feat(commit): support TruncateTable/Abort/RollbackToAsLatest and add ut

[lucasfang]

Purpose

Linked issue: close #xxx

This change aligns the C++ FileStoreCommit with Java Paimon by adding three missing commit operations, plus unit-test coverage for the commit path.

• TruncateTable(commit_identifier): truncates the whole table by overwriting all partitions with no new files. It reuses the existing TryOverwrite path with an empty partition list (which matches all partitions), producing an OVERWRITE snapshot that deletes every existing file. Mirrors Java FileStoreCommit.truncateTable.
• Abort(commit_messages): cleans up after an unsuccessful commit by best-effort deleting the data and index files referenced by the given commit messages (new files, changelog files, compact-after files, and new index files from both the data and compact increments). Delete failures are ignored, matching Java deleteQuietly. A non-CommitMessageImpl message is rejected with an error. Mirrors Java FileStoreCommit.abort.
• RollbackToAsLatest(target_snapshot_id): rolls the table back to a target snapshot and materializes it as a new latest snapshot. It reads the surviving ADD data files of both the current latest snapshot and the target snapshot, computes their difference (files only in latest become DELETE, files only in target become ADD), and writes the base and delta manifests. It then commits a single OVERWRITE snapshot whose visible state equals the target: the new snapshot inherits the target's schema id, index manifest, total record count, watermark, statistics, and properties, and its nextRowId is the larger of the latest and target nextRowId to keep row-id allocation monotonic. Mirrors Java FileStoreCommit.rollbackToAsLatest.

Unlike Java, the C++ commit path has no commit-callback mechanism, so for RollbackToAsLatest the Java pre/post callbacks and rollbackIndexChanges are intentionally omitted; the new snapshot's index manifest points directly to the target's, which is equivalent for the on-disk result. Two private helpers are added to FileStoreCommitImpl: ReadAddManifestEntries (mirrors FileEntry.mergeEntries followed by keeping only ADD entries) and MaxNextRowId (mirrors Java maxNextRowId).

The change also adds unit tests for FileStoreCommitImpl and introduces a dedicated sequence_snapshot_properties_test.cpp suite.

Tests

Added GoogleTest cases (run via paimon-core-test):

• FileStoreCommitImplTest.TestTruncateTable — truncates a committed table and asserts a new OVERWRITE snapshot whose delta manifest deletes all files (0 added, 3 deleted for the append_09 fixture).
• FileStoreCommitImplTest.TestAbortDeletesDataAndIndexFiles — materializes the data/index files referenced by a commit message and verifies Abort removes them.
• FileStoreCommitImplTest.AbortIgnoresMissingFilesAndFailsForNonImplMessage — aborting missing files is a no-op; a non-impl message returns an error.
• FileStoreCommitImplTest.TestRollbackToAsLatest — appends three snapshots, then rolls back to snapshot 1 (exercising the DELETE-only delta branch) and rolls forward to snapshot 3 (exercising the ADD-only delta branch). Asserts a new OVERWRITE snapshot with the target's schema id, total record count, index manifest, nextRowId == max(latest, target), and a surviving file set equal to the target.
• FileStoreCommitImplTest.TestRollbackToAsLatestDeletionVectorOnlyChange — when the target and latest share identical data files and differ only in the deletion vector (index) manifest, the rollback produces an empty data delta (DeltaRecordCount == 0) and the new snapshot inherits the target's index manifest, dropping the deletion vector.
• FileStoreCommitImplTest.TestRollbackToAsLatestNoLatestSnapshotReturnsError — rolling back an empty table returns an error.
• FileStoreCommitImplTest.TestRollbackToAsLatestTargetNotExistReturnsError — rolling back to a non-existent target snapshot returns an error.
• Additional FileStoreCommitImplTest coverage (drop-partition validation, filter-and-commit, overwrite ownership/upgrade, conflict checks, etc.).
• New SequenceSnapshotPropertiesTest suite covering max-sequence-number parsing, extraction from files, and merge behavior.

API and Format

Yes — public API additions. Three new pure-virtual methods on paimon::FileStoreCommit (include/paimon/file_store_commit.h):

• virtual Status TruncateTable(int64_t commit_identifier) = 0;
• virtual Status Abort(const std::vector<std::shared_ptr<CommitMessage>>& commit_messages) = 0;
• virtual Result<bool> RollbackToAsLatest(int64_t target_snapshot_id) = 0;

No storage format or protocol change. TruncateTable and RollbackToAsLatest produce standard OVERWRITE snapshots; there is no on-disk layout change.

Documentation

New public API methods, documented via doxygen comments in the header. No new user-guide page is required.

Generative AI tooling

Generated-by: Qoder

[Copilot]

Pull request overview

This PR extends the C++ FileStoreCommit commit path with table-wide truncation and best-effort abort cleanup, and adds unit tests covering these new behaviors along with additional commit-path correctness checks.

Changes:

• Add new public FileStoreCommit APIs: TruncateTable() and Abort(...), with corresponding FileStoreCommitImpl implementations.
• Implement truncate as an overwrite of all partitions with empty changes; implement abort as best-effort deletion of newly produced data/index files from commit messages.
• Add/extend unit tests (including a new sequence_snapshot_properties_test.cpp) and wire the new test into the CMake test sources.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
include/paimon/file_store_commit.h	Adds new public commit APIs and documents their semantics.
src/paimon/core/operation/file_store_commit_impl.h	Declares the new override methods in FileStoreCommitImpl.
src/paimon/core/operation/file_store_commit_impl.cpp	Implements TruncateTable and Abort behaviors in the commit implementation.
src/paimon/core/operation/file_store_commit_impl_test.cpp	Adds unit tests for truncation, abort cleanup, and additional commit-path validation behaviors.
src/paimon/core/operation/commit/sequence_snapshot_properties_test.cpp	Introduces UT coverage for sequence snapshot properties parsing/merging.
src/paimon/CMakeLists.txt	Registers the new test source in the test build.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[zjw1111]

Reviewed the three new commit operations against the Java FileStoreCommitImpl. No blocking correctness issues; the implementations track Java line-by-line and the tests are behavior-level (write files → operate → assert file/snapshot state).

Verified as correct:

• TruncateTable: empty partitions={} maps to "all partitions" in both the data scan (FileStoreScan::CreatePartitionPredicate) and the index filter (CommitScanner::ReadAllIndexEntriesFromPartitions), equivalent to Java tryOverwritePartition(null, ...).
• Abort: delete set = newFiles + changelogFiles + compactAfter + compact.changelogFiles (data) and newIndexFiles from both increments; correctly excludes compactBefore. Matches Java exactly.
• RollbackToAsLatest: Snapshot ctor argument order matches the regular commit path; total_record_count/index_manifest/watermark/statistics/properties from target, delta_record_count from the diff, commit_identifier=int64_max, next_row_id=max(latest,target) — all consistent with Java. Skipping pre/post-commit callbacks is consistent with the C++ path not implementing callbacks yet.

One inline comment on scope. Test blind spots (non-blocking):

• RollbackToAsLatest: no coverage for the success=false (concurrent-conflict) branch and the temp-manifest-list leak on that path (leak matches Java).
• Abort: no coverage for a delete that fails on IO/permission and is silently ignored (only the missing-file no-op is covered).
• TruncateTable: only single-truncate covered; no empty-table case.

[zjw1111]

Two follow-ups after the latest fixes.

[zjw1111]

Minor: the PR description has a lot of hard line breaks in the middle of sentences. Please drop them and let the text wrap naturally — the forced wrapping reads awkwardly in the rendered view.

[zjw1111]

+1