Hi Saad Your argument here >> The >>theory is that since prefetch is an async operation, a lot of the reads in >>the checkAndPut for the region in question start reading from S3 which is >>slow. So the write lock obtained for the checkAndPut is held for a longer >>duration than normal. This has cascading upstream effects. Does that sound >>plausible?
Seems very much plausible. So before even the prefetch happens say for 'block 1' - and you have already issues N checkAndPut calls for the rows in that 'block 1' - all those checkAndPut will have to read that block from S3 to perform the get() and then apply the mutation. This may happen for multiple threads at the same time because we are not sure when the prefetch would have actually been completed. I don know what are the general read characteristics when a read happens from S3 but you could try to see how things work when a read happens from S3 and after the prefetch completes ensure the same checkandPut() is done (from cache this time) to really know the difference what S3 does there. Regards Ram On Fri, Mar 2, 2018 at 2:57 AM, Saad Mufti <saad.mu...@gmail.com> wrote: > So after much investigation I can confirm: > > a) it was indeed one of the regions that was being compacted, major > compaction in one case, minor compaction in another, the issue started just > after compaction completed blowing away bucket cached blocks for the older > HFile's > b) in another case there was no compaction just a newly opened region in a > region server that hadn't finished perfetching its pages from S3 > > We have prefetch on open set to true. Our load is heavy on checkAndPut .The > theory is that since prefetch is an async operation, a lot of the reads in > the checkAndPut for the region in question start reading from S3 which is > slow. So the write lock obtained for the checkAndPut is held for a longer > duration than normal. This has cascading upstream effects. Does that sound > plausible? > > The part I don't understand still is all the locks held are for the same > region but are all for different rows. So once the prefetch is completed, > shouldn't the problem clear up quickly? Or does the slow region slow down > anyone trying to do checkAndPut on any row in the same region even after > the prefetch has completed. That is, do the long held row locks prevent > others from getting a row lock on a different row in the same region? > > In any case, we trying to use > https://issues.apache.org/jira/browse/HBASE-16388 support in HBase 1.4.0 > to > both insulate the app a bit from this situation and hoping that it will > reduce pressure on the region server in question, allowing it to recover > faster. I haven't quite tested that yet, any advice in the meantime would > be appreciated. > > Cheers. > > ---- > Saad > > > > On Thu, Mar 1, 2018 at 9:21 AM, Saad Mufti <saad.mu...@gmail.com> wrote: > > > Actually it happened again while some minior compactions were running, so > > don't think it related to our major compaction tool, which isn't even > > running right now. I will try to capture a debug dump of threads and > > everything while the event is ongoing. Seems to last at least half an > hour > > or so and sometimes longer. > > > > ---- > > Saad > > > > > > On Thu, Mar 1, 2018 at 7:54 AM, Saad Mufti <saad.mu...@gmail.com> wrote: > > > >> Unfortunately I lost the stack trace overnight. But it does seem related > >> to compaction, because now that the compaction tool is done, I don't see > >> the issue anymore. I will run our incremental major compaction tool > again > >> and see if I can reproduce the issue. > >> > >> On the plus side the system stayed stable and eventually recovered, > >> although it did suffer all those timeouts. > >> > >> ---- > >> Saad > >> > >> > >> On Wed, Feb 28, 2018 at 10:18 PM, Saad Mufti <saad.mu...@gmail.com> > >> wrote: > >> > >>> I'll paste a thread dump later, writing this from my phone :-) > >>> > >>> So the same issue has happened at different times for different > regions, > >>> but I couldn't see that the region in question was the one being > compacted, > >>> either this time or earlier. Although I might have missed an earlier > >>> correlation in the logs where the issue started just after the > compaction > >>> completed. > >>> > >>> Usually a compaction for this table's regions take around 5-10 minutes, > >>> much less for its smaller column family which is block cache enabled, > >>> around a minute or less, and 5-10 minutes for the much larger one for > which > >>> we have block cache disabled in the schema, because we don't ever read > it > >>> in the primary cluster. So the only impact on reads would be from that > >>> smaller column family which takes less than a minute to compact. > >>> > >>> But the issue once started doesn't seem to recover for a long time, > long > >>> past when any compaction on the region itself could impact anything. > The > >>> compaction tool which is our own code has long since moved to other > >>> regions. > >>> > >>> Cheers. > >>> > >>> ---- > >>> Saad > >>> > >>> > >>> On Wed, Feb 28, 2018 at 9:39 PM Ted Yu <yuzhih...@gmail.com> wrote: > >>> > >>>> bq. timing out trying to obtain write locks on rows in that region. > >>>> > >>>> Can you confirm that the region under contention was the one being > major > >>>> compacted ? > >>>> > >>>> Can you pastebin thread dump so that we can have better idea of the > >>>> scenario ? > >>>> > >>>> For the region being compacted, how long would the compaction take > (just > >>>> want to see if there was correlation between this duration and > timeout) > >>>> ? > >>>> > >>>> Cheers > >>>> > >>>> On Wed, Feb 28, 2018 at 6:31 PM, Saad Mufti <saad.mu...@gmail.com> > >>>> wrote: > >>>> > >>>> > Hi, > >>>> > > >>>> > We are running on Amazon EMR based HBase 1.4.0 . We are currently > >>>> seeing a > >>>> > situation where sometimes a particular region gets into a situation > >>>> where a > >>>> > lot of write requests to any row in that region timeout saying they > >>>> failed > >>>> > to obtain a lock on a row in a region and eventually they experience > >>>> an IPC > >>>> > timeout. This causes the IPC queue to blow up in size as requests > get > >>>> > backed up, and that region server experiences a much higher than > >>>> normal > >>>> > timeout rate for all requests, not just those timing out for failing > >>>> to > >>>> > obtain the row lock. > >>>> > > >>>> > The strange thing is the rows are always different but the region is > >>>> always > >>>> > the same. So the question is, is there a region component to how > long > >>>> a row > >>>> > write lock would be held? I looked at the debug dump and the > RowLocks > >>>> > section shows a long list of write row locks held, all of them are > >>>> from the > >>>> > same region but different rows. > >>>> > > >>>> > Will trying to obtain a write row lock experience delays if no one > >>>> else > >>>> > holds a lock on the same row but the region itself is experiencing > >>>> read > >>>> > delays? We do have an incremental compaction tool running that major > >>>> > compacts one region per region server at a time, so that will drive > >>>> out > >>>> > pages from the bucket cache. But for most regions the impact is > >>>> > transitional until the bucket cache gets populated by pages from the > >>>> new > >>>> > HFile. But for this one region we start timing out trying to obtain > >>>> write > >>>> > locks on rows in that region. > >>>> > > >>>> > Any insight anyone can provide would be most welcome. > >>>> > > >>>> > Cheers. > >>>> > > >>>> > ---- > >>>> > Saad > >>>> > > >>>> > >>> > >> > > >
