3.4.2

Release Date: April 10, 2025

Improvements

• FE supports graceful shutdown to improve system availability. When exiting FE via ./stop_fe.sh -g, FE will first return a 500 status code to the front-end Load Balancer via the /api/health API to indicate that it is preparing to shut down, allowing the Load Balancer to switch to other available FE nodes. Meanwhile, FE will continue to run ongoing queries until they finish or timeout (default timeout: 60 seconds). #56823

Bug Fixes

The following issues have been fixed:

• Partition pruning might not work if the partition column is a generated column. #54543
• Incorrect parameter handling in the concat function could cause a BE crash during query execution. #57522
• The ssl_enable property did not take effect when using Broker Load to load data. #57229
• When NULL values exist, querying subfields of STRUCT-type columns could cause a BE crash. #56496
• When modifying the bucket distribution of a table with the statement ALTER TABLE {table} PARTITIONS (p1, p1) DISTRIBUTED BY ..., specifying duplicate partition names could result in failure to delete internally generated temporary partitions. #57005
• In a shared-data cluster, running SHOW PROC '/current_queries' resulted in the error "Error 1064 (HY000): Sending collect query statistics request fails". #56597
• Running INSERT OVERWRITE loading tasks in parallel caused the error "ConcurrentModificationException: null", resulting in loading failure. #56557
• After upgrading from v2.5.21 to v3.1.17, running multiple Broker Load tasks concurrently could cause exceptions. #56512

Behavior Changes

• The default value of the BE configuration item avro_ignore_union_type_tag has been changed to true, enabling the direct parsing of ["NULL", "STRING"] as STRING type data, which better aligns with typical user requirements. #57553
• The default value of the session variable big_query_profile_threshold has been changed from 0 to 30 (seconds). #57177
• A new FE configuration item enable_mv_refresh_collect_profile has been added to control whether to collect Profile information during materialized view refresh. The default value is false (previously, the system collected Profile by default). #56971

3.3.12

3.3.12

Release date: April 3, 2025

New Features

• Supports the percentile_approx_weighted function. #56654
• Supports modifying properties of Hive Catalog and Hudi Catalog. #56212
• Paimon Catalog supports manifest cache. #55788
• Supports SHOW PARTITIONS for tables in Paimon Catalog. #55785
• Supports statistics collection for Paimon Catalog. #55757

Improvements

• Various improvements and bug fixes related to statistics. #57147 #57238 #57170 #57154 #57124 #57047 #56956 #57031 #56904 #56950 #56671 #55922
• Optimized error messages when table creation fails. #57055
• Enhanced retry mechanism for Broker Load. #56987
• Improved performance of array_generate. #57252
• Aborted ongoing Compaction tasks for deleted partitions. #56943
• Optimized error messages when ALTER TABLE fails. #57054
• Removed unnecessary reverse step from array_agg() to improve performance. #56958
• Added checksum verification for replicas in Primary Key tables. #56519
• Masked sensitive information in the FILES function output. #56684
• Reduced noisy logs related to materialized views. #56672
• Upgraded Iceberg version to 1.7.1. #55271

Bug Fixes

• INSERT INTO FILES did not support CSV delimiter conversion. #57126
• Issues with Iceberg REST Catalog. #55416
• Predicate was lost during rewrite for view-based materialized views. #57153
• Paimon Catalog failed to read tables with schema changes. #56796
• Timezone conversion issue in Paimon Catalog. #56879
• SHOW MATERIALIZED VIEWS did not display default_catalog information. #56362
• In Trino dialect mode, time strings containing 'T' were not accepted. (Solution: replaced parse_datetime with str_to_jodatime.) #56565
• Incorrect result of first_value function. #56467
• Incorrect result of concat_ws function. #56384

Behavior Changes

• Added authentication to the FE Profile interface. #56914
• Changed default value of session variable big_query_profile_threshold from 0 to 30. #56520

3.4.1

Release Date: March 12, 2025

New Features and Enhancements

• Data lake analytics supports Deletion Vector in Delta Lake.
• Supports secure views. By creating a secure view, you can prevent users without the SELECT privilege on the referenced base tables from querying the view (even if they have the SELECT privilege on the view).
• Supports for Sketch HLL (ds_hll_count_distinct). Compared to approx_count_distinct, this function provides higher-precision approximate deduplication.
• Storage Volume in the shared-data clusters supports Azure Data Lake Storage Gen2.
• Supports SSL authentication for connections to StarRocks via the MySQL protocol, ensuring that data transmitted between the client and the StarRocks cluster cannot be read by unauthorized users.

