[G1-Sync] Manual knowledge update

2026-05-09 21:08:02 +09:00
parent f0befc887a
commit 93ec7e9056
363 changed files with 68333 additions and 64 deletions
@@ -0,0 +1,248 @@
+---
+id: db-vacuum-autovacuum
+title: Vacuum / Autovacuum — Bloat / Wraparound 방지
+category: Coding
+status: draft
+source_trust_level: B
+verification_status: conceptual
+created_at: 2026-05-09
+updated_at: 2026-05-09
+tags: [database, postgres, vacuum, vibe-coding]
+tech_stack: { language: "Postgres", applicable_to: ["Backend"] }
+applied_in: []
+aliases: [VACUUM, autovacuum, bloat, dead tuple, transaction wraparound, freeze]
+---
+
+# Vacuum / Autovacuum
+
+> Postgres MVCC = UPDATE/DELETE 가 dead tuple 생성. **VACUUM 이 정리**. Autovacuum 자동 — 그러나 큰 테이블 / write 많을 때 manual 필요. Bloat / wraparound 위험.
+
+## 📖 핵심 개념
+- Dead tuple: 삭제 / 업데이트 된 row.
+- Bloat: dead tuple 누적 → 테이블 거대.
+- VACUUM: dead tuple 마킹 + 재사용.
+- VACUUM FULL: 테이블 rewrite (lock).
+- ANALYZE: 통계 업데이트.
+
+## 💻 코드 패턴
+
+### Autovacuum 기본
+```
+default 활성. 각 테이블 별:
+- 50 row + 20% 변경 → autovacuum
+- 50 row + 10% 변경 → autoanalyze
+```
+
+### 큰 테이블 = autovacuum 안 따라옴
+```
+1억 row × 20% = 2천만 → autovacuum 한 번에 큰 cost
+→ 자주 작게 하라.
+```
+
+```sql
+ALTER TABLE orders SET (
+  autovacuum_vacuum_scale_factor = 0.05,    -- 5% 마다 (default 20%)
+  autovacuum_analyze_scale_factor = 0.02,
+  autovacuum_vacuum_cost_limit = 2000,      -- I/O cost limit ↑
+);
+```
+
+### 수동 VACUUM
+```sql
+VACUUM (ANALYZE, VERBOSE) orders;
+-- Bloat 정리 + 통계 + log
+
+-- Concurrent (lock 적음)
+VACUUM (ANALYZE) orders;  -- 이미 concurrent default
+
+-- Full (lock 큼 — 잘 안 씀)
+VACUUM FULL orders;  -- 테이블 lock + rewrite
+```
+
+### Bloat 측정
+```sql
+SELECT
+  schemaname, relname,
+  n_live_tup, n_dead_tup,
+  round(100.0 * n_dead_tup / nullif(n_live_tup + n_dead_tup, 0), 2) AS dead_pct,
+  last_vacuum, last_autovacuum
+FROM pg_stat_user_tables
+ORDER BY n_dead_tup DESC LIMIT 20;
+```
+
+→ dead_pct > 20% = vacuum 필요.
+
+### pgstattuple (정확한 bloat)
+```sql
+CREATE EXTENSION pgstattuple;
+SELECT * FROM pgstattuple('orders');
+-- tuple_count, dead_tuple_count, free_space
+```
+
+### Index bloat
+```sql
+CREATE EXTENSION pgstattuple;
+SELECT * FROM pgstatindex('orders_pkey');
+-- index_size, leaf_fragmentation
+```
+
+→ Index bloat → REINDEX.
+
+### REINDEX (CONCURRENTLY)
+```sql
+REINDEX INDEX CONCURRENTLY orders_pkey;
+-- lock 없이 새 인덱스 빌드 + swap
+```
+
+### pg_repack (extension)
+```bash
+pg_repack -d mydb -t orders
+# Online table rewrite — bloat 제거 lock 없이
+```
+
+→ VACUUM FULL 의 zero-downtime 대안.
+
+### Transaction wraparound (위험)
+```
+Postgres = 32-bit transaction ID. 2B → wrap.
+→ 모든 row 의 xmin 이 frozen 안 되면 — DB freeze.
+
+Autovacuum 자동 freeze. but very busy DB 일 때 추적 필요.
+```
+
+```sql
+-- 가장 오래된 frozen 까지의 거리
+SELECT
+  datname,
+  age(datfrozenxid) AS xid_age,
+  2147483647 - age(datfrozenxid) AS remaining
+FROM pg_database
+ORDER BY age(datfrozenxid) DESC;
+```
+
+→ 200M+ = 주의. 1.5B = 위험. 2B = freeze.
