ingest(databases): online schema changes + autovacuum/wraparound

[choiyounggi]

What

Adds two databases-domain wiki pages, derived from the Hatchet "Postgres survival guide" and cross-checked against the official PostgreSQL docs:

• schema-design/online-schema-changes — applying ALTER TABLE / CREATE INDEX to a live table without the ACCESS EXCLUSIVE outage: the non-volatile DEFAULT fast path, ADD CONSTRAINT ... NOT VALID + VALIDATE (only SHARE UPDATE EXCLUSIVE), the CHECK-proves-NOT NULL trick, CREATE INDEX CONCURRENTLY, ADD FOREIGN KEY's lighter lock, lock_timeout for lock-queue pileups, and expand-and-contract to decouple the DB migration from the app deploy.
• operations/autovacuum-and-wraparound (new operations category) — keeping autovacuum ahead of a hot table and off the wraparound cliff: per-table scale_factor/cost_limit tuning, age(datfrozenxid)/relfrozenxid + n_dead_tup monitoring, the ~3M-XID read-only cliff and superuser VACUUM recovery (not FULL/FREEZE), and VACUUM FULL vs pg_repack.

Plumbing: registers both in databases/index.md (new operations category), appends log.md, and adds two-way related: links from column-data-types and index-write-cost.

Decision Log

• Intent: fill two databases gaps a grep of the domain confirmed absent (not valid/expand and contract 0 hits; vacuum only mentioned as index-bloat/visibility side notes, no wraparound/autovacuum-tuning page). Both are production-severe for write-heavy workloads.
• Routing: online DDL → existing schema-design category (adjacent to column-data-types's live type-change note). Autovacuum/wraparound is operational maintenance, which no existing category (indexing/query-optimization/schema-design/transactions/sqlite) owns → new operations category, per the ingest "create a category only if none fits" rule.
• Rejected: folding autovacuum into index-write-cost (that page is about index write amplification, a different trigger — would violate one-case-per-page); duplicating SKIP-LOCKED / CONCURRENTLY content already covered in transactions/isolation-level-selection and indexing/index-write-cost (linked instead).
• Sourcing: every claim cites postgresql.org (sql-altertable, sql-createindex, routine-vacuuming); blog-specific numbers ("10×", "90%") were excluded as unverifiable.

Cross-Check: independently fact-checked against the official PostgreSQL docs by a separate agent — 8 of 9 load-bearing claims CONFIRMED verbatim (ADD FOREIGN KEY SHARE ROW EXCLUSIVE, VALIDATE SHARE UPDATE EXCLUSIVE, non-volatile default fast path, CHECK→SET NOT NULL scan skip, binary-coercible type change, ~3M-XID cliff, per-table storage-param syntax, CONCURRENTLY semantics). One error caught and fixed pre-merge: autovacuum_vacuum_scale_factor default is 0.2, not 0.1 (and the derived 1M-row example → ~200k dead tuples, not ~100k).

Verification

• All related: ids and inline [page-id] refs resolve to existing pages.
• Both new pages registered in databases/index.md with load when lines; log.md entry appended.
• Body ≤120 lines each; no banned vague qualifiers in directives.