Re: False positive increasing
Jeff is probably correct. I skimmed over the fact that it's just increasing by one every few minutes so I went on about a different scenario. On Mon, Jul 3, 2017, at 01:46 PM, Jeff Jirsa wrote: > > > On 2017-07-03 06:55 (-0700), Jean Carlo> wrote: > > Hello > > > > Lately I am observing that the false positives of one of my nodes are > > increasing in a continous way (1 per 5min) > > > > There's probably one partition that has a false positive entry, and you > read it once every 5 minutes. Bloom filters are probabilistic, false > positives are OK, it just causes a little bit of extra disk IO. > > > Bloom filter false positives: 532 > > Bloom filter false ratio: 0.01449 > > Bloom filter space used: 1.34 MB > > Bloom filter off heap memory used: 1.33 MB > > > > At the same time I can see that the duration of GC has increased also > > > > There is a link between the increasment of the GC and the bloom filter ? > > > > Probably not in any meaningful way (like mentioned above, false positive > causes some extra disk IO to check one extra sstable, but it's not going > to really impact GC in any meaningful way if it's truly a false > positive). > > > - > To unsubscribe, e-mail: user-unsubscr...@cassandra.apache.org > For additional commands, e-mail: user-h...@cassandra.apache.org > - To unsubscribe, e-mail: user-unsubscr...@cassandra.apache.org For additional commands, e-mail: user-h...@cassandra.apache.org
Re: False positive increasing
On 2017-07-03 06:55 (-0700), Jean Carlowrote: > Hello > > Lately I am observing that the false positives of one of my nodes are > increasing in a continous way (1 per 5min) > There's probably one partition that has a false positive entry, and you read it once every 5 minutes. Bloom filters are probabilistic, false positives are OK, it just causes a little bit of extra disk IO. > Bloom filter false positives: 532 > Bloom filter false ratio: 0.01449 > Bloom filter space used: 1.34 MB > Bloom filter off heap memory used: 1.33 MB > > At the same time I can see that the duration of GC has increased also > > There is a link between the increasment of the GC and the bloom filter ? > Probably not in any meaningful way (like mentioned above, false positive causes some extra disk IO to check one extra sstable, but it's not going to really impact GC in any meaningful way if it's truly a false positive). - To unsubscribe, e-mail: user-unsubscr...@cassandra.apache.org For additional commands, e-mail: user-h...@cassandra.apache.org
Re: False positive increasing
Hi, The number of false positives may be increasing because more filters are being consulted on each query. The number of filters consulted on each query is a function of number of sstables consulted. You may be seeing an increase in number of tables consulted if compaction is falling behind. I'm not an expert on the operational playbook for compaction falling behind, but you can change the compaction throttle, disable gossip so compaction can catch up (this can go wrong), add capacity by adding nodes. If it's just one node you may also want to look into why that node is a hot spot. Is there a single large partition that could be causing issues? Ariel On Mon, Jul 3, 2017, at 09:55 AM, Jean Carlo wrote: > Hello > Lately I am observing that the false positives of one of my nodes are > increasing in a continous way (1 per 5min)> > Bloom filter false positives: 532 > Bloom filter false ratio: 0.01449 > Bloom filter space used: 1.34 MB > Bloom filter off heap memory used: 1.33 MB > At the same time I can see that the duration of GC has increased also> > There is a link between the increasment of the GC and the bloom > filter ?> > Jean Carlo >
False positive increasing
Hello Lately I am observing that the false positives of one of my nodes are increasing in a continous way (1 per 5min) Bloom filter false positives: 532 Bloom filter false ratio: 0.01449 Bloom filter space used: 1.34 MB Bloom filter off heap memory used: 1.33 MB At the same time I can see that the duration of GC has increased also There is a link between the increasment of the GC and the bloom filter ? Jean Carlo