September 18, 2026 - v3.5.1

[v3.5.1-v6.4.0]⚓︎

Added the optional configuration setting max_concurrent_partitions, which makes the number of Hydrolix partitions the connector reads concurrently configurable for each catalog. The connector previously used a fixed value of 3, and leaving the setting unset keeps that behavior.

Fixed startswith, endswith, and contains pushdown matching more rows than Spark does when the search string contains a SQL LIKE metacharacter such as _ or %. For columns that aren't indexed in every partition, the connector now pushes these predicates down as startsWith, endsWith, and position, which have no wildcard behavior and cost less for each row. Columns indexed in every partition keep LIKE, the only predicate the query engine can use to probe a fulltext index for block skipping, with \, %, and _ in the search string escaped.
Fixed pushed-down aggregations failing at planning time on tables with very large partition counts. Before running a pushed-down aggregation, the connector asks the query head for the output types of the aggregates. Previously, the query head planned every partition in the table to resolve the column types. That request is now wrapped in DESCRIBE, so the query head derives the types from the query header without planning any partitions.