SQL Write

Syntax

  1. INSERT { INTO | OVERWRITE } table_identifier [ part_spec ] [ column_list ] { value_expr | query };

For more information, please check the syntax document:

Flink INSERT Statement

INSERT INTO

Use INSERT INTO to apply records and changes to tables.

  1. INSERT INTO my_table SELECT ...

INSERT INTO supports both batch and streaming mode. In Streaming mode, by default, it will also perform compaction, snapshot expiration, and even partition expiration in Flink Sink (if it is configured).

For multiple jobs to write the same table, you can refer to dedicated compaction job for more info.

Clustering

In Paimon, clustering is a feature that allows you to cluster data in your Append Table based on the values of certain columns during the write process. This organization of data can significantly enhance the efficiency of downstream tasks when reading the data, as it enables faster and more targeted data retrieval. This feature is only supported for Append Table(bucket = -1) and batch execution mode.

To utilize clustering, you can specify the columns you want to cluster when creating or writing to a table. Here’s a simple example of how to enable clustering:

  1. CREATE TABLE my_table (
  2.     a STRING,
  3.     b STRING,
  4.     c STRING,
  5. ) WITH (
  6.   'sink.clustering.by-columns' = 'a,b',
  7. );

You can also use SQL hints to dynamically set clustering options:

  1. INSERT INTO my_table /*+ OPTIONS('sink.clustering.by-columns' = 'a,b') */
  2. SELECT * FROM source;

The data is clustered using an automatically chosen strategy (such as ORDER, ZORDER, or HILBERT), but you can manually specify the clustering strategy by setting the sink.clustering.strategy. Clustering relies on sampling and sorting. If the clustering process takes too much time, you can decrease the total sample number by setting the sink.clustering.sample-factor or disable the sorting step by setting the sink.clustering.sort-in-cluster to false.

You can refer to FlinkConnectorOptions for more info about the configurations above.

Overwriting the Whole Table

For unpartitioned tables, Paimon supports overwriting the whole table. (or for partitioned table which disables dynamic-partition-overwrite option).

Use INSERT OVERWRITE to overwrite the whole unpartitioned table.

  1. INSERT OVERWRITE my_table SELECT ...

Overwriting a Partition

For partitioned tables, Paimon supports overwriting a partition.

Use INSERT OVERWRITE to overwrite a partition.

  1. INSERT OVERWRITE my_table PARTITION (key1 = value1, key2 = value2, ...) SELECT ...

Dynamic Overwrite

Flink’s default overwrite mode is dynamic partition overwrite (that means Paimon only deletes the partitions appear in the overwritten data). You can configure dynamic-partition-overwrite to change it to static overwritten.

  1. -- MyTable is a Partitioned Table
  3. -- Dynamic overwrite
  4. INSERT OVERWRITE my_table SELECT ...
  6. -- Static overwrite (Overwrite whole table)
  7. INSERT OVERWRITE my_table /*+ OPTIONS('dynamic-partition-overwrite' = 'false') */ SELECT ...

Truncate tables

Flink 1.17-

You can use INSERT OVERWRITE to purge tables by inserting empty value.

  1. INSERT OVERWRITE my_table /*+ OPTIONS('dynamic-partition-overwrite'='false') */ SELECT * FROM my_table WHERE false;

Flink 1.18+

  1. TRUNCATE TABLE my_table;

Purging Partitions

Currently, Paimon supports two ways to purge partitions.

  1. Like purging tables, you can use INSERT OVERWRITE to purge data of partitions by inserting empty value to them.

  2. Method #1 does not support to drop multiple partitions. In case that you need to drop multiple partitions, you can submit the drop_partition job through flink run.

  1. -- Syntax
  2. INSERT OVERWRITE my_table /*+ OPTIONS('dynamic-partition-overwrite'='false') */ 
  3. PARTITION (key1 = value1, key2 = value2, ...) SELECT selectSpec FROM my_table WHERE false;
  5. -- The following SQL is an example:
  6. -- table definition
  7. CREATE TABLE my_table (
  8.     k0 INT,
  9.     k1 INT,
  10.     v STRING
  11. ) PARTITIONED BY (k0, k1);
  13. -- you can use
  14. INSERT OVERWRITE my_table /*+ OPTIONS('dynamic-partition-overwrite'='false') */ 
  15. PARTITION (k0 = 0) SELECT k1, v FROM my_table WHERE false;
  17. -- or
  18. INSERT OVERWRITE my_table /*+ OPTIONS('dynamic-partition-overwrite'='false') */ 
  19. PARTITION (k0 = 0, k1 = 0) SELECT v FROM my_table WHERE false;

Updating tables

Important table properties setting:

  1. Only primary key table supports this feature.
  2. MergeEngine needs to be deduplicate or partial-update to support this feature.
  3. Do not support updating primary keys.

Currently, Paimon supports updating records by using UPDATE in Flink 1.17 and later versions. You can perform UPDATE in Flink’s batch mode.

  1. -- Syntax
  2. UPDATE table_identifier SET column1 = value1, column2 = value2, ... WHERE condition;
  4. -- The following SQL is an example:
  5. -- table definition
  6. CREATE TABLE my_table (
  7.     a STRING,
  8.     b INT,
  9.     c INT,
  10.     PRIMARY KEY (a) NOT ENFORCED
  11. ) WITH ( 
  12.     'merge-engine' = 'deduplicate' 
  13. );
  15. -- you can use
  16. UPDATE my_table SET b = 1, c = 2 WHERE a = 'myTable';

Deleting from table

Flink 1.17+

Important table properties setting:

  1. Only primary key tables support this feature.
  2. If the table has primary keys, MergeEngine needs to be deduplicate to support this feature.
  3. Do not support deleting from table in streaming mode.
  1. -- Syntax
  2. DELETE FROM table_identifier WHERE conditions;
  4. -- The following SQL is an example:
  5. -- table definition
  6. CREATE TABLE my_table (
  7.     id BIGINT NOT NULL,
  8.     currency STRING,
  9.     rate BIGINT,
  10.     dt String,
  11.     PRIMARY KEY (id, dt) NOT ENFORCED
  12. ) PARTITIONED BY (dt) WITH ( 
  13.     'merge-engine' = 'deduplicate' 
  14. );
  16. -- you can use
  17. DELETE FROM my_table WHERE currency = 'UNKNOWN';

Partition Mark Done

For partitioned tables, each partition may need to be scheduled to trigger downstream batch computation. Therefore, it is necessary to choose this timing to indicate that it is ready for scheduling and to minimize the amount of data drift during scheduling. We call this process: “Partition Mark Done”.

Example to mark done:

  1. CREATE TABLE my_partitioned_table (
  2.     f0 INT,
  3.     f1 INT,
  4.     f2 INT,
  5.     ...
  6.     dt STRING
  7. ) PARTITIONED BY (dt) WITH (
  8.     'partition.timestamp-formatter'='yyyyMMdd',
  9.     'partition.timestamp-pattern'='$dt',
  10.     'partition.time-interval'='1 d',
  11.     'partition.idle-time-to-done'='15 m'
  12. );
  1. Firstly, you need to define the time parser of the partition and the time interval between partitions in order to determine when the partition can be properly marked done.
  2. Secondly, you need to define idle-time, which determines how long it takes for the partition to have no new data, and then it will be marked as done.
  3. Thirdly, by default, partition mark done will create _SUCCESS file, the content of _SUCCESS file is a json, contains creationTime and modificationTime, they can help you understand if there is any delayed data. You can also configure other actions.