...
Discussion thread | https://lists.apache.org/thread/d8toll8042k3m4z3pvn1pwg0n8twm0wt |
---|---|
Vote thread | TBD |
ISSUE | TBD |
Release | Paimon-0.5 |
Motivation
In data warehouse, it's a common requirement to query historical data. So we need to periodically merge incremental data with historical data to generate complete data of current time. For example, Hive 2.2 supports using MERGE INTO statement to merge source table's records into target table by performing update, insert and delete according to different filter, so a good way is taking historical table as target and incremental table as source, then using MERGE INTO to perform update (for existing records) and insert (for non-existing records). But when there exists a large amount of data, the merging may cost much time.
...
To solve this problem, we propose to introduce a new mechanism `Tag`. A tag is created from snapshot and can keep longer. The purposes of tag are:
Fault recovery (or we can say disaster recovery). Users can ROLL BACK to a tag if needed. If user rollbacks to a tag, the table will hold the data in the tag and the data committed after the tag will be deleted. In this scenario the tag is like backup, but more lightweight because the tag only record the metadata of data files instead of copying them.
Record versions of data at a longer interval (typically daily or weekly). With tag, user can query the old data in batch mode. In this scenario the tag is like long-lived snapshot.
Proposed Changes
Storage
Like snapshot, a new directory `/tag` will be created under table directory for storing tags. The qualified path for a tag file is `/path/to/table/tag/tag-<id>`, and the id is increased from 1.
New Classes
We need an entrance class `Tag` to store the information of a tag:
...
The schema of TagsTable is in section `Public Interfaces`.
Time Travel to Tags Support
Time Travel to Tag (Batch Reading)
Currently, we support time traveling to a snapshot. Now we can support to travel to a savepoint.
For spark, we can use SQL `SELECT * FROM t VERSION AS OF <version-string>`. Currently, the version string only supports a long value which represents snapshot. Now we suppose to use `snapshot.<id>` to specify traveling to snapshot or use `savepoint.<id>` to specify traveling to savepoint. The old style `<id>` will be considered as snapshot id.
For Flink, we can use dynamic option `scan.snapshot-id=<snapshot-id>`. We suppose to introduce a new core option `scan.savepoint-id` to support traveling to a savepoint.
Time Travel to Timestamp with Savepoint
Currently, we support time traveling to a time point. Paimon Scanner will find the snapshot of which commit time is closest to the specified time. The problem is all of the snapshots before the time point may have expired. So we suppose to try to find the closest savepoint in this scenario.
Usage
Periodically savepoints
Public Interfaces
Table Options
key | Requires | Type | note |
---|---|---|---|
savepoint.create-time | false | String | The time point to create savepoint periodically. Must be the format 'hh:mm'. |
savepoint.create-interval | false | Duration | Interval between two savepoints. At least 1d and must be intergers of days. |
Flink Actions
Now,
Call Procedures (Future work)
- Spark supports extension to extend SQL syntax. We can provide savepoint procedures after we support CALL statement for Spark.
- Supporting CALL is in the road map of Flink. We can provide savepoint procedures after then.
Proposed Changes
Storage
Like snapshot, a new directory `/savepoint` will be created under table directory for storing savepoints. The qualified path for a savepoint file is `/path/to/table/savepoint/savepoint-<id>`.
Creation & Deletion
Currently, we can provide two ways for user to control the creation and deletion of savepoints.
...
In the future, after Flink support stored procedure, we can provide procedures to managing savepoints.
Interaction with Snapshot
- When creating savepoint, we will do a full-compact and pick the compacted snapshot to save.
- When expiring snapshot, Paimon will check whether the data files are used by living snapshots. If not, they will be deleted. After we introduce savepoint, we should also check if the data files are used by savepoints.
- When deleting savepoint, we will check and delete unused datafiles (like we expire snapshot).
System Table
We suppose to introduce a system table `SavepointsTable`. The schema is:
Code Block | ||||
---|---|---|---|---|
| ||||
savepoint_id BIGINT schema_id BIGINT save_time BIGINT record_count BIGINT |
Usage
Time Travel to Savepoint (Batch Reading)
Currently, we support time traveling to a snapshot. Now we can support to travel to a savepoint.
For spark, we can use SQL `SELECT * FROM t VERSION AS OF <version-string>`. Currently, the version string only supports a long value which represents snapshot. Now we suppose to use `snapshot.<id>` to specify traveling to snapshot or use `savepoint.<id>` to specify traveling to savepoint. The old style `<id>` will be considered as snapshot id.
For Flink, we can use dynamic option `scan.snapshot-id=<snapshot-id>`. We suppose to introduce a new core option `scan.savepoint-id` to support traveling to a savepoint.
Time Travel to Timestamp with Savepoint
...
Compatibility, Deprecation, and Migration Plan
None.
Test Plan
UT tests: verify creating and reading savepoint
...
- time travel to savepoint
- expiration of snapshots won't delete data files pointed by savepoints
Rejected Alternatives
Support starting from a savepoint in streaming reading
Current design of savepoint just store full data manifests, so it's not able to support streaming reading now.