hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ryan Rawson <ryano...@gmail.com>
Subject Re: [jira] Updated: (HBASE-1563) incrementColumnValue does not write to WAL
Date Tue, 23 Jun 2009 22:58:00 GMT
we're going to have to rethink the inplace increment. That causes a
potential race. I want to write a bench suite to try to replicate it.

On Jun 23, 2009 3:45 PM, "Jonathan Gray (JIRA)" <jira@apache.org> wrote:

   Attachment: HBASE-1563-v3.patch

- Removes unnecessary Get in client code
- Removed duplicate server calls in the client code, default just calls full
w/ true
- Replaced increment javadoc with "Atomically increments a column value. If
the column value already exists and is not a big-endian long, this could
throw an exception."  Is that more clear?

@Ryan I got most the way through splitting up Store.incrementColumnValue
logic so we figure out what we have to do first, then write to wal, then
perform the insert.  The issue is when we increment in-place in the
memcache.  If we truly don't want to write until the wal append, we'd have
to make a copy of the KV, perform the increment on it, and then either
increment the original memcache value or swap.  This more or less negates
the in-place increment optimization.  Thoughts?

> incrementColumnValue does not write to WAL >
------------------------------------------ > > ...
>         Attachments: HBASE-1563-v1.patch, HBASE-1563-v2.patch,

> > > Incrementing never writes to the WAL. Under failure scenarios, you
will lose all increments s...

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message