Skip to main content

StarRocks version 3.3

3.3.1 (Yanked)​

Release date: July 18, 2024

tip

This version has been taken offline due to compatibility issues in Primary Key tables.

  • Problem: After the cluster is upgraded from versions earlier than v3.1.11 and v3.2.4 to v3.3.1, index decompression failures will lead to failures of queries against Primary Key tables.

  • Impact scope: This problem only affects queries against Primary Key tables.

  • Temporary workaround: You can downgrade the cluster to v3.3.0 or earlier to avoid this issue. It will be fixed in v3.3.2.

New Features​

  • [Preview] Supports temporary tables.
  • [Preview] JDBC Catalog supports Oracle and SQL Server.
  • [Preview] Unified Catalog supports Kudu.
  • INSERT INTO on Primary Key tables supports Partial Updates by specifying the column list.
  • User-defined variables support the ARRAY type. #42631
  • Stream Load supports converting JSON-type data and loading it into columns of STRUCT/MAP/ARRAY types. #45406
  • Supports global dictionary cache.
  • Supports deleting partitions in batch. #44744
  • Supports queries on Iceberg views. #46273
  • Supports managing column-level permissions in Apache Ranger. (Column-level permissions for materialized views and views must be set under the table object.) #47702
  • Supports Partial Updates in Column mode For Primary Key tables in shared-data clusters. #46516
  • Stream Load supports data compression during transmission, reducing network bandwidth overhead. Users can specify different compression algorithms using parameters compression and Content-Encoding. Supported compression algorithms including GZIP, BZIP2, LZ4_FRAME, and ZSTD. #43732

Improvements​

  • Optimized the IdChain hashcode implementation to reduce the FE restart time. #47599
  • Improved error messages for the csv.trim_space parameter in the FILES() function, checking for illegal characters and providing reasonable prompts. #44740
  • Stream Load supports using \t and \n as row and column delimiters. Users do not need to convert them to their hexadecimal ASCII codes. #47302

Bug Fixes​

Fixed the following issues:

  • Schema Change failures due to file location changes caused by Tablet migration during the Schema Change process. #45517
  • Cross-cluster Data Migration Tool fails to create tables in the target cluster due to control characters such as \, \r in the default values of fields. #47861
  • Persistent bRPC failures after BE restarts. #40229
  • The user_admin role can change the root password using the ALTER USER command. #47801
  • Primary key index write failures cause data write errors. #48045

Behavior Changes​

  • Intermediate result spilling is enabled by default when sinking data to Hive and Iceberg. #47118
  • Changed the default value of the BE configuration item max_cumulative_compaction_num_singleton_deltas to 500. #47621
  • When users create a partitioned table without specifying the bucket number, if the number of partitions exceeds 5, the rule for setting the bucket count is changed to max(2*BE or CN count, bucket number calculated based on the largest historical partition data volume). The previous rule was to calculate the bucket number based on the largest historical partition data volume). #47949
  • Specifying a column list in the INSERT INTO statement on a Primary Key table will perform Partial Updates instead of Full Upsert in earlier versions.

Downgrade notes​

To downgrade a cluster from v3.3.1 or later to v3.2, users must clean all temporary tables in the cluster by following these steps:

  1. Disallow users to create new temporary tables:

    ADMIN SET FRONTEND CONFIG("enable_experimental_temporary_table"="false");

  2. Check if there are any temporary tables in the cluster:

    SELECT * FROM information_schema.temp_tables;

  3. If there are temporary tables in the system, clean them up using the following command (the SYSTEM-level OPERATE privilege is required):

    CLEAN TEMPORARY TABLE ON SESSION 'session';

3.3.0​

Release date: June 21, 2024

New Features and Improvements​

Shared-data Cluster​

  • Optimized the performance of Schema Evolution in shared-data clusters, reducing the time consumption of DDL changes to a sub-second level. For more information, see Schema Evolution.

  • To satisfy the requirement for data migration from shared-nothing clusters to shared-data clusters, the community officially released the StarRocks Data Migration Tool. It can also be used for data synchronization and disaster recovery between shared-nothing clusters.

