[
https://issues.apache.org/jira/browse/HBASE-1738?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
stack updated HBASE-1738:
-------------------------
Attachment: 1738-v3.patch
Here is v3:
{code}
M src/test/org/apache/hadoop/hbase/regionserver/TestMemStore.java
Reset memstore scanners at end of each full scan.
Reenable part of test that was broken because we weren't getting
all values.
M src/java/org/apache/hadoop/hbase/regionserver/MemStore.java
Have MemStoreScanners become Observers on MemStore. If MemStore
snapshots or clears snapshots, have Scanners update themselves.
Changes MemStore to have a single scanner for both kvset and for
snapshot rather than one for each. Now we meld the results from
both in the one scanner.
New inner FirstOnRow class that is updated atomically.
{code}
> Scanner doesnt reset when a snapshot is created, could miss new updates into
> the 'kvset' (active part)
> ------------------------------------------------------------------------------------------------------
>
> Key: HBASE-1738
> URL: https://issues.apache.org/jira/browse/HBASE-1738
> Project: Hadoop HBase
> Issue Type: Bug
> Affects Versions: 0.20.0
> Reporter: ryan rawson
> Assignee: stack
> Fix For: 0.20.0
>
> Attachments: 1738-v2.patch, 1738-v3.patch, 1738.patch
>
>
> when a Scanner is created, it creates 2 MemStoreScanners on the kvset and the
> snapshot (internal names of Memstore)... if the snapshot is originally empty,
> it only creates the 1, for kvset. When the snapshot is created, the
> outstanding Scanners now have a pointer to the tree that is now the snapshot,
> but no pointer to the kvset.
> When the flush completes, the scanner will reset the memstore scanners and
> 'see' the new values again.
> If there is a large delay between snapshot and finalization of the flush,
> there can be a large period of time a scanner doesnt see 'new' values that
> are being inserted. the canonical 'bad' case where this can do things is the
> META scanner, and we end up with double assignment.
> The snapshot is really lightweight, it only takes out a small lock in
> memstore, so im not sure there is an easy mechanism to hook to without
> building out a bit more code or restructuring the memstore scanner.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