+
+```sql
+-- 명시 freeze
+VACUUM FREEZE orders;
+```
+
+### Long-running transaction = autovacuum 차단
+```sql
+-- 오래된 transaction
+SELECT pid, usename, query, state, age(NOW(), query_start) AS age
+FROM pg_stat_activity
+WHERE state != 'idle'
+ORDER BY query_start;
+```
+
+→ 30분+ = 의심. 2시간+ = autovacuum 차단.
+
+```ts
+// App 에서 transaction 짧게
+async function work() {
+  await db.transaction(async (tx) => {
+    // ❌ 안 — fetch 외부 API
+    // const data = await fetch(...);
+    
+    // 짧게 — write 만
+    await tx.insert(...);
+  });
+  // 외부 호출 후 commit
+}
+```
+
+### Maintenance window
+```sql
+-- Off-peak 시간 더 적극 vacuum
+ALTER TABLE big_table SET (
+  autovacuum_vacuum_cost_delay = 0  -- 빠르게
+);
+```
+
+### Monitoring
+```sql
+-- Autovacuum 진행
+SELECT pid, datname, query, query_start
+FROM pg_stat_activity
+WHERE query LIKE 'autovacuum:%';
+
+-- 횟수
+SELECT relname, n_tup_ins, n_tup_upd, n_tup_del,
+       autovacuum_count, autoanalyze_count,
+       last_autovacuum, last_autoanalyze
+FROM pg_stat_user_tables
+ORDER BY n_tup_upd + n_tup_del DESC LIMIT 20;
+```
+
+### Tunables
+```
+autovacuum_max_workers = 3       -- worker 수
+autovacuum_naptime = 1min        -- check 주기
+maintenance_work_mem = 256MB     -- vacuum 메모리
+
+# 큰 cluster:
+autovacuum_max_workers = 6
+maintenance_work_mem = 2GB
+```
+
+### HOT update (no index update)
+```sql
+-- Index 컬럼 안 변경 + page 안 free space 있음 → HOT update
+-- → bloat 적음 + 빠름
+```
+
+```sql
+-- 측정
+SELECT relname, n_tup_upd, n_tup_hot_upd,
+       round(100.0 * n_tup_hot_upd / nullif(n_tup_upd, 0), 2) AS hot_pct
+FROM pg_stat_user_tables ORDER BY n_tup_upd DESC LIMIT 20;
+```
+
+→ hot_pct 높음 = 좋음.
+
+```sql
+-- Fillfactor (page 안 free space 남기기)
+ALTER TABLE orders SET (fillfactor = 80);  -- 20% 비움
+```
+
+### Alarm
+```yaml
+- alert: HighDeadTuples
+  expr: pg_stat_user_tables_n_dead_tup / pg_stat_user_tables_n_live_tup > 0.3
+  for: 1h
+
+- alert: TransactionWraparound
+  expr: pg_database_xid_age > 1500000000
+  for: 10m
+```
+
+## 🤔 의사결정 기준
+| 상황 | 액션 |
+|---|---|
+| Dead pct > 20% | Manual VACUUM |
+| Bloat 큼 | pg_repack |
+| Index bloat | REINDEX CONCURRENTLY |
+| Wraparound 임박 | VACUUM FREEZE |
+| Long transaction | App fix — short tx |
+| 큰 write 테이블 | scale_factor 낮게 |
+
+## ❌ 안티패턴
+- **VACUUM FULL prod**: lock — table 못 사용. pg_repack.
+- **Autovacuum 끄기**: bloat 폭발.
+- **Long transaction (30분+)**: autovacuum 차단.
+- **모든 table 같은 setting**: 큰 vs 작은 다름.
+- **Wraparound 모니터링 X**: freeze 위험.
+- **REINDEX 직접 prod**: lock. CONCURRENTLY.
+- **Fillfactor 100% prod**: HOT update X — bloat.
+
+## 🤖 LLM 활용 힌트
+- 큰 table = scale_factor 낮게 + cost limit 높게.
+- Long transaction = 적.
+- pg_repack 가 zero-downtime maintenance.
+- Wraparound + bloat alarm 항상.
+
+## 🔗 관련 문서
+- [[DB_Postgres_EXPLAIN]]
+- [[DB_Query_Optimization]]
+- [[DB_Lock_Analysis]]