[ancongui]

Summary

Starting a run got slower with every run ever started: the usage accounting trigger looked up its reservation by scanning every row of operation_reservations (36–90 ms per inserted run at 524,000 reservations).

Cause. weave_usage_charge passes TG_TABLE_NAME (type name, collation "C") down the call chain, so PL/pgSQL gives every text parameter and variable of weave_reservation_touch, weave_reserve_refresh and weave_allocation_delta collation "C". Compared with the default-collation key columns, that rules out the primary key. auto_explain inside the trigger shows three affected statements:

Function	Statement	Before
weave_reserve_refresh	SELECT amount,outbox_slots FROM operation_reservations WHERE kind=kind_name AND resource_id=identifier	Seq Scan, kind = 'run'::text COLLATE "C"
weave_reserve_refresh	EXISTS(... operation_control_context ... AND kind=kind_name ...)	COLLATE "C" key condition
weave_allocation_delta	SELECT/DELETE ... operation_allocations WHERE relation=relation_name AND row_key=identity_key	Seq Scan, both columns COLLATE "C"

weave_control_begin/weave_control_finish receive literals or table-read kind values and are not affected.

Fix. Forward migration 0031_operations_key_collation patches both shared functions in place (pg_get_functiondef + replace, as in 0021/0025), comparing in the default collation. It fails closed unless each pattern occurs exactly twice; CREATE OR REPLACE keeps owner, grants and SET options. No table, data or role change. SCHEMA_VERSION moves to the new head.

Tests

• New tests/integration/test_reservation_lookup.py: seeds 6,000 reservations and 6,000 allocations in another project, inserts a run through the trigger with auto_explain, and asserts index scans on operation_reservations_pkey / operation_allocations_pkey with both key columns as index conditions, no COLLATE "C" condition on any of the three tables, and unchanged usage accounting (assert_usage before and after).
• Fails without the migration (Seq Scan), passes with it.
• tests/integration/operations_support.py carries recompute_usage/assert_usage verbatim from feat/operations-observability, which is not merged yet; that branch's file is a superset.
• make check: all requested checks passed.
• Integration on real PostgreSQL 17 (Docker context colima-weave-tests): test_reservation_lookup, test_operations, test_large_row_accounting, test_run_lifecycle, test_outbox, test_terminal_capacity, test_boot, test_incident_migration, test_schedule_migration, test_runs, test_waits: 115 passed.

Upgrade note

Operators must run the explicit migration before restarting the server; CHANGELOG (Unreleased) and docs/operations/upgrades.md are updated.

Coordination

feat/operations-observability also adds a 0031_* migration. Whichever merges second renumbers to 0032_* and chains its down_revision.