New Features

AI & Search

• Support ANN indexes on Merge-on-Write tables (#67155)
• Support multimodal file embedding (#66461)

Query & Execution

• Add adaptive global runtime filter tree publishing to support large cluster (#65599)
• Push down limits into CTE producers (#63675)
• Add partition-filter SQL block rules (#62196)

Load & Streaming

• Add adaptive random-bucket load routing (#65450)
• Support forwarding group-commit Stream Load requests (#63594)
• Support compute-group routing and planning for Stream Load (#65571)
• Support OceanBase CDC streaming jobs (#65588)
• Detect PostgreSQL schema changes from Relation events (#64850)
• Support MySQL ADD and DROP schema changes in streaming jobs (#65325)

Cloud Native

• Add MetaService RPC rate limiting in FE (#65694)
• Add memory governance for external metadata caches (#66717)
• Support SHOW COMPUTE GROUPS in non-cloud mode (#66697)
• Support tablet-level compaction through SQL (#66611)
• Add resource group success-quorum checks (#66751)

Lakehouse

• Support altering Iceberg and Paimon table properties (#66428)
• Support Alibaba Cloud OSS Tables REST catalog (#66823)

Security & Authentication

• Add TLS support for internal Doris communication (#64016)
• Add internal HTTP authentication between FE and MetaService (#66090)

Other

• Add queue_time_ms to audit logs (#66641)

Improvement

Query & Execution

• Optimize MD5 with an AVX2 batch path (#63484)
• Improve percentile aggregate performance (#62520)
• Optimize hash join probe-side output performance (#65600)
• Optimize shuffle-key selection for Repeat decomposition (#66532)
• Deduplicate grouping scalar functions in Repeat projections (#65880)
• Infer set-operation distinctness from NDV statistics (#64618)
• Avoid underestimating NOT IN predicate row counts (#66632)
• Add instance IDs to hash join profiles (#66097)
• Add spill-read deserialization timing (#67041)
• Make random shuffle prefer local channels (#59431)

Storage & Compaction

• Add remote-index byte metrics for the file cache (#65398)
• Add queue-time metrics for delete-bitmap tasks (#65523)
• Speed up file-cache LRU restoration at startup (#65174)
• Limit the file-cache LRU recorder shadow queue (#64798)
• Skip redundant TTL scans for non-TTL tablets (#65434)
• Add an option to disable file-cache writes during TopN lazy materialization (#66414)
• Disable file-cache writes after a remote-scan threshold is reached (#66479)
• Improve memory-limit refresh when cgroup settings change in serverless mode (#66463)
• Backport nullable-column and projection optimizations (#66586)

Load & Streaming

• Optimize nullable-string deserialization in CSV and text load (#64476)
• Backpressure asynchronous group commit by per-table WAL count (#65362)

Cloud Native

• Add S3 rate-limit observability (#64038)
• Add warm-up job count metrics on BE (#66136)
• Use rendezvous hashing for colocate-tablet placement (#64638)
• Add MetaService rate-limit dry-run observability (#66969)
• Add cloud-tablet rebalancer metrics (#66576)
• Presize global cloud-tablet route sets (#66447)
• Skip non-MoW tablets when calculating TopN delete-bitmap scores (#65964)

Lakehouse

• Cache credentials providers in the Hadoop S3A glue layer (#65165)
• Use credential-aware Hadoop FileSystem cache keys (#65586)
• Accelerate fixed-binary decimal decoding in Parquet (#66379)
• Improve nullable selection planning for fragmented Parquet batches (#66397)

Bugfix

AI & Search

• Fix crashes caused by null-query inverted-index predicates (#65138)
• Split bound multi-segment inverted-index readers (#63138)
• Fix CLucene multi-segment readBlock handling (#66736)
• Reject inverted-index terms containing embedded NUL bytes (#67179)
• Disable the DSL cache for score queries (#65436)
• Validate nested Variant MATCH predicates (#66207)
• Fix index lookup after indexes are removed (#66316)
• Preserve atomicity when appending nested Variant values (#66421)
• Recognize nullable array elements as NestedGroup types (#66196)
• Prefer sparse Variant fields over root values (#65660)
• Serve shredded Variant leaves without rebuilding the root (#66941)
• Preserve NestedGroup access paths through explode (#67022)
• Avoid recursion when the ANN IVF-list cache is absent (#67024)

