db-specialized-fit

db-specialized-fit (M20) — vector · time-series · graph · search

M20 covers purpose-built engines whose correctness depends on parameters a generic relational audit ignores. Sub-modules carry a letter (M20a..M20d); the scorer maps them to the parent M20 in the paradigm profile (Vector / Time-series / Graph categories in references/scoring-model.md).

M20a — Vector (pgvector, Qdrant, Pinecone, Weaviate)

Feeds: design (Métrica & dimensión, Modelo-version, Metadata/filtro) + performance (Índice & params, Búsqueda filtrada, Recall-vs-latencia).

• Dimension match — column/collection dim equals the embedding model's output dim (e.g. vector(1536) for text-embedding-3-small). A mismatch is a hard bug (design, sev 5).
• Distance metric match — the index metric (cosine / L2 / inner-product) matches how the model was trained; a mismatch silently wrecks recall. Sev-5 only when the model is declared in-repo; else directional and do not cap (per scoring-model honesty rule).
• Index present & tuned — HNSW (m, ef_construction, ef_search) or IVFFlat (lists/probes) declared, not a brute-force seq scan on a large table (performance).
• Model version captured — embeddings are tied to a model version so a re-embed is possible (design).
• Filtered search — metadata used for pre/post-filtering is itself indexed (performance).

M20b — Time-series / OLAP (TimescaleDB, InfluxDB, ClickHouse)

Feeds: design (Hypertable-fit, Retención, Precisión-ts & tz) + performance (Chunk/retención, Continuous-agg, Compresión, Query).

• Hypertable / partition fit — large append-only time data is a hypertable / partitioned by time, with a sane chunk interval (not one giant chunk, not millions of tiny ones).
• Continuous aggregates / rollups declared for dashboard queries instead of scanning raw rows.
• Compression / TTL retention policy present for cold chunks; raw retention bounded.
• Timestamp precision & timezone — timestamptz/UTC, not naive local time (shares the M4 rule).

M20c — Graph (Neo4j)

Feeds: design (Modelado-aristas, Nodos/traversal) + performance (Índice-lookup, Traversal, Supernodo).

• Edge modeling — relationships carry the right direction/type; relationship properties exist where traversals need them; no relationship modeled as a node-pair table that defeats the graph engine.
• Index-backed lookups — entry-point node properties used to start traversals have indexes/constraints (CREATE CONSTRAINT ... IS UNIQUE), so traversals don't begin with a full label scan.
• Supernode risk — a node with a very high relationship degree that will dominate traversal cost (performance). Sev escalation requires live degree evidence; else directional.

M20d — Search (Elasticsearch, OpenSearch)

Feeds: design (mapping/analyzer correctness) + performance (query/shard layout).

• Mapping & analyzer — fields have explicit mappings (not relying on dynamic mapping for analyzed text); analyzer matches language; keyword vs text chosen deliberately.
• No mapping explosion / unbounded dynamic fields; shard count sane for index size.

Tier-0 static checks

Parse declared schema/config: pgvector vector(N) columns + CREATE INDEX ... USING hnsw/ivfflat; Qdrant/Pinecone collection config (size, distance); Timescale create_hypertable/add_*_policy; ClickHouse ENGINE = MergeTree ... ORDER BY; Neo4j constraint/index DDL; ES mappings/settings JSON. Cross-check declared embedding-model dim/metric against any model id referenced in repo code.

Tier-1 verification query

• pgvector: SELECT indexdef FROM pg_indexes WHERE indexdef ILIKE '%hnsw%' OR indexdef ILIKE '%ivfflat%'; and \d <table> for the vector(N) dim.
• Timescale: SELECT * FROM timescaledb_information.hypertables; and ...continuous_aggregates;.
• Neo4j: SHOW INDEXES; SHOW CONSTRAINTS; and per-node degree for supernode (MATCH (n) RETURN ...).
• ES: GET /<index>/_mapping, GET /<index>/_settings. When the engine isn't reachable → needs_api.

Findings

Emit per schema/finding.schema.json. Example ids:

• M20a.embeddings.dim_mismatch — vector(768) but model emits 1536 (fail, severity 5, axis design, confidence established when the model id is in-repo).
• M20a.embeddings.metric_mismatch — index uses L2 but model trained for cosine (fail, severity 5 only if model declared, else warn/directional, axis both).
• M20a.embeddings.no_ann_index — seq scan vector search (warn, severity 4, axis performance).
• M20b.metrics.no_hypertable — large time table not a hypertable (warn, severity 4, axis both).
• M20c.users.unindexed_traversal_root — traversal start property unindexed (warn, severity 3, perf). Each finding: evidence.observed quotes the real DDL/collection config verbatim (secrets redacted); verification.reproduce is a runnable command above using $DATABASE_URL; verification.method is schema_introspect, index_check, ddl_parse, or explain_plan; expected_impact carries {axis, confidence, magnitude, rationale} — banded, never a naked %.

Honesty

• Never assert a recall number or latency for an index configuration — magnitude is banded only.
• Metric/dimension mismatch caps only when the embedding model is declared in-repo (durable fact). Inferred mismatches are directional and never cap. Supernode/hot-partition sev-5 needs live degree evidence; otherwise directional or needs_api, never a silent pass.