0685 — A claim poll reads its snapshot before any transaction and opens the reclaim transaction only on an expired lease

Context

Every cursor poll opened two transactions before it knew whether there was anything to do: the reclaim transaction (BEGIN, UPDATE … SKIP LOCKED, ROLLBACK) ran unconditionally ahead of the fresh claim, and the fresh claim opened its own transaction before taking the (head, xmax) snapshot read. An idle poll was six round trips for one read-only statement; a fresh claim was nine.

Measured on a local docker Postgres (RTT ~65µs, 41 partitions, BatchLimit 100, bench/claim): the claim’s SQL executes in ~150µs server-side across all its statements; the rest of a 1120µs claim is round trips, one WAL fsync, and payload transfer. The cursor CTE statement itself runs in 29µs, so its shape is not where time goes.

Decision

• readClaimSnapshot runs first, autocommit, one statement: the visible MAX(id), the snapshot’s xmax, the cursor’s proof columns, and EXISTS (expired lease) as reclaimable.
• The reclaim transaction opens only when reclaimable is true. A reclaim that finds nothing (a peer won under SKIP LOCKED) falls through to the fresh claim with the same snapshot.
• The caught-up short-circuit follows the reclaim branch, outside any transaction. Only then does the fresh claim BEGIN, straight into the cursor statement, taking the snapshot values as parameters.
• The gate CTE lists its three candidate (head, xmax) pairs as a VALUES table under one fence predicate, replacing two CASE arms.

Consequences

• Idle poll: one round trip, no transaction (was six). Fresh claim: six round trips (was nine). Reclaim path: seven (was six; it pays the snapshot read it does not use).
• The fence stays sound: the pair is still read in an earlier statement than the cursor statement, and a read-only autocommit statement can never assign a txid, so the poll’s own transaction can never sit below xmax. Under a REPEATABLE READ default the two statements no longer share one snapshot, which the old shape would have.
• Reclaim detection lags by the snapshot-to-reclaim gap, microseconds; granularity was already once per poll. An expired lease held locked forever keeps reclaimable true and runs the reclaim every poll, the old behavior.
• A missing cursor row now errors before the reclaim branch; before, an expired lease could still be handed out without one. [0387]‘s loud error is the intended behavior.
• Supersedes the “idle poll short-circuits after the snapshot statement” mechanism of [0396] in shape, not in intent: the short-circuit still runs after the read-only snapshot, now with no transaction around it.
• Rejected: bounding MAX(id) by settled_head for partition pruning (the InitPlan becomes a correlated SubPlan, measured slower; the probe already stops at the first partition holding a row) and reading the sequence’s last_value as head (the sequence read happens after the statement’s snapshot, so an id issued in between by a transaction with xid >= xmax escapes the fence). Collapsing the fresh claim to one pipelined batch is parked in ROADMAP with its prototype.

Superseded by 0714: active observations must allocate their own transaction id; snapshot xmax does not bound all already-running producers.