Even more interesting behavior : a repair on a CF has consequences on other CFs. I didn't expect that.
There are no writes being issued to the cluster yet the logs indicate that - SSTableReader has opened dozens and dozens of files, most of them unrelated to the CF being repaired - compactions are taking place continuously on CFs other than the one being repaired, even CFs in other keyspaces - I see "Sending AEService tree" messages for CF not being repaired. After a very long time, I got some AES messages indicating that streaming from node C had finished and then many minutes after that node B. And yet the pending stream count on node B hasn't changed. The *-data.db files for the CF being repaired are about 70MB on-disk. Maybe when a stream is fully received on node B, netstats indicates that no streams are pending but since they are not acknowledged, node A doesn't ? 2011/8/16 Philippe <watche...@gmail.com> > I'm still trying different stuff. Here are my latest findings, maybe > someone will find them useful: > > - I have been able to repair some small column families by issuing a > repair [KS] [CF]. When testing on the ring with no writes at all, it still > takes about 2 repairs to get "consistent" logs for all AES requests. > - Launching a repair one the smallest CF of the biggest KS has > triggered a flurry of compactions and streams. Some of those streams are > for > other CF in that keyspace !? > - During repairs (one at a time cluster-wide), I get 25-50% io waits & > 35%-50% cpu usage on a 6 core SATA-disk setup > > What is surprising to me (bug?) is that netstats shows me streams going > from node A to node B at 0% progress. But netstats on node B doesn't show me > any streams coming in. I'm thinking that repairs may be never ending and > that may be messing up my compactions hence the huge pile up of compactions > until the disk fulls. > I know there's an issue related to failed streams & repairs, could I be > hitting it ? > > Thanks > > 2011/8/14 Philippe <watche...@gmail.com> > >> @Teijo : thanks for the procedure, I hope I won't have to do that >> >> Peter, I'll answer inline. Thanks for the detailed answer. >> >> >>> > the number of SSTables for some keyspaces goes dramatically up (from 3 >>> or 4 >>> > to several dozens). >>> >>> Typically with a long running compaction, such as that triggered by >>> repair, that's what happens as flushed memtables accumulate. In >>> particular for memtables with frequent flushes. >>> >>> Are you running with concurrent compaction enabled? >>> >> Yes, it is enabled. On my 0.8 cluster, cassandra.yaml has this (it's >> commented). BTW, I have 6 cores on each server. >> >> #concurrent_compactors: 1 >> >> > the commit log keeps increasing in size, I'm at 4.3G now, it went up to >>> 40G >>> > when the compaction was throttled at 16MB/s. On the other nodes it's >>> around >>> > 1GB at most >>> Hmmmm. The Commit Log should not be retained longer than what is >>> required for memtables to be flushed. Is it possible you have had an >>> out-of-disk condition and flushing has stalled? Are you seeing flushes >>> happening in the log? >>> >> No I don't believe there was ever an out of disk. Yes it is flushing for >> the first couple of hours. >> Then, when repair seems locked up, my log is mostly filled with lines such >> as this >> INFO [ScheduledTasks:1] 2011-08-14 23:15:47,267 StatusLogger.java (line >> 88) [My_Keyspace].[My_Columnfamily] 45,105541 50/50 >> 20/20 >> Why is that ? >> >> > the data directory is bigger than on the other nodes. I've seen it go up >>> to >>> > 480GB when the compaction was throttled at 16MB/s >>> How much data are you writing? Is it at all plausible that the huge >>> spike is a reflection of lots of overwriting writes that aren't being >>> compacted? >>> >> No, there's no bulk loading going on at the moment and I'm pretty sure >> there wasn't when it spiked up to that load. >> I've never measured the load because it's a mix of counter increments and >> new counters all the time. It's not that much though. >> >> >>> Normally when disk space spikes with repair it's due to other nodes >>> streaming huge amounts (maybe all of their data) to the node, leading >>> to a temporary spike. But if your "real" size is expected to be 60, >>> 480 sounds excessive. Are you sure other nodes aren't running repairs >>> at the same time and magnifying each other's data load spikes? >>> >> Yes, the two other nodes were running repairs. I had them scheduled at 8 >> hour intervals but they must have started. >> When data is streamed from one to another, does that data go into the >> commit log as a regular write ? >> How much of a negative impact can that have on the repair going on on >> this node ? >> >> > What's even weirder is that currently I have 9 compactions running but >>> CPU >>> > is throttled at 1/number of cores half the time (while > 80% the rest >>> of the >>> > time). Could this be because other repairs are happening in the ring ? >>> You mean compaction is taking less CPU than it "should"? >>> >> Yes >> >> >>> No, this should not be due to other nodes repairing. However it sounds >>> to me like you are bottlenecking on I/O and the repairs and >>> >> Yes, I/O is really high on the node right now. Around 50% I/O waits. >> >> >>> compactions are probably proceeding extremely slowly, probably being >>> completely drowned out by live traffic (which is probably having an >>> abnormally high performance impact due to data size spike). >>> >> Yes, the live traffic is 3 to 10x times slower during repair. Ouch... I >> hope I won't to do this too often while in production ! >> >> >>> >>> What's your read concurrency configured on the node? What does "iostat >>> -x -k 1" show in the average queue size column? >> >> Average queue size on the disk (RAID-1 + separate LVM volumes for data, >> commit log, caches, logs)) varies between 2 and 90. I'd say the average is >> around 30-40. Very high variation. >> >> >>> Is "nodetool -h >>> localhost tpstats" showing that ReadStage is usually "full" (@ your >>> limit)? >>> >> No backlog at all in tpstats >> >> I've figured out how AES is logging its actions and it looks like it >> really is going through every CF in every keyspace and doing a tree request >> for every token range >> So it really looks like it's just taking forever to compact stuff as it's >> repairing. >> I saw in another email that repairing was taking 2-3mn/ GB... it looks >> like a lot more for my ring. Anybody else have numbers ? >> >> Thanks >> > >