[bitner]

Correctness release (0.1.2). 0.1.1 can return wrong results (missing rows) and can crash the cluster. Every case below was reproduced against 0.1.1 and is covered by a test here. Full list in CHANGELOG.md.

Crashes

• Any planning-time ERROR (e.g. SELECT 1/0) in a backend that had loaded the extension aborted the backend, so the whole cluster restarted. The planner hooks are now #[pg_guard]ed.
• REINDEX of a table_range index crashed the same way: the SPI build planned SQL over a heap whose index was being reindexed.

Wrong results

• Cached plans (PREPARE, PL/pgSQL) kept pruning after an insert widened a summary. Widening now invalidates the heap's relcache entry, not just the index's.
• The per-transaction dedup of invalidations lost them across statements and rolled-back savepoints. Removed: PostgreSQL already coalesces them per command.
• Concurrent inserts could overwrite each other's widening. Now a compare-and-swap under the buffer lock.
• Failed widenings were swallowed. They now raise, or mark the column unbounded (e.g. a value too large for the metapage).
• Partial indexes are now rejected.
• Builds used an MVCC snapshot via SPI. They now use the table AM's index-build scan, which covers rows committed after a REPEATABLE READ snapshot and rows still visible to older snapshots.
• Bounds depended on DateStyle/extra_float_digits. These are now pinned while bounds are written and read.
• Cross-type constants were re-parsed as the column type. Now compared via the operator family's cross-type proc.
• Comparisons under a different collation from the summary's were pruned. They are now kept.

Other

• Builds report the real heap_tuples (0.1.1 set reltuples to 0).
• uuid/boolean columns now build: min()/max() aggregates are no longer used.
• All extension SQL is schema-qualified. This lets geometry builds run under PG17+'s restricted search_path and stops a caller's search_path from substituting functions.
• amsummarizing = true, so updates of indexed columns can stay HOT.
• On-page format v2. v1 summaries are ignored until REINDEX. sql/pg_table_range--0.1.1--0.1.2.sql warns how many indexes need it.

Upgrade

Install, restart PostgreSQL (a backend still running 0.1.1 cannot read v2 summaries and would skip widening them), then ALTER EXTENSION pg_table_range UPDATE, then REINDEX each table_range index.

Testing

• cargo pgrx test: 43/43 on PG18, and on PG16 with PostGIS installed. Not run locally on PG17.
• New tests/multisession.sh (cross-backend invalidation, REPEATABLE READ build, concurrent widening) passes on 0.1.2 and fails on 0.1.1. It is wired into CI; that step has not run on GitHub yet.
• README benchmarks re-run (PG 18.4, release build). table_range planning is within noise of the no-pruning baseline and faster than CHECK constraint exclusion. Insert latency and WAL are unchanged.

Known limits (marked shortcut: in code)

• A widening insert still writes a full-page WAL image (~8 KB).
• Bounds are stored as text.
• Geometry extents go through SQL.
• Under READ COMMITTED, a widening that commits between planning and execution can be missed by that one execution.

🤖 Generated with Claude Code