Bound PostgreSQL KV expiry housekeeping

CHANGES.md

@@ -8,6 +8,28 @@ Version 2.0.35
 
 To be released.
 
+### @fedify/postgres
+
+ -  `PostgresKvStore` no longer scans the entire table for expired entries
+    after every write.  Cleanup now deletes up to 1,000 rows per batch,
+    including across partitioned or inherited tables, with a delay between
+    attempts that shortens after a full batch.  Store
+    instances sharing a table do not sweep it concurrently. With a pooled
+    client, writes no longer wait for cleanup, and a cleanup failure no longer
+    fails the write. Caller-owned transactions and reserved connections still
+    await cleanup and propagate its errors to avoid hiding a rollback.  Reads
+    continue to exclude expired entries before physical deletion.
+    [[#1308], [#1314]]
+ -  Fixed `PostgresKvStore.set()` retaining an old creation time when
+    replacing an existing key.  TTLs are now measured from the SQL statement
+    clock on both insert and update, and expiration checks use that same clock
+    inside caller-owned transactions. Replacing an expired value works before
+    its row is physically deleted, and a replacement written with a zero TTL
+    still reads as absent.  [[#1308], [#1314]]
+
+[#1308]: https://github.com/fedify-dev/fedify/issues/1308
+[#1314]: https://github.com/fedify-dev/fedify/pull/1314
+
 ### @fedify/testing
 
  -  Fixed missing runtime dependencies in the npm package.  `@fedify/vocab`

changes.d/postgres/postgres-kv-expiry.md

@@ -0,0 +1,21 @@
+---
+links:
+  '#1308': https://github.com/fedify-dev/fedify/issues/1308
+  '#1314': https://github.com/fedify-dev/fedify/pull/1314
+---
+ -  `PostgresKvStore` no longer scans the entire table for expired entries
+    after every write.  Cleanup now deletes up to 1,000 rows per batch,
+    including across partitioned or inherited tables, with a delay between
+    attempts that shortens after a full batch.  Store
+    instances sharing a table do not sweep it concurrently. With a pooled
+    client, writes no longer wait for cleanup, and a cleanup failure no longer
+    fails the write. Caller-owned transactions and reserved connections still
+    await cleanup and propagate its errors to avoid hiding a rollback.  Reads
+    continue to exclude expired entries before physical deletion.
+    [[#1308], [#1314]]
+ -  Fixed `PostgresKvStore.set()` retaining an old creation time when
+    replacing an existing key.  TTLs are now measured from the SQL statement
+    clock on both insert and update, and expiration checks use that same clock
+    inside caller-owned transactions. Replacing an expired value works before
+    its row is physically deleted, and a replacement written with a zero TTL
+    still reads as absent.  [[#1308], [#1314]]