StarRocks Files

Release Date: September 30, 2025

Behavior Change

• Lake internal tablet parallel scan (enable_lake_tablet_internal_parallel) is now enabled by default, increasing per‑query internal parallelism (may raise peak resource usage) #62360

Bug Fixes

The following issues have been fixed:

Data Lake Analytics

• Delta Lake partition column names were forcibly converted to lowercase, causing mismatch with actual column names #62970
• Iceberg manifest cache eviction race could trigger a NullPointerException #63052
• Uncaught generic exceptions during Iceberg scan phase interrupted scan range submission and produced no metrics #63019

Materialized Views (MV)

• Complex multi-layer projected views used in MV rewrite produced invalid plans or missing column statistics #63014 #62230
• Case mismatch of Hive external table MV partition columns was incorrectly rejected #62623
• MV refresh used only the creator’s default role, causing insufficient privilege in “no default role” or LDAP setups (role activation strategy & config introduced) #62461
• Case-insensitive conflicts in list-partitioned MV partition names led to duplicate name errors #62443
• Residual version mapping after failed MV restore caused subsequent incremental refresh to be skipped, returning empty results #62643
• Abnormal partitions after MV recovery caused FE restart NullPointerException #62563
• Non-global aggregation queries incorrectly applied aggregation pushdown rewrite producing invalid plans #63105

Storage / Metadata

• Tablet deletion state was only updated in memory (shutdown) and not persisted, so GC still treated it as running and skipped reclamation #63623 #63620
• Concurrent query plus drop tablet led to early delvec cleanup and “no delete vector found” errors #63307
• Base and cumulative sstable sharing the same max_rss_rowid in PK index compaction were misordered, risking lost delete semantics #63362
• Possible BE crash when LakePersistentIndex destructor ran after a failed initialization #62297
• Graceful shutdown of publish thread pool silently discarded queued tasks without marking failures, creating version holes and a false “all succeeded” impression #62683
• Newly cloned replica on a newly added BE during rebalance was immediately judged redundant and removed, preventing data migration to the new node #62894
• Missing lock when reading tablet max version caused inconsistent replication transaction decisions #62280

Query & Optimization

• Combination of date_trunc equality and raw column range predicate was reduced to a point interval, returning empty result sets (e.g. date_trunc('month', dt)='2025-09-01' AND dt>'2025-09-23') #63570
• Pushdown of non-deterministic predicates (random/time functions) produced inconsistent results #63533
• Missing consumer node after CTE reuse decision produced incomplete execution plans #63188
• Type mismatch crashes when table functions and low-cardinality (dictionary) encoding coexisted #62500 #62384

Ingestion & Export

• Oversized CSV split into parallel fragments caused every fragment to skip header rows, leading to data loss (only the first fragment should skip) #62789
• SHOW CREATE ROUTINE LOAD without explicit DB returned job from another database with the same name #62792
• NullPointerException when sameLabelJobs became null during concurrent load job cleanup #63181

Cluster Operations & Management

• CN normal restart or crash path incorrectly executed scale-in deregistration, harming topology consistency #63002 #63010
• Backend decommission blocked even when all tablets were already in recycle bin (no force completion) #63267
• OPTIMIZE TABLE task stuck in PENDING after thread pool rejection #62556
• Dirty tablet metadata cleanup used GTID arguments in the wrong order #62285