0262 — latest_key trades a second synchronous write per keyed publish for an O(1) latest-per-key read

Context. The correlated “is this the latest row for its key” scan is the ground-truth definition of latest but costs O(partitions since the row) to evaluate, with no early termination for a never-superseded key.

Decision. A latest_key(topic_id, compaction_key, latest_id) table is upserted synchronously in the same transaction as every keyed write: appendMessage’s INSERT gained RETURNING id, followed by INSERT ... ON CONFLICT (topic_id, compaction_key) DO UPDATE SET latest_id = EXCLUDED.latest_id WHERE latest_key.latest_id < EXCLUDED.latest_id. The read path in readMessages/FanOut became m.compaction_key IS NULL OR m.id = (SELECT latest_id FROM latest_key WHERE topic_id = $N AND compaction_key = m.compaction_key) — the old unbounded scan was deleted outright, with no fallback path.

Consequences. The latest-per-key read is O(1) regardless of partition count. Every keyed publish pays a second write in its transaction — measured as noise when uncontended, and a real serialization cost on a single hot key (recorded separately). Unkeyed traffic short-circuits on compaction_key IS NULL and pays nothing.