Gather statistics for CBO

This topic describes the basic concept of StarRocks CBO and how to collect statistics for the CBO. StarRocks 2.4 introduces histograms to gather accurate data distribution statistics.

What is CBO

Cost-based optimizer (CBO) is critical to query optimization. After an SQL query arrives at StarRocks, it is parsed into a logical execution plan. The CBO rewrites and transforms the logical plan into multiple physical execution plans. The CBO then estimates the execution cost of each operator in the plan (such as CPU, memory, network, and I/O) and chooses a query path with the lowest cost as the final physical plan.

StarRocks CBO was launched in StarRocks 1.16.0 and is enabled by default from 1.19 onwards. Developed upon the Cascades framework, StarRocks CBO estimates cost based on various statistical information. It is capable of choosing an execution plan with the lowest cost among tens of thousands of execution plans, significantly improving the efficiency and performance of complex queries.

Statistics are important for the CBO. They determine whether cost estimation is accurate and useful. The following sections detail the types of statistical information, the collection policy, and how to collect statistics and view statistical information.

Types of statistical information

StarRocks collects a variety of statistics as the input for cost estimation.

Basic statistics

By default, StarRocks periodically collects the following basic statistics of tables and columns:

• row_count: total number of rows in a table

• data_size: data size of a column

• ndv: column cardinality, which is the number of distinct values in a column

• null_count: the amount of data with NULL values in a column

• min: minimum value in a column

• max: maximum value in a column

Basic statistics are stored in the _statistics_.table_statistic_v1 table. You can view this table in the _statistics_ database of your StarRocks cluster.

Histogram

StarRocks 2.4 introduces histograms to complement basic statistics. Histogram is considered an effective way of data representation. For tables with skewed data, histograms can accurately reflect data distribution.

StarRocks uses equi-height histograms, which are constructed on several buckets. Each bucket contains an equal amount of data. For data values that are frequently queried and have a major impact on selectivity, StarRocks allocates separate buckets for them. More buckets mean more accurate estimation, but may also cause a slight increase in memory usage. You can adjust the number of buckets and most common values (MCVs) for a histogram collection task.

Histograms are applicable to columns with highly skewed data and frequent queries.If your table data is uniformly distributed, you do not need to create histograms. Histograms can be created only on columns of numeric, DATE, DATETIME, or string types.

Currently, StarRocks supports only manual collection of histograms. Histograms are stored in the histogram_statistics table of the _statistics_ database.

Collection types and methods

Data size and data distribution constantly change in a table. Statistics must be updated regularly to represent that data change. Before creating a statistics collection task, you must choose a collection type and method that best suit your business requirements.

StarRocks supports full and sampled collection, both can be performed automatically and manually. By default, StarRocks automatically collects full statistics of a table. It checks for any data updates every 5 minutes. If data change is detected, data collection will be automatically triggered. If you do not want to use automatic full collection, you can set the FE configuration item enable_collect_full_statistic to false and customize a collection task.

Collection type	Collection method	Description	Advantage and disadvantage
Full collection	Automatic/manual	Scans the full table to gather statistics.Statistics are collected partition. If a partition has no data change, data will not be collected from this partition, reducing resource consumption.Full statistics are stored in the _statistics_.column_statistics table.	Advantage: The statistics are accurate, which helps the CBO make accurate estimation.Disadvantage: It consumes system resources and is slow. From 2.5 onwards, StarRocks allows you to specify an automatic collection period, which reduces resource consumption.
Sampled collection	Automatic/manual	Evenly extracts N rows of data from each partition of a table.Statistics are collected by table. Basic statistics of each column are stored as one record. Cardinality information (ndv) of a column is estimated based on the sampled data, which is not accurate. Sampled statistics are stored in the _statistics_.table_statistic_v1 table.	Advantage: It consumes less system resources and is fast.Disadvantage: The statistics are not complete, which may affect the accuracy of cost estimation.

Collect statistics

StarRocks offers flexible statistics collection methods. You can choose automatic, manual, or custom collection, whichever suits your business scenarios.

Automatic collection

For basic statistics, StarRocks automatically collects full statistics of a table by default, without requiring manual operations. For tables on which no statistics have been collected, StarRocks automatically collects statistics within the scheduling period. For tables on which statistics have been collected, StarRocks updates the total number of rows and modified rows in the tables, and persists this information regularly for judging whether to trigger automatic collection.

From 2.5 onwards, StarRocks allows you to specify a collection period for automatic full collection, which prevents cluster performance jitter caused by automatic full collection. This period is specified by FE parameters statistic_auto_analyze_start_time and statistic_auto_analyze_end_time.

