Hey Stack, Along with the log replaying part, logic is also needed for log roll over. This, I think, easier compared to the merging of the logs. Any edits less than the last sequence number on the file system can be removed from all the WALs.
Cheers, Karthik On 1 October 2011 18:05, Jesse Yates <[email protected]> wrote: > I think adding the abstraction layer and making it not only pluggable, but > configurable would be great. > > It would be nice to be able to tie into a service that logs directly to > disk, rather than go through HDFS giving some potentially awesome speedup > at > the cost of having to write a logging service that handles replication, > etc. > Side note, Accumulo is using their own service to storing the WAL, rather > than HDFS and I suspect that plays a big role in people's claim of its > ability to do 'outperform' HBase. > > -Jesse Yates > > On Sat, Oct 1, 2011 at 2:04 PM, Stack <[email protected]> wrote: > > > Yes. For sure. Would need to check that the split can deal w/ > > multiple logs written by the one server concurrently (sort by sequence > > edit id after sorting on all the rest that makes up a wal log key). > > > > St.Ack > > > > On Sat, Oct 1, 2011 at 1:36 PM, karthik tunga <[email protected]> > > wrote: > > > Hey, > > > > > > Doesn't multiple WALs need some kind of merging when recovering from a > > crash > > > ? > > > > > > Cheers, > > > Karthik > > > > > > > > > On 1 October 2011 15:17, Stack <[email protected]> wrote: > > > > > >> +1 on making WAL pluggable so we can experiment. Being able to write > > >> multiple WALs at once should be easy enough to do (the WAL split code > > >> should be able to handle it). Also a suggestion made a while back was > > >> making it so hbase could be configured to write two filesystems -- > > >> there'd be hbase.rootdir as now -- and then we'd allow specifying > > >> another fs to use for writing WALs (If not specified, we'd just use > > >> hbase.rootdir for all filesystem interactions as now). > > >> > > >> St.Ack > > >> > > >> On Sat, Oct 1, 2011 at 10:56 AM, Dhruba Borthakur <[email protected]> > > >> wrote: > > >> > I have been experimenting with the WAL settings too. It is obvious > > that > > >> > turning off the wal makes ur transactions go faster, HDFS write/sync > > are > > >> not > > >> > yet very optimized for high throughput small writes. > > >> > > > >> > However, irrespective of whether I have one wal or two, I have > seeing > > the > > >> > same throughput. I have experimented with an HDFS setting that > allows > > >> > writing/sync to multiple replicas in parallel, and that has > increased > > >> > performance for my test workload, see > > >> > https://issues.apache.org/jira/browse/HDFS-1783. > > >> > > > >> > About using one wal or two, it will be nice if we can separate out > the > > >> wal > > >> > API elegantly and make it pluggable. In that case, we can experiment > > >> HBase > > >> > with multiple systems. Once we have it pluggable, we can make the > > habse > > >> wal > > >> > go to a separate HDFS (pure SSD based maybe?). > > >> > > > >> > -dhruba > > >> > > > >> > > > >> > On Sat, Oct 1, 2011 at 8:09 AM, Akash Ashok <[email protected] > > > > >> wrote: > > >> > > > >> >> Hey, > > >> >> I've see that setting writeToWAL(false) boosts up the writes like > > crazy. > > >> I > > >> >> was just thinking having MuiltipleWAL on HBase. I understand that > > this > > >> is a > > >> >> consideration in BigTable paper that a WAL per region is not used > > >> because > > >> >> it > > >> >> might result in a lot of disk seeks when there are large number of > > >> reasons. > > >> >> But how about having as many WALs as the number of HardDrives in > the > > >> >> system. > > >> >> I see that the recommended configs for HBase are 4 - 12 hard drives > > per > > >> >> node. This might kick the writes up a notch. > > >> >> > > >> >> Would like to know the general opinion on this one? > > >> >> > > >> >> Cheers, > > >> >> Akash A > > >> >> > > >> > > > >> > > > >> > > > >> > -- > > >> > Connect to me at http://www.facebook.com/dhruba > > >> > > > >> > > > > > >