Bug Fixes

The following issues have been fixed:

• An issue where OLAP views affected the materialized view processing logic. #52989
• Write transactions would fail if one replica was not found, regardless of how many replicas had successfully committed. (After the fix, the transaction succeeds as long as the majority replicas succeed. #55212
• Stream Load fails when a node with an Alive status of false was scheduled. #55371
• Files in cluster snapshots were mistakenly deleted. #56338

Behavior Changes

• Graceful shutdown is now enabled by default (previously it was disabled). The default value of the related BE/CN parameter loop_count_wait_fragments_finish has been changed to 2, meaning that the system will wait up to 20 seconds for running queries to complete. #56002

3.3.11

3.3.11

Release date: March 7, 2025

Improvements

• Files supports exporting JSON type data into Parquet files. #56406
• Optimized Data Cache WarmUp performance for cloud-native tables in shared-data clusters. #56190
• Supports parsing AT TIME ZONE expressions and the from_iso8601_timestamp function in Trino. #56311 #55573
• Partial Updates for Primary Key tables within shared-data clusters supports Condition Updates. #56132
• Extended support for statistics collection across all types of SQL statements. #56257
• Supports configuring the maximum number of returned rows for SHOW PROC '/transaction'. #55933
• Supports creating asynchronous materialized views on Oracle-type JDBC Catalog tables. #55372
• MemTracker on BE WebUI supports pagination with 25 rows per page. #56206

Bug Fixes

Fixed the following issues:

• FE does not support casting constant TIME data types into DATETIME. #55804
• Stream Load transaction interface does not support the starrocks_fe_table_load_rows and starrocks_fe_table_load_bytes metrics. #44991
• Changes to automatic statistics collection do not take effect. #56173
• Materialized views in abnormal states caused issues with SHOW MATERIALIZED VIEWS. #55995
• Text-based materialized view rewrite does not work across different databases. #56001
• Metadata compatibility issues in JDBC Catalogs. #55993
• Issues of handling the JSON data type in JDBC Catalogs. #56008
• Incorrect Sort Key settings during Schema Change. #55902
• Credential information leak issue in Broker Load. #55358

Behavior Changes

• Added authentication to the query_detail interface in FE. #55919

3.2.15

Release date: February 14, 2025

New Features

• Window functions support max_by and min_by. #54961
  Improvements
• Added StarClient timeout parameters. #54496
  • star_client_read_timeout_seconds
  • star_client_list_timeout_seconds
  • star_client_write_timeout_seconds
• Tables with List partitioning strategies support partition pruning for DELETE statements. #55400

Bug Fixes

Fixed the following issues:

• Stream Load fails when a node with an Alive status of false was scheduled. #55371
• An error is returned during partial updates on Primary Key tables with Stream Load. #53403
• bRPC error persists after BE node restart. #40229

3.4.0

Release date: January 24, 2025

Data Lake Analytics

• Optimized Iceberg V2 query performance and lowered memory usage by reducing repeated reads of delete-files.
• Supports column mapping for Delta Lake tables, allowing queries against data after Delta Schema Evolution. For more information, see Delta Lake catalog - Feature support.
• Data Cache related improvements:
  • Introduces a Segmented LRU (SLRU) Cache eviction strategy, which significantly defends against cache pollution from occasional large queries, improves cache hit rate, and reduces fluctuations in query performance. In simulated test cases with large queries, SLRU-based query performance can be improved by 70% or even higher. For more information, see Data Cache - Cache replacement policies.
  • Unified the Data Cache instance used in both shared-data architecture and data lake query scenarios to simplify the configuration and improve resource utilization. For more information, see Data Cache.
  • Provides an adaptive I/O strategy optimization for Data Cache, which flexibly routes some query requests to remote storage based on the cache disk's load and performance, thereby enhancing overall access throughput.
• Supports automatic collection of external table statistics through automatic ANALYZE tasks triggered by queries. It can provide more accurate NDV information compared to metadata files, thereby optimizing the query plan and improving query performance. For more information, see Query-triggered collection.
• Provides Time Travel query capability for Iceberg, allowing data to be read from a specified BRANCH or TAG by specifying TIMESTAMP or VERSION.
• Supports asynchronous delivery of query fragments for data lake queries. It avoids the restriction that FE must obtain all files to be queried before BE can execute a query, thus allowing FE to fetch query files and BE to execute queries in parallel, and reducing the overall latency of data lake queries involving a large number of files that are not in the cache. Meanwhile, it reduces the memory load on FE due to caching the file list and improves query stability. (Currently, the optimization for Hudi and Delta Lake is implemented, while the optimization for Iceberg is still under development.)

