db-partitioning-sharding (M16)

Scaling topology is Performance & Scale (axis performance); feeds relational Escala w12 (shared with M17/M2/M9) and the Shard-key / Partición&hot categories in NoSQL profiles. The two failure directions are symmetric: scaling too late (a monster table that should be partitioned) and scaling too early (sharding a 5 GB database that a single node handles trivially).

What it checks

1. Partitioning fit — large append-only / time-series tables (events, logs, metrics) that would benefit from Postgres declarative range/list partitioning (cheap pruning, fast retention drops) but are a single heap.
2. Hot partition / skewed key — a partition or shard key with low cardinality or temporal skew (e.g. partitioning by tenant_id where one tenant is 90% of traffic, or all writes hitting "today's" partition). On wide-column stores an unbounded/hot partition on an event table is severity:5 (perf) with live write-rate evidence — otherwise directional.
3. Premature sharding — application-level sharding / multiple shards introduced with no size or throughput justification, adding cross-shard-query and rebalancing cost for no benefit.

Score / axis

Feeds performance only (relational Escala w12; Shard-key document / Partición&hot KV+WC / Escala vector+graph).

Tier-0 (static)

Detect partitioning DDL (PARTITION BY, partition children), the chosen partition/shard key and its apparent cardinality, large-table candidates from naming/columns (timestamp + high insert intent), and shard-fanout code. Table size, row counts, and per-partition write rate are runtime → needs_api at Tier-0 (never a silent pass).

Tier-1/2 (verification query, Postgres)

-- partition inventory + sizes (Tier-1):
SELECT inhparent::regclass AS parent, inhrelid::regclass AS partition,
       pg_size_pretty(pg_total_relation_size(inhrelid)) AS sz
FROM pg_inherits ORDER BY pg_total_relation_size(inhrelid) DESC;
-- candidate (unpartitioned) large tables:
SELECT relname, pg_size_pretty(pg_total_relation_size(oid)) AS sz, reltuples::bigint AS est_rows
FROM pg_class WHERE relkind='r' ORDER BY pg_total_relation_size(oid) DESC LIMIT 20;

Method schema_introspect / query_stat. A large unpartitioned table confirms a partitioning-fit finding as established; per-partition skew needs Tier-2 write stats — without them, hot-partition is directional and never caps.

Findings

Emit findings per schema/finding.schema.json. Examples:

• M16.events.unpartitioned_large_table — large append-only table as a single heap (severity:3, warn, axis performance, confidence established Tier-1 / needs_api Tier-0, fixable: advisory).
• M16.metrics.hot_partition_today — all writes to the current partition / skewed key (severity:3, warn, directional static / established Tier-2; severity:5 only with live write-rate on a WC event table, fixable: advisory).
• M16.app.premature_sharding — sharding with no size/throughput justification (severity:2, warn, directional, fixable: advisory).

Each finding: evidence.observed quotes the DDL / shard code / catalog row verbatim; verification.reproduce is the query above referencing $DATABASE_URL; expected_impact is banded + confidence-tagged (no naked %).

Honesty

• Partitioning has real overhead (planning time, cross-partition constraints) — recommend it only for tables genuinely large or with retention/pruning needs, never as a default.
• Never fabricate a row count or table size to justify (or refute) partitioning — size claims need Tier-1; absent it, emit needs_api.
• Hot-partition sev-5 requires live write-rate evidence; a static key-cardinality guess stays directional and cannot cap.
• These are architectural changes → advisory (or proposed for a concrete declarative-partition migration), never auto.