INFR-393 work.
Do not read full document unless:
- You need exact RAID scoring tables or the install checklist — then open the
Confluence origin page.
Linked documents:
- ai/sources/confluence/2026-07-17-object-storage-plan.md (storage 방안 3안 비교)
Open Questions
- CacheVault presence (decides RAID10 vs RAID6) and backplane mapping unconfirmed.
- Server IPs intentionally omitted here; see origin page.
Details
Decision style notable for portfolio use: every open choice is compared with 1-5 star tables (RAID levels, DB kinds, auth options, forge software), each with an explicit "why", and rejected options (RAID5, ZFS-on-HWRAID, SSO, native clustering) documented with reasons and external references (Gitea HA docs, Forgejo/Codeberg cluster discussions).