Still a 'hole' in scanners, even after HBASE-532
------------------------------------------------
Key: HBASE-588
URL: https://issues.apache.org/jira/browse/HBASE-588
Project: Hadoop HBase
Issue Type: Bug
Reporter: stack
Priority: Blocker
Before HBASE-532, as soon as a flush started, we called snapshot. Snapshot
used to copy current live memcache into a 'snapshot' TreeMap inside in
Memcache. This snapshot TreeMap was an accumulation of all snapshots since
last flush. Whenever we took out a scanner, we'd do a copy of this snapshot
into a new backing map carried by the scanner (Every outstanding Scanner had
complete copy). Memcache snapshots were cleared when a flush started.
Flushing could take near no time to up to tens of seconds during which an
scanners taken out meantime would not see the edits in the snapshot currently
being flushed and gets or getFull would also return incorrect answers because
the content of the snapshot was not available to them.
HBASE-532 made it so the snapshot was available until flush was done -- until a
file had made it out to disk. This fixed gets and getFull and any scanners
taken out during flushing. But there is still a hole. Any outstanding
scanners will be going against the state of Store Readers at time scanner was
opened; they will not see the new flush file.
Chatting about this on IRC, Jim suggests that we pass either memcache or
current snapshot to each Scanner (Pass the snapshot if not empty). The notion
is that the Scanner would hold on to the Scanner reference should it be cleared
by flushing. Upside is that scanner wouldn't have to be concerned with the new
flush that has been put out to disk. Downsides are that Scanner data could be
way stale if for instance the memcache was near to flushing but we hadn't done
it yet. And we wouldn't be clearing the snapshot promptly so would be some
memory pressure.
Another suggestion is that flushing send an event. Listeners such as
outstanding scanners would notice event and open the new Reader. Would have to
skip forward in the new Reader to catch up with the current set but shouldn't
be bad. Same mechanism could be used to let compactions be moved into place
while scanners were outstanding closing down all existing readers skipping to
the current 'next' location in the new compacted store file.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