Performance Improvement and Query Optimization

• [Experimental] Offers a preliminary Query Feedback feature for automatic optimization of slow queries. The system will collect the execution details of slow queries, automatically analyze its query plan for potential opportunities for optimization, and generate a tailored optimization guide for the query. If CBO generates the same bad plan for subsequent identical queries, the system will locally optimize this query plan based on the guide. For more information, see Query Feedback.
• [Experimental] Supports Python UDFs, offering more convenient function customization compared to Java UDFs. For more information, see Python UDF.
• [Experimental] Supports Arrow Flight interface for more efficient reading of large data volumes in query results. It also allows BE, instead of FE, to process the returned results, greatly reducing the pressure on FE. It is especially suitable for business scenarios involving big data analysis and processing, and machine learning.
• Enables the pushdown of multi-column OR predicates, allowing queries with multi-column OR conditions (for example, a = xxx OR b = yyy) to utilize certain column indexes, thus reducing data read volume and improving query performance.
• Optimized TPC-DS query performance by roughly 20% under the TPC-DS 1TB Iceberg dataset. Optimization methods include table pruning and aggregated column pruning using primary and foreign keys, and aggregation pushdown.

Shared-data Enhancements

• Supports Query Cache, aligning the shared-nothing architecture.
• Supports synchronous materialized views, aligning the shared-nothing architecture.

Storage Engine

• Unified all partitioning methods into the expression partitioning and supported multi-level partitioning, where each level can be any expression. For more information, see Expression Partitioning.
• [Preview] Supports all native aggregate functions in Aggregate tables. By introducing a generic aggregate function state storage framework, all native aggregate functions supported by StarRocks can be used to define an Aggregate table.
• Supports vector indexes, enabling fast approximate nearest neighbor searches (ANNS) of large-scale, high-dimensional vectors, which are commonly required in deep learning and machine learning scenarios. Currently, StarRocks supports two types of vector indexes: IVFPQ and HNSW.

Loading

• INSERT OVERWRITE now supports a new semantic - Dynamic Overwrite. When this semantic is enabled, the ingested data will either create new partitions or overwrite existing partitions that correspond to the new data records. Partitions not involved will not be truncated or deleted. This semantic is especially useful when users want to recover data in specific partitions without specifying the partition names. For more information, see Dynamic Overwrite.
• Optimized the data ingestion with INSERT from FILES to replace Broker Load as the preferred loading method:
  • FILES now supports listing files in remote storage, and providing basic statistics of the files. For more information, see FILES - list_files_only.
  • INSERT now supports matching columns by name, which is especially useful when users load data from numerous columns with identical names. (The default behavior matches columns by their position.) For more information, see Match column by name.
  • INSERT supports specifying PROPERTIES, aligning with other loading methods. Users can specify strict_mode, max_filter_ratio, and timeout for INSERT operations to control and behavior and quality of the data ingestion. For more information, see INSERT - PROPERTIES.
  • INSERT from FILES supports pushing down the target table schema check to the Scan stage of FILES to infer a more accurate source data schema. For more information, see see Push down target table schema check.
  • FILES supports unionizing files with different schema. The schema of Parquet and ORC files are unionized based on the column names, and that of CSV files are unionized based on the position (order) of the columns. When there are mismatched columns, users can choose to fill the columns with NULL or return an error by specifying the property fill_mismatch_column_with. For more information, see Union files with different schema.
  • FILES supports inferring the STRUCT type data from Parquet files. (In earlier versions, STRUCT data is inferred as STRING type.) For more information, see Infer STRUCT type from Parquet.
• Supports merging multiple concurrent Stream Load requests into a single transaction and committing data in a batch, thus improving the throughput of real-time data ingestion. It is designed for high-concurrency, small-batch (from KB to tens of MB) real-time loading scenarios. It can reduce the excessive data versions caused by frequent loading operations, resource consumption during Compaction, and IOPS and I/O latency brought by excessive small files.

