[
https://issues.apache.org/jira/browse/HBASE-30381?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Xiao Liu updated HBASE-30381:
-----------------------------
Description:
h2. Problem
MOB snapshot scanning fails on branch-2.5 through both:
* \{{TableSnapshotInputFormat}}, which generates an input split for the MOB
pseudo-region.
* \{{TableSnapshotScanner}} with \{{snapshotAlreadyRestored=true}}, which
includes the MOB pseudo-region in its region list.
Opening the pseudo-region as a regular region fails with:
{noformat}
java.io.IOException: Region directory does not exist:
.../data/<namespace>/<table>/<mob-region-encoded-name>
{noformat}
HBASE-30365 and HBASE-30368 fixed these paths on newer branches, but the fixes
are missing from branch-2.5.
h2. Root cause and affected versions
The missing MOB pseudo-region filtering has existed since at least 2.5.0.
Previously, region initialization could automatically create the missing
directory, masking the problem by scanning an empty region.
HBASE-29662 was backported to branch-2.5 as commit 7b58903e6ea and first
included in 2.5.13. It prevents region initialization from silently creating
missing region directories, exposing this existing defect as a scan failure.
This is not caused by HBASE-30336 on branch-2.5. The branch already skips MOB
regions when creating regular region directories in
\{{RestoreSnapshotHelper.cloneRegion()}}. MOB files are restored under
\{{mobdir}}, so the scanners must exclude the pseudo-region.
h2. Reproduction
# Create a MOB-enabled table, write MOB data, and take a snapshot.
# Scan the snapshot using \{{TableSnapshotInputFormat}}; alternatively, restore
it with \{{RestoreSnapshotHelper.copySnapshotForScanner()}} and scan it using
\{{TableSnapshotScanner}} with \{{snapshotAlreadyRestored=true}}.
# The scan fails when it attempts to open the MOB pseudo-region.
A minimal local reproduction confirms both failures. Adding the MOB filters
makes both paths pass. A comparison using \{{HRegionFileSystem}} from
immediately before and after HBASE-29662 also confirms the transition from
automatic directory creation to the exception.
h2. Proposed fix
Backport the MOB pseudo-region filters from:
* HBASE-30365: \{{TableSnapshotInputFormatImpl.getRegionInfosFromManifest()}}.
* HBASE-30368: \{{TableSnapshotScanner.isValidRegion()}}.
> MOB snapshot scanning fails on branch-2.5 after HBASE-29662
> -----------------------------------------------------------
>
> Key: HBASE-30381
> URL: https://issues.apache.org/jira/browse/HBASE-30381
> Project: HBase
> Issue Type: Bug
> Components: mapreduce, mob, snapshots
> Affects Versions: 2.5.13, 2.5.14, 2.5.15, 2.5.16
> Reporter: Xiao Liu
> Priority: Major
> Fix For: 2.5.17
>
>
> h2. Problem
> MOB snapshot scanning fails on branch-2.5 through both:
> * \{{TableSnapshotInputFormat}}, which generates an input split for the MOB
> pseudo-region.
> * \{{TableSnapshotScanner}} with \{{snapshotAlreadyRestored=true}}, which
> includes the MOB pseudo-region in its region list.
> Opening the pseudo-region as a regular region fails with:
> {noformat}
> java.io.IOException: Region directory does not exist:
> .../data/<namespace>/<table>/<mob-region-encoded-name>
> {noformat}
> HBASE-30365 and HBASE-30368 fixed these paths on newer branches, but the
> fixes are missing from branch-2.5.
> h2. Root cause and affected versions
> The missing MOB pseudo-region filtering has existed since at least 2.5.0.
> Previously, region initialization could automatically create the missing
> directory, masking the problem by scanning an empty region.
> HBASE-29662 was backported to branch-2.5 as commit 7b58903e6ea and first
> included in 2.5.13. It prevents region initialization from silently creating
> missing region directories, exposing this existing defect as a scan failure.
> This is not caused by HBASE-30336 on branch-2.5. The branch already skips MOB
> regions when creating regular region directories in
> \{{RestoreSnapshotHelper.cloneRegion()}}. MOB files are restored under
> \{{mobdir}}, so the scanners must exclude the pseudo-region.
> h2. Reproduction
> # Create a MOB-enabled table, write MOB data, and take a snapshot.
> # Scan the snapshot using \{{TableSnapshotInputFormat}}; alternatively,
> restore it with \{{RestoreSnapshotHelper.copySnapshotForScanner()}} and scan
> it using \{{TableSnapshotScanner}} with \{{snapshotAlreadyRestored=true}}.
> # The scan fails when it attempts to open the MOB pseudo-region.
> A minimal local reproduction confirms both failures. Adding the MOB filters
> makes both paths pass. A comparison using \{{HRegionFileSystem}} from
> immediately before and after HBASE-29662 also confirms the transition from
> automatic directory creation to the exception.
> h2. Proposed fix
> Backport the MOB pseudo-region filters from:
> * HBASE-30365: \{{TableSnapshotInputFormatImpl.getRegionInfosFromManifest()}}.
> * HBASE-30368: \{{TableSnapshotScanner.isValidRegion()}}.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
