Fix join amplification caused by WinMagic subquery rewrites. (#67288)
Retain sliding-window rows until they can be safely evicted and respect window frames when simplifying window functions. (#67329, #67743)
Preserve nullability in DISTINCT window rewrites and cast comparisons, and preserve session guards during common expression extraction. (#67721, #67843, #67863)
Preserve weighted averages when eliminating GROUP BY and correct participation checks in aggregate inference. (#67832, #68175)
Normalize projected COUNT slots before NULL-safety checks. (#68068)
Prevent runtime filters from being pushed below Repeat when the target slot is absent from any grouping set. (#67992)
Reject storage-level MIN and MAX pushdown through unsafe casts. (#67980)
Fix partial TopN pushdown through joins, preserve required slots during TopN lazy materialization, and disable TopN lazy materialization for tables without light schema change. (#68235, #68072, #67928)
Preserve UNION constant rows across rewrites and retain Generate child outputs required by subqueries. (#68172, #68173)
Resolve correlated qualified columns before dereference expressions. (#68062)
Hide internal HAVING helper columns from query output. (#68075)
Preserve duplicate output column positions in table-valued functions. (#67829)
Align constant folding and interval folding with BE execution. (#68071, #68174)
Track user variables in the SQL cache and apply sql_select_limit to materialized CTE results. (#68069, #68077)
Propagate constants out of one-row relations to enable partition pruning. (#67400)
Improve error messages for unsupported subqueries. (#67693)
Fix execution of topn_array and topn_weighted on BOOLEAN inputs while preserving BOOLEAN array result types. (#68409)
Fix candidate loss in topn, topn_array, and topn_weighted when space_expand_rate is zero. Non-positive values now allow unlimited intermediate candidate retention; top_num still limits the final result. (#68545)
Return NULL from TRY_CAST on arithmetic overflow. (#67985)
Fix overflow, batch integer output, and NULL-argument row handling in array_range. (#67689, #68644)
Fix nested array_map lambda captures, appending to constant NULL arrays, and array functions operating on outer-NULL containers with hidden nested data. (#68471, #68442, #68696)
Preserve the leading NULLs of each row in array_cum_sum. (#68605)
Fix Decimal rounding with zero decimal places and rounding overflow at scale boundaries. (#68412, #68415)
Fix LIKE matches with extra trailing characters and avoid using NGram Bloom Filter indexes when patterns may contain escapes. (#68159, #68319)
Fix empty constant regular expression patterns, UTF-8 splitting with an empty regular expression, and extended regular expression fallback for non-constant patterns. (#68179, #68417, #68591)
Fix argument-count validation for json_set, json_insert, and json_replace, and fix constant nullable JSON path evaluation. (#67914, #68525)
Prevent JSON text from being truncated at U+0000 inside a string and preserve NUL bytes in count-by-enum keys. (#68735, #68419)
Preserve NULL propagation in date floor functions and account for microseconds in week floor and ceil operations. (#68423, #68416)
Fix range checks for oversized Unix timestamps and correct the aliases of next_day and previous_day. (#68364, #68542)
Fix odd-byte IPv6 truncation and zero-address IPv4 CIDR handling. (#68414, #68420)
Fix NULL handling in parse_data_size and MAKE_SET. (#68413, #68411)
Fix BIGINT overflow during base conversion in conv. (#68424)
Preserve all samples when merging percentile states, align NULL ordering in complex aggregates, and restore BOOLEAN support in linear_histogram. (#68475, #68422, #68676)
Fix crashes when reading zero-length serialized BITMAP, HLL, or QuantileState cells. (#68497)
Fix historical Iceberg scans after schema evolution, V2 reads across nested schema evolution, and schema binding for historical scans. (#67479, #67690, #68568)
Support fractional timestamps correctly in Iceberg time travel. (#67704)
Fix planning failures for Iceberg VARIANT subpaths and preserve nullability in schema metadata. (#68121, #67856)
Report errors for missing Iceberg snapshot files and invalid name mappings instead of returning incorrect data. (#68143, #68003)
Fix equality-delete key matching after schema type promotion and preserve file partition specifications when tables evolve to an unpartitioned layout. (#68383, #68481)
Run Iceberg execute actions in the corresponding catalog authentication context. (#67695)
Preserve Paimon CAST predicate semantics, estimate row counts from snapshot metadata, and invalidate the SDK cache on metadata refresh. (#67933, #68155, #68238)
Scope Hive file cache entries by table and fix name and row-count cache invalidation after external metadata changes. (#68575, #67417, #68196)
Release Iceberg FileIO and Hudi fsView resources on cache eviction. (#66913)
Align Parquet and ORC timestamp semantics and fix external binary and timestamp mappings and binary value ownership. (#68441, #68532, #68297)
Preserve safe predicate ordering for Parquet VARIANT metadata pruning and enforce a Parquet metadata size limit. (#68144, #68047)
Fix lost and duplicate rows when splitting CSV and JSON files using multi-character line delimiters. (#68600)
Clamp HTTP range reads to the file size and keep S3 response streams usable when an error body exceeds the read buffer. (#67683, #68544)
Fix BE access to AWS in EKS Pod Identity environments. (#67555)
Select Lance vector indexes compatible with the requested distance metric and fix Lance JNI compatibility with glibc 2.17. (#67553, #67862)
