corgy-w opened a new pull request, #11962:
URL: https://github.com/apache/seatunnel/pull/11962
Closes #11961
### Purpose of this pull request
Separate savepoint storage from checkpoint storage and introduce a versioned
savepoint bundle format (`engine-wire-v1`) with manifest checksums and
versioned readers, so that savepoints are never lost by checkpoint
retention/terminal cleanup and the wire format has an explicit compatibility
contract.
See the linked issue for the full problem statement and design.
### Does this PR introduce _any_ user-facing change?
Yes, compared to `dev` (unreleased branch):
- **Previous behavior**: savepoints are written into the checkpoint
directory (`<namespace>/<job-id>/`) with the runtime serialization format,
participate in `max-retained` rotation, and are deleted together with
checkpoints when the job reaches `FINISHED`/`CANCELED` after a
savepoint-restore cycle.
- **New behavior**: savepoints are written as isolated bundles under
`<namespace>/savepoint/<job-id>/<savepoint-id>/` (payloads staged first,
`_metadata.ser` as the commit marker with `formatVersion=1` and per-file length
+ SHA-256 checksums). They are excluded from checkpoint rotation and terminal
cleanup, kept until explicitly deleted. Restore selects the newest completed
bundle; savepoints written by earlier engines (still inside the checkpoint
directory, no version marker) remain readable on a best-effort basis. Bundles
written by a newer engine are rejected with an explicit error.
`restore.mode`/`savepoint.path` env options remain **planned-not-implemented**
(documented).
The `CheckpointStorage` SPI, CLI/REST parameters and checkpoint storage
layout are unchanged.
### How was this patch tested?
Unit/integration tests (all green locally, JDK 11):
- `CheckpointWireCompatibilityTest` (8): legacy-v0 fixture replay, byte
stability, `isRestored` excluded from v1, enum name encoding, unknown-enum
explicit error
- `SavepointReaderRegistryTest` (3): v1 readable, below-window migration
error, above-window rejection
- `LocalSavepointStorageTest` (5) / `HdfsSavepointStorageTest` (2):
commit/list/read/delete, staging invisibility before commit, abort, corruption
detection, duplicate id rejection, checkpoint/savepoint isolation
- `CheckpointStorageTest` (5), `SavePointTest` (6, 1 skipped on Windows),
`CheckpointManagerTest` (6), `CheckpointCoordinatorTest` (14):
savepoint-and-restore end-to-end with the new bundle path; updated
`CheckpointStorageTest` asserts isolation
- `SavepointRestoreSelectionTest` (1): two savepoints coexist as isolated
bundles, restore picks the newest after deleting the older one
Build commands used (local notes: JDK 21 breaks the bundled
google-java-format, use JDK 11; build with `-pl` so shaded modules resolve from
the local repo):
```bash
JAVA_HOME=<jdk11> ./mvnw -pl seatunnel-engine/seatunnel-engine-server
spotless:apply
JAVA_HOME=<jdk11> ./mvnw -pl seatunnel-engine/seatunnel-engine-server
-Dskip.spotless=true \
-Dtest='SavePointTest,CheckpointStorageTest,CheckpointCoordinatorTest,CheckpointManagerTest,CheckpointWireCompatibilityTest,SavepointReaderRegistryTest,SavepointRestoreSelectionTest'
test
JAVA_HOME=<jdk11> ./mvnw -pl
seatunnel-engine/seatunnel-engine-storage/checkpoint-storage-plugins/checkpoint-storage-local-file,seatunnel-engine/seatunnel-engine-storage/checkpoint-storage-plugins/checkpoint-storage-hdfs
-Dskip.spotless=true \
-Dtest='LocalFileStorageTest,LocalSavepointStorageTest,AbstractFileCheckPointTest,HdfsSavepointStorageTest'
test
```
### Check list
* [x] If necessary, please update the documentation to describe the new
feature. https://github.com/apache/seatunnel/tree/dev/docs
(`state-storage-and-recovery.md` en/zh, `rest-api-job-lifecycle.md` en/zh)
* [x] If necessary, please update `incompatible-changes.md` to describe the
incompatibility caused by this PR.
* [x] No new Jar binary packages added (no License Notice needed).
* [x] No connector code changes (plugin-mapping.properties / seatunnel-dist
pom / label-scope-conf / e2e / plugin_config not applicable).
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