  • [Preview] AWS Express One Zone Storage can be used as storage volumes, significantly improving read and write performance. For more information, see CREATE STORAGE VOLUME.

  • Optimized the garbage collection (GC) mechanism in shared-data clusters. Supports manual compaction for data in object storage. For more information, see Manual Compaction.

  • Optimized the Publish execution of Compaction transactions for Primary Key tables in shared-data clusters, reducing I/O and memory overhead by avoiding reading primary key indexes.

  • Supports Internal Parallel Scan within tablets. This optimizes query performance in scenarios where there are very few buckets in the table, which limits query parallelism to the number of tablets. Users can enable the Parallel Scan feature by setting the following system variables:

    SET GLOBAL enable_lake_tablet_internal_parallel = true;
    SET GLOBAL tablet_internal_parallel_mode = "force_split";

Data Lake Analytics​

  • Data Cache enhancements
    • Added the Data Cache Warmup command CACHE SELECT to fetch hotspot data from data lakes, which speeds up queries and minimizes resource usage. CACHE SELECT can work with SUBMIT TASK to achieve periodic cache warmup. This feature supports both tables in external catalogs and internal tables in shared-data clusters.
    • Added metrics and monitoring methods to enhance the observability of Data Cache.
  • Parquet reader performance enhancements
    • Optimized Page Index, significantly reducing the data scan size.
    • Reduced the occurrence of reading unnecessary pages when Page Index is used.
    • Uses SIMD to accelerate the computation to determine whether data rows are empty.
  • ORC reader performance enhancements
    • Uses column ID for predicate pushdown to read ORC files after Schema Change.
    • Optimized the processing logic for ORC tiny stripes.
  • Iceberg table format enhancements
    • Significantly improved the metadata access performance of the Iceberg Catalog by refactoring the parallel Scan logic. Resolved the single-threaded I/O bottleneck in the native Iceberg SDK when handling large volumes of metadata files. As a result, queries with metadata bottlenecks now experience more than a 10-fold performance increase.
    • Queries on Parquet-formatted Iceberg v2 tables support equality deletes.
  • [Experimental] Paimon Catalog enhancements
    • Materialized views created based on the Paimon external tables now support automatic query rewriting.
    • Optimized Scan Range scheduling for queries against the Paimon Catalog, improving I/O concurrency.
    • Support for querying Paimon system tables.
    • Paimon external tables now support DELETE Vectors, enhancing query efficiency in update and delete scenarios.
  • Enhancements in collecting external table statistics
    • ANALYZE TABLE can be used to collect histograms of external tables, which helps prevent data skews.
    • Supports collecting statistics of STRUCT subfields.
  • Table sink enhancements
    • The performance of the Sink operator is doubled compared to Trino.
    • Data can be sunk to Textfile- and ORC-formatted tables in Hive catalogs and storage systems such as HDFS and cloud storage like AWS S3.
  • [Preview] Supports Alibaba Cloud MaxCompute catalogs, with which you can query data from MaxCompute without ingestion and directly transform and load the data from MaxCompute by using INSERT INTO.
  • [Experimental] Supports ClickHouse Catalog.
  • [Experimental] Supports Kudu Catalog.