Conditions that will trigger automatic collection:

• Table data has changed since previous statistics collection.

• The collection time falls within the range of the configured collection period. (The default collection period is all day.)

• The update time of the previous collecting job is earlier than the latest update time of partitions.

• The health of table statistics is below the specified threshold (statistic_auto_collect_ratio).

Formula for calculating statistics health:

If the number of partitions with data updated is less than 10, the formula is 1 - (Number of updated rows since previous collection/Total number of rows). If the number of partitions with data updated is greater than or equal to 10, the formula is 1 - MIN(Number of updated rows since previous collection/Total number of rows, Number of updated partitions since previous collection/Total number of partitions).

In addition, StarRocks allows you to configure collection policies based on table size and table update frequency:

• For tables with small data volume, **statistics are collected in real time with no restrictions, even though the table data is frequently updated. The statistic_auto_collect_small_table_size parameter can be used to determine whether a table is a small table or a large table. You can also use statistic_auto_collect_small_table_interval to configure collection intervals for small tables.

• For tables with large data volume, the following restrictions apply:

  • The default collection interval is not less than 12 hours, which can be configured using statistic_auto_collect_large_table_interval.

  • When the collection interval is met and the statistics health is lower than the threshold for automatic sampled collection (statistic_auto_collect_sample_threshold), sampled collection is triggered.

  • When the collection interval is met and the statistics health is higher than the threshold for automatic sampled collection (statistic_auto_collect_sample_threshold) and lower than the automatic collection threshold (statistic_auto_collect_ratio), full collection is triggered.

  • When the size of the largest partition to collect data (statistic_max_full_collect_data_size) is greater than 100 GB, sampled collection is triggered.

  • Only statistics of partitions whose update time is later than the time of the previous collection task are collected. Statistics of partitions with no data change are not collected.

tip

After the data of a table is changed, manually triggering a sampled collection task for this table will make the update time of the sampled collection task later than the data update time, which will not trigger automatic full collection for this table in this scheduling period.

Automatic full collection is enabled by default and run by the system using the default settings.

The following table describes the default settings. If you need to modify them, run the ADMIN SET CONFIG command.

FE configuration item	Type	Default value	Description
enable_statistic_collect	BOOLEAN	TRUE	Whether to collect statistics. This switch is turned on by default.
enable_collect_full_statistic	BOOLEAN	TRUE	Whether to enable automatic full collection. This switch is turned on by default.
statistic_collect_interval_sec	LONG	300	The interval for checking data updates during automatic collection. Unit: seconds.
statistic_auto_collect_ratio	FLOAT	0.8	The threshold for determining whether the statistics for automatic collection are healthy. If statistics health is below this threshold, automatic collection is triggered.
statistic_auto_analyze_start_time	STRING	00:00:00	The start time of automatic collection. Value range: 00:00:00 - 23:59:59.
statistic_auto_analyze_end_time	STRING	23:59:59	The end time of automatic collection. Value range: 00:00:00 - 23:59:59.
statistic_auto_collect_small_table_size	LONG	5368709120	The threshold for determining whether a table is a small table for automatic full collection. A table whose size is greater than this value is considered a large table, whereas a table whose size is less than or equal to this value is considered a small table. Unit: Byte. Default value: 5368709120 (5 GB).
statistic_auto_collect_small_table_interval	LONG	0	The interval for automatically collecting full statistics of small tables. Unit: seconds.
statistic_auto_collect_large_table_interval	LONG	43200	The interval for automatically collecting full statistics of large tables. Unit: seconds. Default value: 43200 (12 hours).
statistic_auto_collect_sample_threshold	DOUBLE	0.3	The statistics health threshold for triggering automatic sampled collection. If the health value of statistics is lower than this threshold, automatic sampled collection is triggered.
statistic_max_full_collect_data_size	LONG	107374182400	The size of the largest partition for automatic collection to collect data. Unit: Byte. Default value: 107374182400 (100 GB). If a partition exceeds this value, full collection is discarded and sampled collection is performed instead.
statistic_full_collect_buffer	LONG	20971520	The maximum buffer size taken by automatic collection tasks. Unit: Byte. Default value: 20971520 (20 MB).
statistic_collect_max_row_count_per_query	INT	5000000000	The maximum number of rows to query for a single analyze task. An analyze task will be split into multiple queries if this value is exceeded.
statistic_collect_too_many_version_sleep	LONG	600000	The sleep time of automatic collection tasks if the table on which the collection task runs has too many data versions. Unit: ms. Default value: 600000 (10 minutes).

