feat(wiki): databases/selection — datastore choice by workload (relational/document/vector/graph)

[choiyounggi]

Summary

New wiki category databases/selection — architecture-level guidance on which database type to choose for a workload, motivated by the PostgreSQL / JSONB·MongoDB / Vector DB / Graph DB role split. Researched via Brave Search; every page carries checkable sources.

Pages (4 new):

• choosing-a-datastore-by-workload — default to one relational system of record; three-question gate before adding any specialized store; workload→store mapping; derived-projection + one-way-sync rule
• relational-jsonb-vs-document-store — split queried core into columns, variance into JSONB; JSONB vs MongoDB decision table incl. write-amplification and sharding cases
• vector-search-engine-selection — start with pgvector; measured move-indicators (RAM footprint of HNSW, rebuild stalls, OLTP contention, hybrid search); vector store as re-derivable index
• graph-workloads-relational-vs-graph-db — edge table + recursive CTEs first; traversal-shape decision table; ≥3-hop / pattern-matching / graph-algorithm adoption criteria; CTE cycle/depth pitfalls

Plumbing: wiki/databases/index.md gets the selection section + routing line, root INDEX.md databases row extended, two-way related: links to 4 existing pages, log.md ingest entry.

INGEST_REPORT

Verified best-practice

Each directive is backed by sources returned from live Brave searches (2026-09-03): martinfowler.com PolyglotPersistence, postgresql.org JSON-type docs, pgvector GitHub README, Neo4j graph-vs-relational, plus current benchmark/comparison articles (tensoria.fr 100M-vector benchmark, nisai.dev 2026 comparison, qdrant.tech pgvector-tradeoffs — cited even though it argues the vendor's side, for the sync-pain evidence). No URL was invented; all are cited in each page's Sources section with what they support. Checks pass locally: wiki-lint-prohibitions.js → 0 violations, 75 directive units (unchanged, no test bump needed); wiki-structure-checks.js → 275 pages, 0 findings; tests/wiki-lint-prohibitions.bats, tests/pre-flush-pr-gate.bats, tests/review-routing.bats all green.

Existing-layer check

Pages read: databases index (all category tables), databases-schema-design-column-data-types, databases-schema-design-requirements-to-tables, databases-schema-design-foreign-keys-and-referential-actions, databases-indexing-partial-and-expression-indexes, databases-transactions-isolation-level-selection (format reference). The existing databases categories (indexing, query-optimization, schema-design, operations, data-survey, sqlite, transactions) all assume the relational DB is already chosen — none covers which store type to pick. Overlap handled by deference: column-type mechanics defer to column-data-types (noted in the index load-when line), JSONB indexing defers to partial-and-expression-indexes, both linked two-way instead of duplicated.

Open-PR check

gh pr list --state open → empty. No open knowledge/* or wiki PR overlaps; no sibling-duplicate risk.

Routing decision

Domain databases (root INDEX routing line extended accordingly). No existing category covers store-type selection under another name, so a new selection category was created per wiki-ingest step 3, with four single-trigger pages rather than one oversized page (each has a distinct "load when" situation; the overview page cross-links the three specialized ones).

🤖 Generated with Claude Code

https://claude.ai/code/session_01K7rjymauWGk7kPdNc8iKSa