Skip to main content
See GitHub release: v0.26.1

​
New Features ✨

  • Join Scan now runs the faceted form pdb.agg(...) OVER () (i.e. pdb.agg(...) as a window function) over a join. The document covers every row the join produces, while ORDER BY ... LIMIT still picks the rows returned, so a page of search results across tables carries its facets in the same query. The document can also feed other target list expressions, such as pdb.agg(...) OVER () ->> 'value'. The spec follows the same rules as pdb.agg() over joins; see Limitations for the shapes this path turns down (#6359).
  • Aggregate Scan now reports workers requested and launched, the worker selection method, and estimated parallel costs in EXPLAIN ANALYZE.

​
Performance Improvements 🚀

  • Range-partitioned joins (paradedb.enable_range_partitioned_join) no longer walk every document of a segment that crosses the first partition’s edge to find the rows with a NULL join key. Selective queries on that partition now cost the same as on the others.
  • Join Scan now lazily fetches row ctids during visibility checks and projection. For blocks marked all-visible in PostgreSQL’s visibility map, index fast-field ctid lookups are avoided entirely, reading significantly fewer buffers in queries that filter or sort before projecting rows (#6665).
  • A join of three or more range-partitioned tables now keeps its first join task-local when the two tables share a partition_by key. This removes one network shuffle of the larger table.
  • Small aggregations avoid parallel worker startup when serial execution is cheaper. Worker estimates account for query traversal, aggregation work, and heap visibility checks.
  • Queries made up of term, match, phrase, and match-all clauses use index statistics to estimate matches, reducing planning time. Boolean combinations and score wrappers around these queries use the same path.

​
Stability Improvements 💪

  • Retrying index creation after a rolled-back subtransaction, such as a PL/pgSQL EXCEPTION block or a ROLLBACK TO SAVEPOINT, now persists its tokenizer settings correctly, so the committed index remains usable from new connections.
  • Aggregate Scan now runs a GROUP BY query that filters a grouping key to one value, or that selects a column the primary key decides. These queries fell back to Postgres, and pdb.agg() failed (#6606). This also fixes wrong rows for GROUP BY date(ts) under a window function (#6633), and an error on a GROUP BY key that is a cast (#6632).
  • Aggregate Scan now runs a GROUP BY query with a HAVING condition that has no aggregate, such as a condition on a grouping key, and a grouped subquery with a condition of the outer query on a grouping key. Postgres moves such a condition to WHERE, and the scan declined the query with HAVING clause is not supported. A pdb.agg() query with such a condition failed with the same message (#6636).
  • A search query whose heap filter compares a column to a parameter (lib = $2 in a generic plan, or a nested-loop parameter from a LATERAL join) no longer crashes the backend or fails with unrecognized node type when the scan returns paradedb.score() or paradedb.snippet(), when an aggregate has a FILTER clause, or when EXPLAIN ANALYZE prints a wrapped aggregate (#6492).
  • A date of infinity or -infinity no longer fails index builds, inserts or searches. Infinite date, timestamp and timestamptz constants in a search, and a daterange with an infinite bound, now work too (#6722).