You can rely on automatic jobs for a majority of statistics collection, but if you have specific requirements, you can manually create a task by executing the ANALYZE TABLE statement or customize an automatic task by executing the CREATE ANALYZE statement.

Manual collection

You can use ANALYZE TABLE to create a manual collection task. By default, manual collection is a synchronous operation. You can also set it to an asynchronous operation. In asynchronous mode, after you run ANALYZE TABLE, the system immediately returns whether this statement is successful. However, the collection task will be running in the background and you do not have to wait for the result. You can check the status of the task by running SHOW ANALYZE STATUS. Asynchronous collection is suitable for tables with large data volume, whereas synchronous collection is suitable for tables with small data volume. Manual collection tasks are run only once after creation. You do not need to delete manual collection tasks.

Manually collect basic statistics

ANALYZE [FULL|SAMPLE] TABLE tbl_name (col_name [,col_name])
[WITH SYNC | ASYNC MODE]
PROPERTIES (property [,property]);

Parameter description:

• Collection type

  • FULL: indicates full collection.
  • SAMPLE: indicates sampled collection.
  • If no collection type is specified, full collection is used by default.

• col_name: columns from which to collect statistics. Separate multiple columns with commas (,). If this parameter is not specified, the entire table is collected.

• [WITH SYNC | ASYNC MODE]: whether to run the manual collection task in synchronous or asynchronous mode. Synchronous collection is used by default if you do not specify this parameter.

• PROPERTIES: custom parameters. If PROPERTIES is not specified, the default settings in the fe.conf file are used. The properties that are actually used can be viewed via the Properties column in the output of SHOW ANALYZE STATUS.

PROPERTIES	Type	Default value	Description
statistic_sample_collect_rows	INT	200000	The minimum number of rows to collect for sampled collection.If the parameter value exceeds the actual number of rows in your table, full collection is performed.

Examples

Manual full collection

-- Manually collect full stats of a table using default settings.
ANALYZE TABLE tbl_name;

-- Manually collect full stats of a table using default settings.
ANALYZE FULL TABLE tbl_name;

-- Manually collect stats of specified columns in a table using default settings.
ANALYZE TABLE tbl_name(c1, c2, c3);

Manual sampled collection

-- Manually collect partial stats of a table using default settings.
ANALYZE SAMPLE TABLE tbl_name;

-- Manually collect stats of specified columns in a table, with the number of rows to collect specified.
ANALYZE SAMPLE TABLE tbl_name (v1, v2, v3) PROPERTIES(
    "statistic_sample_collect_rows" = "1000000"
);

Manually collect histograms

ANALYZE TABLE tbl_name UPDATE HISTOGRAM ON col_name [, col_name]
[WITH SYNC | ASYNC MODE]
[WITH N BUCKETS]
PROPERTIES (property [,property]);

Parameter description:

• col_name: columns from which to collect statistics. Separate multiple columns with commas (,). If this parameter is not specified, the entire table is collected. This parameter is required for histograms.

• [WITH SYNC | ASYNC MODE]: whether to run the manual collection task in synchronous or asynchronous mode. Synchronous collection is used by default if you not specify this parameter.

• WITH N BUCKETS: N is the number of buckets for histogram collection. If not specified, the default value in fe.conf is used.

• PROPERTIES: custom parameters. If PROPERTIES is not specified, the default settings in fe.conf are used.

PROPERTIES	Type	Default value	Description
statistic_sample_collect_rows	INT	200000	The minimum number of rows to collect. If the parameter value exceeds the actual number of rows in your table, full collection is performed.
histogram_buckets_size	LONG	64	The default bucket number for a histogram.
histogram_mcv_size	INT	100	The number of most common values (MCV) for a histogram.
histogram_sample_ratio	FLOAT	0.1	The sampling ratio for a histogram.
histogram_max_sample_row_count	LONG	10000000	The maximum number of rows to collect for a histogram.

The number of rows to collect for a histogram is controlled by multiple parameters. It is the larger value between statistic_sample_collect_rows and table row count * histogram_sample_ratio. The number cannot exceed the value specified by histogram_max_sample_row_count. If the value is exceeded, histogram_max_sample_row_count takes precedence.

The properties that are actually used can be viewed via the Properties column in the output of SHOW ANALYZE STATUS.

Examples

-- Manually collect histograms on v1 using the default settings.
ANALYZE TABLE tbl_name UPDATE HISTOGRAM ON v1;

-- Manually collect histograms on v1 and v2, with 32 buckets, 32 MCVs, and 50% sampling ratio.
ANALYZE TABLE tbl_name UPDATE HISTOGRAM ON v1,v2 WITH 32 BUCKETS 
PROPERTIES(
   "histogram_mcv_size" = "32",
   "histogram_sample_ratio" = "0.5"
);

Custom collection

Customize an automatic collection task

You can use the CREATE ANALYZE statement to customize an automatic collection task.