Performance Improvement and Query Optimization​

  • Optimized performance on ARM.
    • Significantly optimized performance for ARM architecture instruction sets. Performance tests under AWS Graviton instances showed that the ARM architecture was 11% faster than the x86 architecture in the SSB 100G test, 39% faster in the Clickbench test, 13% faster in the TPC-H 100G test, and 35% faster in the TPC-DS 100G test.
  • Spill to Disk is in GA. Optimized the memory usage of complex queries and improved spill scheduling, allowing large queries to run stably without OOM.
  • [Preview] Supports spilling intermediate results to object storage.
  • Supports more indexes.
  • Improved the performance and memory usage of Bitmap functions. Added the capability to export Bitmap data to Hive by using Hive Bitmap UDFs.
  • [Preview] Supports Flat JSON. This feature automatically detects JSON data during data loading, extracts common fields from the JSON data, and stores these fields in a columnar manner. This improves JSON query performance, comparable to querying STRUCT data.
  • [Preview] Optimized global dictionary. Provides a dictionary object to store the mapping of key-value pairs from a dictionary table in the BE memory. A new dictionary_get() function is now used to directly query the dictionary object in the BE memory, accelerating the speed of querying the dictionary table compared to using the dict_mapping() function. Furthermore, the dictionary object can also serve as a dimension table. Dimension values can be obtained by directly querying the dictionary object using dictionary_get(), resulting in faster query speeds than the original method of performing JOIN operations on the dimension table to obtain dimension values.
  • [Preview] Supports Colocate Group Execution. Significantly reduces memory usage for executing Join and Agg operators on the colocate tables, which ensures that large queries can be executed more stably.
  • Optimized the performance of CodeGen. JIT is enabled by default, which achieves a 5X performance improvement for complex expression calculations.
  • Supports using vectorization technology to implement regular expression matching, which reduces the CPU consumption of the regexp_replace function.
  • Optimized Broadcast Join so that the Broadcast Join operation can be terminated in advance when the right table is empty.
  • Optimized Shuffle Join in scenarios of data skew to prevent OOM.
  • When an aggregate query contains Limit, multiple Pipeline threads can share the Limit condition to prevent compute resource consumption.

Storage Optimization and Cluster Management​

  • Enhanced flexibility of range partitioning. Three time functions can be used as partitioning columns. These functions convert timestamps or strings in the partitioning columns into date values and then the data can be partitioned based on the converted date values.
  • FE memory observability. Provides detailed memory usage metrics for each module within the FE to better manage resources.
  • Optimized metadata locks in FE. Provides Lock manager to achieve centralized management for metadata locks in FE. For example, it can refine the granularity of metadata lock from the database level to the table level, which improves load and query concurrency. In a scenario of 100 concurrent load jobs on a small dataset, the load time can be reduced by 35%.
  • Supports adding labels on BEs. Supports adding labels on BEs based on information such as the racks and data centers where BEs are located. It ensures even data distribution among racks and data centers, and facilitates disaster recovery in case of power failures in certain racks or faults in data centers.
  • Optimized the sort key. Duplicate Key tables, Aggregate tables, and Unique Key tables all support specifying sort keys through the ORDER BY clause.
  • [Experimental] Optimized the storage efficiency of non-string scalar data. This type of data supports dictionary encoding, reducing storage space usage by 12%.
  • Supports size-tiered compaction for Primary Key tables. Reduces write I/O and memory overhead during compaction. This improvement is supported in both shared-data and shared-nothing clusters. You can use the BE configuration item enable_pk_size_tiered_compaction_strategy to control whether to enable this feature (enabled by default).
  • Optimized read I/O for persistent indexes in Primary Key tables. Supports reading persistent indexes by a smaller granularity (page) and improves the persistent index's bloom filter. This improvement is supported in both shared-data and shared-nothing clusters.
  • Supports for IPv6. StarRocks now supports deployment on IPv6 networks.