Query & Execution

• Fix analytic partition keys overflowing before conversion to 64-bit columns (#65240)
• Guard execution-profile access during concurrent teardown (#65442)
• Fix duplicate join nodes produced by PushDownAggThroughJoinOnPkFk (#65172)
• Guard UniqueFunction handling in filter and TopN pushdown rules (#62742)
• Enable pre-aggregation for multi-argument aggregates with all-key distinct inputs (#65846)
• Fix missing decimal precision and scale for legacy decimal data (#65419)
• Materialize constant columns before block merge (#65770)
• Use serialized hash keys for complex types (#66777)
• Fix blocking-queue waiter accounting (#65827)
• Fix unsafe eager-aggregation nullability conversions (#66208)
• Fix duplicate aggregate functions pushed through projects (#66531)
• Preserve NULL in pushed-down CHAR MIN and MAX (#65952)
• Fix invalid semi-join transpose when the lower join is a mark join (#66574)
• Drop invalid functional dependencies from the outer side of joins (#65982)
• Prevent unsafe CTE runtime-filter pushdown (#65247)
• Suppress invalid unique constraints in scan functional-dependency derivation (#66801)
• Fix TIMESTAMPTZ nullable type equality (#66722)
• Fix TIMESTAMPTZ values in COALESCE (#66689)
• Preserve TIMESTAMPTZ user-variable types (#66833)
• Handle DST fallback in timestamp-cast monotonicity (#65903)
• Accept spaced TIMESTAMPTZ offsets in range-partition bounds (#66292)
• Return Arrow and Thrift DATETIME values as naive timestamps (#67027, #67232)
• Correct the Flight SQL GetTables schema and TIMESTAMPTZ Arrow reader (#66344)
• Fix count_substrings tail overmatching (#63215)
• Avoid ICU default-locale contention in functions (#66440)
• Keep multi-match regex cache ownership alive (#66909)
• Prevent explode_bitmap crashes when output size exceeds INT_MAX (#66034)
• Skip CHAR payload checks for NULL rows (#67043)
• Restore null-literal return-type handling (#66280)
• Stop strict-mode string casts from attaching NULL rows to the next value (#66952)
• Materialize constant virtual columns before caching (#67112)
• Handle case-insensitive auto-partition function arguments (#67121)
• Preserve nested paths during lazy row-ID fetch (#67205)
• Handle empty row-ID fetch RPC failures in TopN (#66443)
• Key UDF class caches by function ID and clean them up in cloud mode (#67046)

Storage & Compaction

• Preserve shortest round-trip representations for FLOAT and DOUBLE ZoneMaps (#65302)
• Fix floating-point equality during Parquet pruning (#66470)
• Fix sparse vertical-compaction destination offsets (#66306)
• Reject transaction preparation on tablets that are shutting down (#66448)
• Clean up spill directories during query teardown (#66458)
• Avoid scan-executor use-after-free during shutdown (#65220)
• Pass tablet IDs explicitly through file-reader and packed-file cache contexts (#61683, #65701)
• Avoid converting segment-cache blocks to index-cache blocks (#65905)
• Fix BE memory leaks after failed schema-change jobs (#56207)
• Fix PublishVersionDaemon thread-pool leaks (#60720)
• Make nullable hash-state updates exception-safe (#66517)
• Synchronize access to load error logs (#66250)
• Fix cache lookup failures after an OLAP partition is dropped (#65889)
• Skip file cache for HTTP chunk-response readers (#66932)

Load & Streaming

• Respect profile levels for load-writer details (#64978)
• Fix Arrow Stream Load with uppercase column names (#65617)
• Fail fast when a backend restarts during INSERT (#65525)
• Skip decommissioning backends during load routing (#65406)
• Correct quorum participants for incremental load streams (#66016)
• Fix stream-receiver leaks racing with close_load (#65584)
• Fix broker-load label self-conflicts during pending retries (#66469)
• Recover Broker Load jobs whose transactions are already visible (#66987)
• Clear transaction IDs after aborting Broker Load transactions (#66884)
• Encode CDC Stream Load records as UTF-8 (#66771)
• Prevent MySQL CDC data loss on non-GTID keepalive reconnects (#66998)
• Isolate schemas across CDC snapshot splits (#65645)
• Keep CDC end offsets consistent with current progress (#65688)
• Normalize MySQL JDBC URLs used by streaming jobs (#65647)
• Avoid applying a global BE authentication precheck to Stream Load (#66629)
• Report Stream Load size limits in MiB (#66224)
• Remove the unsupported Workload Policy action that mutates session variables (#64856)