Others

• Optimized the graceful exit process of BE and CN by accurately displaying the status of BE or CN nodes during a graceful exit as SHUTDOWN.
• Optimized log printing to avoid excessive disk space being occupied.
• Shared-nothing clusters now support backing up and restoring more objects: logical view, external catalog metadata, and partitions created with expression partitioning and list partitioning strategies.
• [Preview] Supports CheckPoint on Follower FE to avoid excessive memory on Leader FE during CheckPoint, thereby improving the stability of Leader FE.

Downgrade Notes

• Clusters can be downgraded from v3.4.0 only to v3.3.9 and later.

3.4.0-RC01

Release date: January 13, 2025

Data Lake Analytics

• Optimized Iceberg V2 query performance and lowered memory usage by reducing repeated reads of delete-files.
• Supports column mapping for Delta Lake tables, allowing queries against data after Delta Schema Evolution. For more information, see Delta Lake catalog - Feature support.
• Data Cache related improvements:
  • Introduces a Segmented LRU (SLRU) Cache eviction strategy, which significantly defends against cache pollution from occasional large queries, improves cache hit rate, and reduces fluctuations in query performance. In simulated test cases with large queries, SLRU-based query performance can be improved by 70% or even higher. For more information, see Data Cache - Cache replacement policies.
  • Unified the Data Cache instance used in both shared-data architecture and data lake query scenarios to simplify the configuration and improve resource utilization. For more information, see Data Cache.
  • Provides an adaptive I/O strategy optimization for Data Cache, which flexibly routes some query requests to remote storage based on the cache disk's load and performance, thereby enhancing overall access throughput.
• Supports automatic collection of external table statistics through automatic ANALYZE tasks triggered by queries. It can provide more accurate NDV information compared to metadata files, thereby optimizing the query plan and improving query performance. For more information, see Query-triggered collection.

Performance Improvement and Query Optimization

• [Experimental] Offers a preliminary Query Feedback feature for automatic optimization of slow queries. The system will collect the execution details of slow queries, automatically analyze its query plan for potential opportunities for optimization, and generate a tailored optimization guide for the query. If CBO generates the same bad plan for subsequent identical queries, the system will locally optimize this query plan based on the guide. For more information, see Query Feedback.
• [Experimental] Supports Python UDFs, offering more convenient function customization compared to Java UDFs. For more information, see Python UDF.

Shared-data Enhancements

• Supports Query Cache, aligning the shared-nothing architecture.
• Supports synchronous materialized views, aligning the shared-nothing architecture.

Storage Engine

• Unified all partitioning methods into the expression partitioning and supported multi-level partitioning, where each level can be any expression. For more information, see Expression Partitioning.

Loading

• INSERT OVERWRITE now supports a new semantic - Dynamic Overwrite. When this semantic is enabled, the ingested data will either create new partitions or overwrite existing partitions that correspond to the new data records. Partitions not involved will not be truncated or deleted. This semantic is especially useful when users want to recover data in specific partitions without specifying the partition names. For more information, see Dynamic Overwrite.
• Optimized the data ingestion with INSERT from FILES to replace Broker Load as the preferred loading method:
  • FILES now supports listing files in remote storage, and providing basic statistics of the files. For more information, see FILES - list_files_only.
  • INSERT now supports matching columns by name, which is especially useful when users load data from numerous columns with identical names. (The default behavior matches columns by their position.) For more information, see Match column by name.
  • INSERT supports specifying PROPERTIES, aligning with other loading methods. Users can specify strict_mode, max_filter_ratio, and timeout for INSERT operations to control and behavior and quality of the data ingestion. For more information, see INSERT - PROPERTIES.
  • INSERT from FILES supports pushing down the target table schema check to the Scan stage of FILES to infer a more accurate source data schema. For more information, see see Push down target table schema check.
  • FILES supports unionizing files with different schema. The schema of Parquet and ORC files are unionized based on the column names, and that of CSV files are unionized based on the position (order) of the columns. When there are mismatched columns, users can choose to fill the columns with NULL or return an error by specifying the property fill_mismatch_column_with. For more information, see Union files with different schema.
  • FILES supports inferring the STRUCT type data from Parquet files. (In earlier versions, STRUCT data is inferred as STRING type.) For more information, see Infer STRUCT type from Parquet.

Others

• Optimized the graceful exit process of BE and CN by accurately displaying the status of BE or CN nodes during a graceful exit as SHUTDOWN.
• Optimized log printing to avoid excessive disk space being occupied.