Materialized Views​

  • Supports view-based query rewrite. With this feature enabled, queries against views can be rewritten to materialized views created upon those views. For more information, see View-based materialized view rewrite.
  • Supports text-based query rewrite. With this feature enabled, queries (or their sub-queries) that have the same abstract syntax trees (AST) as the materialized views can be transparently rewritten. For more information, see Text-based materialized view rewrite.
  • [Preview] Supports setting transparent rewrite mode for queries directly against the materialized view. When the transparent_mv_rewrite_mode property is enabled, StarRocks will automatically rewrite queries to materialized views. It will merge data from refreshed materialized view partitions with the raw data corresponding to the unrefreshed partitions using an automatic UNION operation. This mode is suitable for modeling scenarios where data consistency must be maintained while also aiming to control refresh frequency and reduce refresh costs. For more information, see CREATE MATERIALIZED VIEW.
  • Supports aggregation pushdown for materialized view query rewrite: When the enable_materialized_view_agg_pushdown_rewrite variable is enabled, users can use single-table asynchronous materialized views with Aggregation Rollup to accelerate multi-table join scenarios. Aggregate functions will be pushed down to the Scan Operator during query execution and rewritten by the materialized view before the Join Operator is executed, significantly improving query efficiency. For more information, see Aggregation pushdown.
  • Supports a new property to control materialized view rewrite. Users can set the enable_query_rewrite property to false to disable query rewrite based on a specific materialized view, reducing query rewrite overhead. If a materialized view is used only for direct query after modeling and not for query rewrite, users can disable query rewrite for this materialized view. For more information, see CREATE MATERIALIZED VIEW.
  • Optimized the cost of materialized view rewrite. Supports specifying the number of candidate materialized views and enhanced the filter algorithms. Introduced materialized view plan cache to reduce the time consumption of the Optimizer at the query rewrite phase. For more information, see cbo_materialized_view_rewrite_related_mvs_limit.
  • Optimized materialized views created upon Iceberg catalogs. Materialized views based on Iceberg catalogs now support incremental refresh triggered by partition updates and partition alignment for Iceberg tables using Partition Transforms. For more information, see Data lake query acceleration with materialized views.
  • Enhanced the observability of materialized views. Improved the monitoring and management of materialized views for better system insights. For more information, see Metrics for asynchronous materialized views.
  • Improved the efficiency of large-scale materialized view refresh. Supports global FIFO scheduling, optimized the cascading refresh strategy for nested materialized views, and fixed some issues that occur in high-frequency refresh scenarios.
  • Supports refresh triggered by multiple fact tables. Materialized views created upon multiple fact tables now support partition-level incremental refresh when data in any of the fact tables is updated, increasing data management flexibility. For more information, see Align partitions with multiple base tables.

SQL Functions​

  • DATETIME fields support microsecond precision. The new time unit is supported in related time functions and during data loading.
  • Added the following functions:

Ecosystem Support​

  • [Experimental] Provides ClickHouse SQL Rewriter, a new tool for converting the syntax in ClickHouse to the syntax in StarRocks.
  • The Flink connector v1.2.9 provided by StarRocks is integrated with the Flink CDC 3.0 framework, which can build a streaming ELT pipeline from CDC data sources to StarRocks. The pipeline can synchronize the entire database, sharded tables, and schema changes in the sources to StarRocks. For more information, see Synchronize data with Flink CDC 3.0 (with schema change supported).

Behavior and Parameter Changes​

Table Creation and Data Distribution​

  • Users must specify Distribution Key when creating a colocate table using CTAS. #45537
  • When users create a non-partitioned table without specifying the bucket number, the minimum bucket number the system sets for the table is 16 (instead of 2 based on the formula 2*BE or CN count). If users want to set a smaller bucket number when creating a small table, they must set it explicitly. #47005

Loading and Unloading​

  • __op is reserved by StarRocks for special purposes and creating columns with names prefixed by __op is forbidden by default. You can allow this such name format by setting FE configuration allow_system_reserved_names to true. Please note that creating such columns in Primary Key tables may result in undefined behaviors. #46239
  • During Routine Load jobs, if the time duration that StarRocks cannot consume data exceeds the threshold specified in the FE configuration routine_load_unstable_threshold_second (Default value is 3600, that is one hour), the status of the job will become UNSTABLE, but the job will continue. #36222
  • The default value of the FE configuration enable_automatic_bucket is changed from false to true. When this item is set to true, the system will automatically set bucket_size for newly created tables, thus enabling automatic bucketing, which is the optimized random bucketing feature. However, in v3.2, setting enable_automatic_bucket to true will take effect. Instead, the system only enables automatic bucketing when bucket_size is specified. This will prevent risks when users downgrade StarRocks from v3.3 to v3.2.

