fix(remote): push the source dump instead of the optimizer's own metadata

[veksen]

Goal

Stop a statistics push from replacing a project's production snapshot with an empty one. Found reviewing #197 after it merged.

What

applyStatistics emitted optimizer.ownMetadata. That is a dump of the optimizing database, not the source:

• QueryOptimizer is constructed against optimizingDbUDRL (remote.ts:113), so setStatistics dumps from that connection (query-optimizer.ts:156) and passes the result as plannerTables.
• Statistics.ownMetadata is plannerTables. Core's own comment says snapshot carries the real numbers and passing the same array as both is "rarely what you want".
• The optimizing database is restored with --exclude-table-data-and-children (schema-link.ts:124), and restoreStats only ever runs inside a transaction that rolls back (genalgo.ts:521).

So the pushed payload reports reltuples = -1 for every table and stats: null for every column. relay.service.ts stores it verbatim, overwriting the single project-level snapshot.

Why it matters now

Nothing reached this path before. applyStatistics is only called from updateStatistics (declared in ClientApi, never invoked by the server), resetStats, and the HTTP import.

#197 seeds the drift baseline on connect, which is what first makes the drift path reachable in production. A drifted project would then push the empty dump, baseline itself against reltuples = -1, and on the next poll see every real table as Size Drift — because Math.max(-1, 1) gives a denominator of 1, so any table over the 1,000-row floor drifts. The result is a full DUMP_STATS_SQL against the customer's production database every 60 seconds, indefinitely.

Not yet live: the deployed analyzer predates #195 and #197. Worth holding the analyzer deploy until this merges.

How

Push statsMode.stats — the dump we were handed — and baseline against the same array. Push nothing for fromAssumption, since synthetic defaults are not a project's production statistics; that also closes a smaller existing hole where resetStats pushed the empty dump for any mode, because the emit sat outside the mode check.

Tests

apply-statistics.test.ts drives applyStatistics with a stubbed optimizer whose ownMetadata reports reltuples = -1, and asserts the emitted payload is the source dump with its real row count. A second case asserts fromAssumption emits nothing. Both fail against the previous behaviour.

[github-actions]

Query Doctor Analysis

3 queries analyzed

0 regressed · 0 improved · 0 new · 0 removed

2 pre-existing issues

• SELECT "guests"."id", "guests"."session_id", "guests"."username", "guests"."avatar_path", "guests"."color", "guests"."side", "guests"."audio_recording_path", "guests"."audio_recording_public", "gue...
  index assets(event_id, inserted_at desc)
  cost 31,003,449 → 1,498 (100% reduction)
• SELECT * FROM guest_ip_addresses WHERE ip_address = '127.0.0.1';
  index guest_ip_addresses(ip_address)
  cost 154,402 → 8 (100% reduction)

_{Using assumed statistics (10000000 rows/table). For better results, sync production stats.}

_{More detail → get_ci_run({ runId: "019fa561-54a1-75df-ae3c-2b4f4d46739a" }) · view run · docs}