[
https://issues.apache.org/jira/browse/HDDS-17098?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Siyao Meng updated HDDS-17098:
------------------------------
Status: Patch Available (was: Open)
> Write a previousSnapshotVersions summary in snapshot local data YAML so OM
> startup can stop reading before the SST file lists
> -----------------------------------------------------------------------------------------------------------------------------
>
> Key: HDDS-17098
> URL: https://issues.apache.org/jira/browse/HDDS-17098
> Project: Apache Ozone
> Issue Type: Sub-task
> Components: Ozone Manager, Snapshot
> Reporter: Siyao Meng
> Assignee: Siyao Meng
> Priority: Major
> Labels: pull-request-available
>
> To build the snapshot version graph at startup,
> {{OmSnapshotLocalDataManager#init}} needs only {{snapshotId}},
> {{previousSnapshotId}}, {{version}} and the {{previousSnapshotVersion}} of
> each version. With HDDS-17079 these are read from parser events and SST file
> entries are skipped, but the whole file is still read, because each
> {{previousSnapshotVersion}} sits inside {{versionSstFileInfos}} next to its
> SST file list, and there is no way to know that the last version has been
> seen until the end of the mapping.
> On a test cluster with about 19,350 snapshots the YAMLs total 126 to 260 GB
> per OM (median about 9 MB per file, almost all SST file entries). With
> HDDS-17078 and HDDS-17079 the YAML phase is disk bound: about 26 minutes to
> read 175.6 GB on HDD.
> Proposal: also write a derived top level map {{previousSnapshotVersions}}
> (version to previousSnapshotVersion). YAML keys are emitted in sorted order,
> so it and the snapshot ids and version all precede {{versionSstFileInfos}}.
> When the summary is present, {{loadWithoutSstFiles}} stops at the
> {{versionSstFileInfos}} key and reads only the first few KB of the file.
> Files without it are still streamed in full.
> Compatibility:
> * Released readers (2.1.x, 2.2.x) construct the document by key and ignore
> unknown keys, so they load new files. Verified by loading a new file with the
> 2.2.1 classes.
> * The summary is left out of the checksum computation, so the checksum is the
> same with or without it, and both old and new code verify old and new files.
> * An older OM rewriting a file drops the summary; the file then falls back to
> the full stream. A stale summary cannot be produced.
> * No layout feature is needed, since the change is additive and every
> released reader tolerates it.
> Tradeoff: for files with the summary, startup no longer parses the SST file
> lists at all, so malformed content there is not detected at startup and
> surfaces on the next full load of that snapshot (invalid SST values already
> behave this way after HDDS-17079). YAMLs are written atomically, so truncated
> files are not expected.
> Existing files keep the old layout until rewritten (snapshot create, purge,
> defrag, orphan version cleanup, previous snapshot change). With defrag off by
> default, most existing files on an upgraded cluster are not rewritten, so the
> gain applies to snapshots created after the upgrade until older ones age out.
> A one time rewrite of existing files is left for a follow-up.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