Query and Semi-structured Data​

  • When a single query is executed within the Pipeline framework, the memory limit is no longer restricted by exec_mem_limit but is only limited by query_mem_limit. A value of 0 for query_mem_limit indicates no limit. #34120
  • NULL values in JSON is treated as SQL NULL values when they are executed by IS NULL and IS NOT NULL operators. For example, parse_json('{"a": null}') -> 'a' IS NULL returns 1, and parse_json('{"a": null}') -> 'a' IS NOT NULL returns 0. #42765 #42909
  • A new session variable cbo_decimal_cast_string_strict is added to control how CBO converts data from the DECIMAL type to the STRING type. If this variable is set to true, the logic built in v2.5.x and later versions prevails and the system implements strict conversion (namely, the system truncates the generated string and fills 0s based on the scale length). If this variable is set to false, the logic built in versions earlier than v2.5.x prevails and the system processes all valid digits to generate a string. The default value is true. #34208
  • The default value of cbo_eq_base_type is changed from varchar to decimal, indicating that the system will compare the DECIMAL-type data with strings as numerical values instead of strings. #43443

Others​

  • The default value of the materialized view property partition_refresh_num has been changed from -1 to 1. When a partitioned materialized view needs to be refreshed, instead of refreshing all partitions in a single task, the new behavior will incrementally refresh one partition at a time. This change is intended to prevent excessive resource consumption caused by the original behavior. The default behavior can be adjusted using the FE configuration default_mv_partition_refresh_number.
  • Originally, the database consistency checker was scheduled based on GMT+8 time zone. Database consistency checker is scheduled based on the local time zone now. #45748
  • By default, Data Cache is enabled to accelerate data lake queries. Users can manually disable it by executing SET enable_scan_datacache = false.
  • If users want to re-use the cached data in Data Cache after downgrading a shared-data cluster from v3.3 to v3.2.8 and earlier, they need to manually rename the Blockfile in the directory starlet_cache by changing the file name format from blockfile_{n}.{version} to blockfile_{n}, that is, to remove the suffix of version information. For more information, refer to the Data Cache Usage Notes. v3.2.9 and later versions are compatible with the file name format in v3.3, so users do not need to perform this operation manually.
  • Supports dynamically modifying FE parameter sys_log_level. #45062
  • The default value of the Hive Catalog property metastore_cache_refresh_interval_sec is changed from 7200 (two hours) to 60 (one minute). #46681

Bug Fixes​

Fixed the following issues:

  • Query results are incorrect when queries are rewritten to materialized views created by using UNION ALL. #42949
  • Extra columns are read when queries with predicates are rewritten to materialized views during query execution. #45272
  • The results of functions next_day and previous_day are incorrect. #45343
  • Schema change fails because of replica migration. #45384
  • Restoring a table with full-text inverted index causes BEs to crash. #45010
  • Duplicate data rows are returned when an Iceberg catalog is used to query data. #44753
  • Low cardinality dictionary optimization does not take effect on ARRAY<VARCHAR>-type columns in Aggregate tables. #44702
  • Query results are incorrect when queries are rewritten to materialized views created by using UNION ALL. #42949
  • If BEs are compiled with ASAN, BEs crash when the cluster is started and the be.warning log shows dict_func_expr == nullptr. #44551
  • Query results are incorrect when aggregate queries are performed on single-replica tables. #43223
  • View Delta Join rewrite fails. #43788
  • BEs crash after the column type is modified from VARCHAR to DECIMAL. #44406
  • When a table with List partitioning is queried by using a not-equal operator, partitions are incorrectly pruned, resulting in wrong query results. #42907
  • Leader FE's heap size increases quickly as many Stream Load jobs using non-transactional interface finishes. #43715

Downgrade notes​

To downgrade a cluster from v3.3.0 or later to v3.2, users must follow these steps:

  1. Ensure that all ALTER TABLE SCHEMA CHANGE transactions initiated in the v3.3 cluster are either completed or canceled before downgrading.

  2. Clear all transaction history by executing the following command:

    ADMIN SET FRONTEND CONFIG ("history_job_keep_max_second" = "0");

  3. Verify that there are no remaining historical records by running the following command:

    SHOW PROC '/jobs/<db>/schema_change';

  4. If you want to downgrade the cluster to a patch version earlier than v3.2.8 or v3.1.14, you must drop all asynchronous materialized views you have created using PROPERTIES('compression' = 'lz4').

  5. Execute the following command to create an image file for your metadata:

    ALTER SYSTEM CREATE IMAGE;

  6. After the new image file is transmitted to the directory meta/image of all FE nodes, you can first downgrade a Follower FE node. If no error is returned, you can then downgrade other nodes in the cluster.