Before creating a custom automatic collection task, you must disable automatic full collection (enable_collect_full_statistic = false). Otherwise, custom tasks cannot take effect.

-- Automatically collect stats of all databases.
CREATE ANALYZE [FULL|SAMPLE] ALL PROPERTIES (property [,property]);

-- Automatically collect stats of all tables in a database.
CREATE ANALYZE [FULL|SAMPLE] DATABASE db_name
PROPERTIES (property [,property]);

-- Automatically collect stats of specified columns in a table.
CREATE ANALYZE [FULL|SAMPLE] TABLE tbl_name (col_name [,col_name])
PROPERTIES (property [,property]);

Parameter description:

• Collection type

  • FULL: indicates full collection.
  • SAMPLE: indicates sampled collection.
  • If no collection type is specified, full collection is used by default.

• col_name: columns from which to collect statistics. Separate multiple columns with commas (,). If this parameter is not specified, the entire table is collected.

• PROPERTIES: custom parameters. If PROPERTIES is not specified, the default settings in fe.conf are used.

PROPERTIES	Type	Default value	Description
statistic_auto_collect_ratio	FLOAT	0.8	The threshold for determining whether the statistics for automatic collection are healthy. If the statistics health is below this threshold, automatic collection is triggered.
statistics_max_full_collect_data_size	INT	100	The size of the largest partition for automatic collection to collect data. Unit: GB.If a partition exceeds this value, full collection is discarded and sampled collection is performed instead.
statistic_sample_collect_rows	INT	200000	The minimum number of rows to collect.If the parameter value exceeds the actual number of rows in your table, full collection is performed.
statistic_exclude_pattern	String	null	The name of the database or table that needs to be excluded in the job. You can specify the database and table that do not collect statistics in the job. Note that this is a regular expression pattern, and the match content is database.table.

Examples

Automatic full collection

-- Automatically collect full stats of all databases.
CREATE ANALYZE ALL;

-- Automatically collect full stats of a database.
CREATE ANALYZE DATABASE db_name;

-- Automatically collect full stats of all tables in a database.
CREATE ANALYZE FULL DATABASE db_name;

-- Automatically collect full stats of specified columns in a table.
CREATE ANALYZE TABLE tbl_name(c1, c2, c3); 

-- Automatically collect stats of all databases, excluding specified database 'db_name'.
CREATE ANALYZE ALL PROPERTIES (
   "statistic_exclude_pattern" = "db_name\."
);

Automatic sampled collection

-- Automatically collect stats of all tables in a database with default settings.
CREATE ANALYZE SAMPLE DATABASE db_name;

-- Automatically collect stats of all tables in a database, excluding specified table 'db_name.tbl_name'.
CREATE ANALYZE SAMPLE DATABASE db_name PROPERTIES (
   "statistic_exclude_pattern" = "db_name.tbl_name"
);

-- Automatically collect stats of specified columns in a table, with statistics health and the number of rows to collect specified.
CREATE ANALYZE SAMPLE TABLE tbl_name(c1, c2, c3) PROPERTIES (
   "statistic_auto_collect_ratio" = "0.5",
   "statistic_sample_collect_rows" = "1000000"
);

View custom collection tasks

SHOW ANALYZE JOB [WHERE predicate]

You can filter results by using the WHERE clause. The statement returns the following columns.

Column	Description
Id	The ID of the collection task.
Database	The database name.
Table	The table name.
Columns	The column names.
Type	The type of statistics, including FULL and SAMPLE.
Schedule	The type of scheduling. The type is SCHEDULE for an automatic task.
Properties	Custom parameters.
Status	The task status, including PENDING, RUNNING, SUCCESS, and FAILED.
LastWorkTime	The time of the last collection.
Reason	The reason why the task failed. NULL is returned if task execution was successful.

Examples

-- View all the custom collection tasks.
SHOW ANALYZE JOB

-- View custom collection tasks of database `test`.
SHOW ANALYZE JOB where `database` = 'test';

Delete a custom collection task

DROP ANALYZE <ID>;

The task ID can be obtained by using the SHOW ANALYZE JOB statement.

Examples

DROP ANALYZE 266030;

View status of collection tasks

You can view the status of all the current tasks by running the SHOW ANALYZE STATUS statement. This statement cannot be used to view the status of custom collection tasks. To view the status of custom collection tasks, use SHOW ANALYZE JOB.

