Manage Snapshots

This section will describe the management and behavior related to snapshots.

Expire Snapshots

Paimon writers generate one or two snapshot per commit. Each snapshot may add some new data files or mark some old data files as deleted. However, the marked data files are not truly deleted because Paimon also supports time traveling to an earlier snapshot. They are only deleted when the snapshot expires.

Currently, expiration is automatically performed by Paimon writers when committing new changes. By expiring old snapshots, old data files and metadata files that are no longer used can be deleted to release disk space.

Snapshot expiration is controlled by the following table properties.

When the number of snapshots is less than snapshot.num-retained.min, no snapshots will be expired(even the condition snapshot.time-retained meet), after which snapshot.num-retained.max and snapshot.time-retained will be used to control the snapshot expiration until the remaining snapshot meets the condition.

The following example show more details(snapshot.num-retained.min is 2, snapshot.time-retained is 1h, snapshot.num-retained.max is 5):

snapshot item is described using tuple (snapshotId, corresponding time)

Please note that too short retain time or too small retain number may result in:

• Batch queries cannot find the file. For example, the table is relatively large and the batch query takes 10 minutes to read, but the snapshot from 10 minutes ago expires, at which point the batch query will read a deleted snapshot.
• Streaming reading jobs on table files (without the external log system) fail to restart. When the job restarts, the snapshot it recorded may have expired. (You can use Consumer Id to protect streaming reading in a small retain time of snapshot expiration).

By default, paimon will delete expired snapshots synchronously. When there are too many files that need to be deleted, they may not be deleted quickly and back-pressured to the upstream operator. To avoid this situation, users can use asynchronous expiration mode by setting snapshot.expire.execution-mode to async.

Rollback to Snapshot

Rollback a table to a specific snapshot ID.

Flink

Run the following command:

<FLINK_HOME>/bin/flink run \
    /path/to/paimon-flink-action-0.8.2.jar \
    rollback_to \
    --warehouse <warehouse-path> \
    --database <database-name> \ 
    --table <table-name> \
    --version <snapshot-id> \
    [--catalog_conf <paimon-catalog-conf> [--catalog_conf <paimon-catalog-conf> ...]]

Java API

import org.apache.paimon.table.Table;
public class RollbackTo {
    public static void main(String[] args) {
        // before rollback:
        // snapshot-3
        // snapshot-4
        // snapshot-5
        // snapshot-6
        // snapshot-7
        table.rollbackTo(5);
        // after rollback:
        // snapshot-3
        // snapshot-4
        // snapshot-5
    }
}

Spark

Run the following sql:

CALL rollback(table => 'test.T', version => '2');

Remove Orphan Files

Paimon files are deleted physically only when expiring snapshots. However, it is possible that some unexpected errors occurred when deleting files, so that there may exist files that are not used by Paimon snapshots (so-called “orphan files”). You can submit a remove_orphan_files job to clean them:

Flink

<FLINK_HOME>/bin/flink run \
    /path/to/paimon-flink-action-0.8.2.jar \
    remove_orphan_files \
    --warehouse <warehouse-path> \
    --database <database-name> \ 
    --table <table-name> \
    [--older_than <timestamp>]

To avoid deleting files that are newly added by other writing jobs, this action only deletes orphan files older than 1 day by default. The interval can be modified by --older_than. For example:

<FLINK_HOME>/bin/flink run \
    /path/to/paimon-flink-action-0.8.2.jar \
    remove_orphan_files \
    --warehouse <warehouse-path> \
    --database <database-name> \ 
    --table T \
    --older_than '2023-10-31 12:00:00'

Spark

CALL sys.remove_orphan_files(table => "tableId", [older_than => "2023-10-31 12:00:00"])