Cloud Native

• Retry MetaService too busy responses (#65378)
• Set a recycler S3 request timeout to avoid slow DeleteObjects failures (#64758)
• Preserve resource IDs when recycling empty rowsets (#65862)
• Skip packed-slice deletion while recycling rowsets (#65533)
• Fix local and remote tablet-size semantics in schema views (#60887)
• Normalize storage and replica fields in SHOW PARTITIONS (#60871)
• Align colocate proc output and tablet-health reporting in cloud mode (#60944)
• Exclude decommissioning backends from cloud backend selection (#65221)
• Prevent urgent loads from preempting schema-change locks (#66082)
• Skip unnecessary MetaService tablet updates after table-property changes (#65983)
• Prevent older clients from treating unknown MetaService status codes as success (#64148)
• Fail closed on unknown MetaService response codes (#66364)
• Skip rowset promotion for incomplete lazy commits (#66253)
• Track deleted-instance recycling status and retain instance tombstones (#66519, #66870)
• Recycle delete-bitmap packed files with correct blob reads and reference counts (#66728, #66919)
• Only mark prepared rowsets before recycling (#65550)
• Skip versioned delete-bitmap cleanup for non-MoW tablets (#67084)
• Separate HTTP encoding for recycler job and check keys (#67243)
• Drop stale CloudReplica routes after compute groups are deleted (#66984)
• Fix SchemaVariablesScanner crashes on incomplete FE responses (#65994)
• Use a bthread-aware shared mutex for tablet-header locks (#66020)
• Fix schema-change alter-version capture in cloud mode (#62506)
• Restrict runtime FE configuration updates to root in cloud mode (#66478)
• Resolve conflicting default replica properties (#65836)
• Exclude zero-size tablets from backend balancing (#66499)
• Inspect and abort failed conflicting transactions during cloud upgrades (#60830)
• Cancel rebuilt virtual-compute-group warm-up jobs when the group is dropped (#65426)
• Fix rollback handling when adding partitions (#64650)

Lakehouse

• Fix MaxCompute IN predicate pushdown polarity (#65083)
• Improve MaxCompute catalog validation (#64119)
• Preserve external partition metadata in multi-catalog queries (#66012)
• Invalidate stale Hive file-cache entries after partition refresh (#65334)
• Safely publish Hadoop catalog properties (#66392)
• Resolve missing remote JDBC table names during schema lookup (#65718)
• Harden JDBC driver URL validation and remove the file-upload HTTP API (#67149)
• Decode ORC timestamps through the standard SerDe path (#64807)
• Preserve nested nullability in Parquet predicate scans (#66799)
• Safely prune nested Parquet columns with Bloom filters (#66471)
• Fix Iceberg Merge-on-Read crashes with row-ID lazy reads (#66506)
• Avoid BE aborts when evaluating Iceberg partition predicates (#66835)
• Use the split file format when scanning Iceberg tables (#65760)
• Delete external data files after failed writes to prevent orphan files (#64678)
• Preserve Paimon partition metadata across splits (#65581)
• Handle special characters in Paimon partition values (#65904)

Security & Authentication

• Require authorization and an explicit configuration gate for _stream_load_forward (#65377)
• Mask Kafka Routine Load sensitive properties (#64786)
• Hide Stream Load tokens and authentication data from BE information endpoints (#60656, #59743)
• Escape audit-log record separators to prevent row forgery (#66580)
• Mask sensitive fields in encryption-key schema tables (#66834)
• Mask private-key passwords in backend configuration (#66836)
• Mask credentials in authentication and Stream Load logs (#66917)
• Stop writing Arrow Flight bearer tokens to fe.log (#67146)

Materialized Views

• Fix scan-materialization pattern-context checks (#65414)
• Distinguish partition compensation from UNION ALL rewrite (#66445)
• Reduce master FE CPU usage when serializing MTMV task TVF information (#66792)
• Optimize MTMV partition-lineage checks (#63899)
• Avoid invalid slot casts in MV null-reject compensation (#66613)
• Refresh MTMVs after excluded trigger tables change (#64041)

Other

• Fix Arrow Flight SQL direct-memory leaks in prepared statements (#65311)
• Fix task finish timestamps when scheduled jobs fail (#66232)
• Guard query-instance metrics before metric repository initialization (#62762)
• Remove query profiles from HDFS file readers (#67293)