Prevent Arrow Flight connections from using the point-query short-circuit path or replaying MySQL SQL cache results. (#67557, #67556)
Release Flight coordinators promptly after query completion and cancel queries when result streams close early. (#67558, #68630)
Deduplicate parallel result endpoints by backend. (#68642)
Validate Arrow result arrays before publishing batches and reject out-of-range timestamps. (#68596, #68604)
Fix year-zero DATE conversion, preserve nested logical type metadata, and return analyzed query schemas and table type metadata. (#68649, #68614, #68631)
Fix MySQL Connector/J cursor fetches hanging on empty results. (#68064)
Prevent snapshot reads from using current materialized views. (#67729)
Fix compensation for invalid partition buckets in partitioned rollup rewrites. (#68067)
Avoid unnecessary full refreshes of asynchronous materialized views when a base table adds a partition without an existing snapshot. (#68249)
Preserve the original statement in asynchronous materialized view refresh tasks. (#68272)
Restore rollup display state after metadata reload and keep rollup jobs waiting when aborting a conflicting transaction fails. (#67596, #68055)
Skip dangling views when querying information_schema.view_dependency and fix quoting of view column names and comments in exported DDL. (#67377, #67398, #68076)
Treat invalid column statistics as UNKNOWN instead of failing the entire ANALYZE job, and fix null-pointer errors when querying a missing ANALYZE job ID. (#68061, #68493)
Commit FE type state only after the transition completes. (#68325)
Treat unparseable or reversed zone map bounds as invalid, preserve every byte of STRING and VARCHAR bounds, and fix NaN handling for legacy FLOAT and DOUBLE zone maps. (#67528, #67550, #67998, #68102)
Fix cumulative point advancement during parallel cumulative compaction and correct compaction thread accounting. (#67425, #67662)
Keep the latest 10 versions unmerged during schema change and release exhausted segment contexts promptly during vertical compaction. (#67589, #67599)
Classify cumulative compaction tasks by row count, avoid repeatedly compacting large rowsets, and include empty rowsets in approximate compaction scores. (#66938, #65601, #66656)
Persist empty inverted-index files and stop writing orphan index files. (#68082)
Preserve VARIANT subpath order and prevent integer element_at indexes from being interpreted as subcolumn paths. (#68284, #68041)
Do not cache failures to load segment row counts in BetaRowset, allowing subsequent retries. (#68683)
Fix concurrent access to the rowset map in the backend tablets scanner. (#68316)
Keep File Cache monitoring off the LRU recorder lock and count evicted bytes only for blocks that were actually downloaded. (#67495, #68157)
Unify File Cache TTL deadlines, restore TTL block types when an expired TTL is extended, and set TTL on cloud schema change output. (#68313, #68092, #67686)
Preserve tablet IDs during synchronous File Cache writes. (#68347)
Trigger periodic cloud checkpoints through edit log rolls. (#67422)
Fix warm-up job scheduling starvation, release destinations on initialization failure, and restore destination locks after FE recovery. (#67643, #68107, #68105)
Avoid conflicts when recycling transactions with the same label and return the existing rowset for concurrent commits. (#67478, #67660)
Exclude stale rowsets from temporary cloud schema change tablets and exclude covered rowsets from compaction minimum timestamps. (#67383, #67751)
Refresh metadata for continuously ingested tablets and repair tablet indexes missing db_id during transaction commit. (#68044, #68127)
Invalidate version caches on visible transaction commit retries. (#68650)
Respect recycle types when deleting partial-update rowsets and preserve clone source mappings for deleted instances. (#68665, #68570)
Fix packed-slice location lifetimes, guard rowset statistics synchronization with task context, and decode transaction IDs safely through Versionstamp. (#67973, #67964, #67954)
Correct the FDB size limit configuration for delete bitmap V2 and fix lookup of prepared transactions by coordinator. (#67938, #68059)
Allocate a fresh StatementContext for each prepared-statement execution to prevent unbounded FE memory growth in long-lived prepared statements. (#67926)
Fix ReadListener leaks when workers reject tasks and prevent concurrent historical profile loaders. (#67990, #67069)
Fix runtime filter RPC callback reuse and memory release, and resolve an AutoReleaseClosure data race. (#67912, #67340)
Fix bvar sampling access to freed thread-local objects, timer callback destruction races, and races between adaptive flush callbacks and thread-pool teardown. (#68045, #68523, #68326)
Initialize the bthread context key before first use and avoid repeated peer connection failures. (#67722, #67647)
Prevent BE crashes on malformed AI adapter responses and correctly extract final text from Responses API output. (#68333, #67795)
Fix double-free errors during PHDR unwinding under ASAN and getentropy crashes on glibc 2.17 under ASAN. (#68229, #67849)
Upgrade libunwind to 1.8.3 to fix aarch64 BE startup crashes on Linux systems with 64 KiB memory pages. (#68339)
Fix dyld cache crashes with large BE binaries on macOS and stop background workers before exiting on startup failure. (#68218, #68653)
Update dependencies including Thrift, AWS SDK, Nimbus, and gRPC, and remove unused LZ4 dependencies. (#67356, #67554, #67847, #68060, #67148, #68557)