[
https://issues.apache.org/jira/browse/HBASE-30403?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wellington Chevreuil reassigned HBASE-30403:
--------------------------------------------
Assignee: Jeremy Schoemaker
> CustomDateTieredCompactionPolicy#getCompactBoundariesForMajor drops the
> cutOffTimestamp boundary when files lack CUSTOM_TIERING_TIME_RANGE
> ------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: HBASE-30403
> URL: https://issues.apache.org/jira/browse/HBASE-30403
> Project: HBase
> Issue Type: Bug
> Components: Compaction
> Reporter: Jeremy Schoemaker
> Assignee: Jeremy Schoemaker
> Priority: Major
> Labels: pull-request-available
>
> getCompactBoundariesForMajor derives the major-compaction boundaries from the
> CUSTOM_TIERING_TIME_RANGE file-info tag of every file in filesToCompact.
> Files written before custom tiering was enabled (or by any writer that does
> not set the tag) contribute nothing, so with a mixed or untagged set the
> min/max traversal produces a boundary list that never contains the
> cutOffTimestamp split. The major compaction then writes a single tier instead
> of separating hot and cold cells.
> Fix (PR https://github.com/apache/hbase/pull/8573): always return
> [Long.MIN_VALUE, cutOffTimestamp] without traversing filesToCompact. This is
> safe because CustomTieringMultiFileWriter#append already buckets each cell by
> these boundaries and only commits a file for a tier that actually received
> data, so an empty tier never produces an empty HFile. The five tests in
> TestCustomCellTieredCompactionPolicy and TestCustomCellTieredCompactor that
> asserted boundary counts derived from the old tag traversal are updated to
> the new intended behavior in the same PR.
> Review thread with wchevreuil on the PR agreed the simplified approach on
> 2026-08-26; this ticket is filed at his request so the change can be merged.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