SHOW ANALYZE STATUS [WHERE predicate];

You can use LILE or WHERE to filter the information to return.

This statement returns the following columns.

List name	Description
Id	The ID of the collection task.
Database	The database name.
Table	The table name.
Columns	The column names.
Type	The type of statistics, including FULL, SAMPLE, and HISTOGRAM.
Schedule	The type of scheduling. ONCE means manual and SCHEDULE means automatic.
Status	The status of the task.
StartTime	The time when the task starts executing.
EndTime	The time when the task execution ends.
Properties	Custom parameters.
Reason	The reason why the task failed. NULL is returned if the execution was successful.

View statistics

View metadata of basic statistics

SHOW STATS META [WHERE];

This statement returns the following columns.

Column	Description
Database	The database name.
Table	The table name.
Columns	The column names.
Type	The type of statistics. FULL means full collection and SAMPLE means sampled collection.
UpdateTime	The latest statistics update time for the current table.
Properties	Custom parameters.
Healthy	The health of statistical information.

View metadata of histograms

SHOW HISTOGRAM META [WHERE];

This statement returns the following columns.

Column	Description
Database	The database name.
Table	The table name.
Column	The columns.
Type	Type of statistics. The value is HISTOGRAM for histograms.
UpdateTime	The latest statistics update time for the current table.
Properties	Custom parameters.

Delete statistics

You can delete statistical information you do not need. When you delete statistics, both the data and metadata of the statistics are deleted, as well as the statistics in expired cache. Note that if an automatic collection task is ongoing, previously deleted statistics may be collected again. You can use SHOW ANALYZE STATUS to view the history of collection tasks.

Delete basic statistics

DROP STATS tbl_name

Delete histograms

ANALYZE TABLE tbl_name DROP HISTOGRAM ON col_name [, col_name];

Cancel a collection task

You can use the KILL ANALYZE statement to cancel a running collection task, including manual and custom tasks.

KILL ANALYZE <ID>;

The task ID for a manual collection task can be obtained from SHOW ANALYZE STATUS. The task ID for a custom collection task can be obtained from SHOW ANALYZE SHOW ANALYZE JOB.

FE configuration items

FE configuration item	Type	Default value	Description
enable_statistic_collect	BOOLEAN	TRUE	Whether to collect statistics. This parameter is turned on by default.
enable_collect_full_statistic	BOOLEAN	TRUE	Whether to enable automatic full statistics collection. This parameter is turned on by default.
statistic_auto_collect_ratio	FLOAT	0.8	The threshold for determining whether the statistics for automatic collection are healthy. If statistics health is below this threshold, automatic collection is triggered.
statistic_max_full_collect_data_size	LONG	107374182400	The size of the largest partition for automatic collection to collect data. Unit: Byte. If a partition exceeds this value, full collection is discarded and sampled collection is performed instead.
statistic_collect_max_row_count_per_query	INT	5000000000	The maximum number of rows to query for a single analyze task. An analyze task will be split into multiple queries if this value is exceeded.
statistic_collect_interval_sec	LONG	300	The interval for checking data updates during automatic collection. Unit: seconds.
statistic_auto_analyze_start_time	STRING	00:00:00	The start time of automatic collection. Value range: 00:00:00 - 23:59:59.
statistic_auto_analyze_end_time	STRING	23:59:59	The end time of automatic collection. Value range: 00:00:00 - 23:59:59.
statistic_sample_collect_rows	LONG	200000	The minimum number of rows to collect for sampled collection. If the parameter value exceeds the actual number of rows in your table, full collection is performed.
statistic_collect_concurrency	INT	3	The maximum number of manual collection tasks that can run in parallel. The value defaults to 3, which means you can run a maximum of three manual collections tasks in parallel. If the value is exceeded, incoming tasks will be in the PENDING state, waiting to be scheduled.
histogram_buckets_size	LONG	64	The default bucket number for a histogram.
histogram_mcv_size	LONG	100	The number of most common values (MCV) for a histogram.
histogram_sample_ratio	FLOAT	0.1	The sampling ratio for a histogram.
histogram_max_sample_row_count	LONG	10000000	The maximum number of rows to collect for a histogram.
statistic_manager_sleep_time_sec	LONG	60	The interval at which metadata is scheduled. Unit: seconds. The system performs the following operations based on this interval:Create tables for storing statistics.Delete statistics that have been deleted.Delete expired statistics.
statistic_analyze_status_keep_second	LONG	259200	The duration to retain the history of collection tasks. The default value is 3 days. Unit: seconds.

References

• To query FE configuration items, run ADMIN SHOW CONFIG.

• To modify FE configuration items, run ADMIN SET CONFIG.