Downgrade Notes

• Clusters can be downgraded from v3.4.0 only to v3.3.9 and later.

3.3.9

3.3.9

Release date: January 12, 2025

New Features

• Supports the translation of Trino SQL into StarRocks SQL. #54185

Improvements

• Corrected FE node names starting with bdbje_reset_election_group to enhance clarity. #54399
• Implemented vectorization for the IF function on ARM architectures. #53093
• ALTER SYSTEM CREATE IMAGE supports creating an image for StarManager. #54370
• Supports deleting cloud-native indexes of Primary Key tables in shared-data clusters. #53971
• Enforced the refresh of materialized views when the FORCE keyword is specified. #52081
• Supports specifying hints in CACHE SELECT. #54697
• Supports loading compressed CSV files using the FILES() function. Supported compression formats include gzip, bz2, lz4, deflate, and zstd. #54626
• Supports assigning multiple values to the same column in an UPDATE statement. #54534

Bug Fixes

Fixed the following issues:

• Unexpected errors when refreshing materialized views built on JDBC catalogs. #54487
• Instability in results when a Delta Lake table joins itself. #54473
• Upload retries fail when backing up data to HDFS. #53679
• BFD initialization errors on the aarch64 architecture. #54372
• Sensitive information recorded in BE logs. #54677
• Errors in Compaction-related metrics in profiles. #54678
• BE crashes caused by creating tables with nested TIME types. #54601
• Query plan errors for LIMIT queries with subquery TOP-N. #54507

Downgrade notes

• Clusters can be downgraded from v3.3.9 only to v3.2.11 and later.

3.2.14

3.2.14

Release date: January 8, 2025

Improvements

• Supports collecting statistics of Paimon tables. #52858
• Included node information and histogram metrics in JSON metrics. #53735

Bug Fixes

Fixed the following issues:

• The score of the Primary Key table index was not updated in the Commit phase. #41737
• Incorrect execution plans for max(count(distinct)) when low-cardinality optimization is enabled. #53403
• When the List partition column has NULL values, queries against the Min/Max value of the partition column will lead to incorrect partition pruning. #53235
• Upload retries fail when backing up data to HDFS. #53679

3.3.8

3.3.8

Release date: January 3, 2025

Improvements

• Added a cluster idle API to assist in determining cluster status. #53850
• Included node information and histogram metrics in JSON metrics. #53735
• Optimized the MemTable for Primary Key tables in shared-data clusters. #54178
• Optimized memory usage and statistics for Primary Key tables in shared-data clusters. #54358
• Introduced a limit on the number of partitions scanned per node for queries requiring full-table or large-scale partition scans, enhancing system stability by reducing scanning pressure on individual BE or CN nodes. #53747
• Supports collecting statistics of Paimon tables. #52858
• Supports configuration of S3 client request timeout for shared-data clusters. #54211

Bug Fixes

Fixed the following issues:

• BE crashes caused by inconsistencies in the DelVec of Primary Key tables. #53460
• Issues with lock release of Primary Key tables in shared-data clusters. #53878
• Errors of UDFs nested in functions are not returned in query failures. #44297
• Transactions are blocked at the Decommission phase because they depend on the original replicas. #49349
• Queries against Delta Lake tables use relative paths instead of filenames for file retrieval. #53949
• An error is returned when querying Delta Lake Shallow Clone tables. #54044
• Case sensitivity issues when reading Paimon using JNI. #54041
• An error is returned during INSERT OVERWRITE operations on Hive tables created in Hive. #53792
• SHOW TABLE STATUS command does not validate view privileges. #53811
• Missing FE metrics. #53058
• Memory leaks in INSERT tasks. #53809
• Concurrency issues caused by missing write locks in replication tasks. #54061
• partition_ttl of tables in the statistics database does not take effect. #54398
• Query Cache-related issues:
  • Crashes when Query Cache is enabled with Group Execution. #54363
  • Runtime Filter crashes. #54305
• Issues with materialized view Union Rewrite. #54293
• Missing padding in string updates for partial updates in Primary Key tables. #54182
• Incorrect execution plans for max(count(distinct)) when low-cardinality optimization is enabled. #53403
• Issues with changing the excluded_refresh_tables parameter of materialized views. #53394

Behavior Changes

• Changed the default value of persistent_index_type for Primary Key tables in shared-data clusters to CLOUD_NATIVE, that is, enabled Persistent Index by default. #52209