Skip to content

[Feature] Bypass manifest lists on exact snapshot live-entry cache hits #375

Description

@wangyong9999

Search before asking

  • I searched existing issues and found no report of this exact-hit planning overhead.

Motivation

Repeated scans of the same snapshot and bucket still read and decode the manifest lists before checking the snapshot live-manifest-entry cache. The cached live entries already contain the state needed for that bucket, so this work adds avoidable CPU and metadata access to repeated point-read planning.

To reproduce, use the public orc/append_09.db/append_09 test fixture with snapshot 5 and bucket 0, enable the snapshot live-manifest-entry cache, and reuse one cache across two scans. Instrument the MANIFEST cache accesses, as in AppendOnlyFileStoreScanTest.TestSnapshotLiveManifestCachePath. Before the proposed change, the second scan revisits the manifest-list path even though its exact snapshot/bucket live-entry lookup hits.

Expected: an eligible exact hit bypasses manifest-list reads and decoding, while each request still applies its own filters. A miss must rebuild the complete bucket. Snapshot selection and the existing snapshot-identity validation must remain intact.

Solution

Resolve the snapshot first and consult the existing snapshot live-entry cache before reading manifest lists. Retain the table file count in the transient cached payload so skipped-file metrics remain correct; report zero scanned manifests for an exact hit. Version the transient payload so incompatible entries fall back to rebuilding.

Draft implementation: #374. Regression tests check the manifest-cache access count, changed predicates, eviction/rebuild, and retained file counts. This change does not introduce a filesystem block cache or alter the persisted table format.

Anything else?

The identical patch passed 88 focused tests in a downstream C++17 Release build. Upstream CI for the draft still requires maintainer approval; no standalone end-to-end performance uplift is claimed for this individual optimization.

Are you willing to submit a PR?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions