Skip to main content
See GitHub release: v0.26.7

Performance Improvements πŸš€

  • CREATE INDEX with partition_by now keeps segments balanced when a partition column is skewed or has few distinct values. Partitions that can’t be cut further on the declared columns are split by heap position (ctid), and each side of a cut gets partitions in proportion to its share of the rows. A column where 99% of rows share one value used to put almost every row in a single segment; those rows are now spread (roughly) evenly across segments (#7271). Existing indexes will need to REINDEX to take advantage of this. Indexes with ctid specified as one of the partition_by columns will not get the fallback behavior, and will continue to have ctid compete equally with the other columns for cut points.
  • Improved merge backpressure during sustained writes by counting all visible segments, including those already being merged, and performing one eligible foreground merge when the index exceeds four times its target segment count.
  • Top-K joins ordered by combined search scores keep the rank join when a single-table filter is a row-level security policy of any shape, a @@@ query whose score is constant (pdb.range, pdb.term_set, pdb.exists), or any query at ::pdb.const(0). Such filters previously forced a hash join over every permitted row. As for every filter the rank join admits, it scores them 0.0.

Stability Improvements πŸ’ͺ

  • Fixed index builds and merges that produce a fast field column with more than 4 GiB of distinct string values in one segment. They failed with assertion failed: end <= orig_range.end.

Breaking Changes 🚨

  • A row-level security policy no longer adds a constant to pdb.score. Previously a policy added a constant to the score in base scans and hash joins and 0.0 in the rank join, so the same statement returned different scores depending on the plan. Conditions the statement itself writes are unchanged.