Are you guys using light weight transactions in your write path? On Thu, Feb 11, 2016 at 12:36 AM, Fabrice Facorat <fabrice.faco...@gmail.com > wrote:
> Are your commitlog and data on the same disk ? If yes, you should put > commitlogs on a separate disk which don't have a lot of IO. > > Others IO may have great impact impact on your commitlog writing and > it may even block. > > An example of impact IO may have, even for Async writes: > > https://engineering.linkedin.com/blog/2016/02/eliminating-large-jvm-gc-pauses-caused-by-background-io-traffic > > 2016-02-11 0:31 GMT+01:00 Mike Heffner <m...@librato.com>: > > Jeff, > > > > We have both commitlog and data on a 4TB EBS with 10k IOPS. > > > > Mike > > > > On Wed, Feb 10, 2016 at 5:28 PM, Jeff Jirsa <jeff.ji...@crowdstrike.com> > > wrote: > >> > >> What disk size are you using? > >> > >> > >> > >> From: Mike Heffner > >> Reply-To: "user@cassandra.apache.org" > >> Date: Wednesday, February 10, 2016 at 2:24 PM > >> To: "user@cassandra.apache.org" > >> Cc: Peter Norton > >> Subject: Re: Debugging write timeouts on Cassandra 2.2.5 > >> > >> Paulo, > >> > >> Thanks for the suggestion, we ran some tests against CMS and saw the > same > >> timeouts. On that note though, we are going to try doubling the instance > >> sizes and testing with double the heap (even though current usage is > low). > >> > >> Mike > >> > >> On Wed, Feb 10, 2016 at 3:40 PM, Paulo Motta <pauloricard...@gmail.com> > >> wrote: > >>> > >>> Are you using the same GC settings as the staging 2.0 cluster? If not, > >>> could you try using the default GC settings (CMS) and see if that > changes > >>> anything? This is just a wild guess, but there were reports before of > >>> G1-caused instabilities with small heap sizes (< 16GB - see > CASSANDRA-10403 > >>> for more context). Please ignore if you already tried reverting back > to CMS. > >>> > >>> 2016-02-10 16:51 GMT-03:00 Mike Heffner <m...@librato.com>: > >>>> > >>>> Hi all, > >>>> > >>>> We've recently embarked on a project to update our Cassandra > >>>> infrastructure running on EC2. We are long time users of 2.0.x and are > >>>> testing out a move to version 2.2.5 running on VPC with EBS. Our test > setup > >>>> is a 3 node, RF=3 cluster supporting a small write load (mirror of our > >>>> staging load). > >>>> > >>>> We are writing at QUORUM and while p95's look good compared to our > >>>> staging 2.0.x cluster, we are seeing frequent write operations that > time out > >>>> at the max write_request_timeout_in_ms (10 seconds). CPU across the > cluster > >>>> is < 10% and EBS write load is < 100 IOPS. Cassandra is running with > the > >>>> Oracle JDK 8u60 and we're using G1GC and any GC pauses are less than > 500ms. > >>>> > >>>> We run on c4.2xl instances with GP2 EBS attached storage for data and > >>>> commitlog directories. The nodes are using EC2 enhanced networking > and have > >>>> the latest Intel network driver module. We are running on HVM > instances > >>>> using Ubuntu 14.04.2. > >>>> > >>>> Our schema is 5 tables, all with COMPACT STORAGE. Each table is > similar > >>>> to the definition here: > >>>> https://gist.github.com/mheffner/4d80f6b53ccaa24cc20a > >>>> > >>>> This is our cassandra.yaml: > >>>> > https://gist.github.com/mheffner/fea80e6e939dd483f94f#file-cassandra-yaml > >>>> > >>>> Like I mentioned we use 8u60 with G1GC and have used many of the GC > >>>> settings in Al Tobey's tuning guide. This is our upstart config with > JVM and > >>>> other CPU settings: > https://gist.github.com/mheffner/dc44613620b25c4fa46d > >>>> > >>>> We've used several of the sysctl settings from Al's guide as well: > >>>> https://gist.github.com/mheffner/ea40d58f58a517028152 > >>>> > >>>> Our client application is able to write using either Thrift batches > >>>> using Asytanax driver or CQL async INSERT's using the Datastax Java > driver. > >>>> > >>>> For testing against Thrift (our legacy infra uses this) we write > batches > >>>> of anywhere from 6 to 1500 rows at a time. Our p99 for batch > execution is > >>>> around 45ms but our maximum (p100) sits less than 150ms except when it > >>>> periodically spikes to the full 10seconds. > >>>> > >>>> Testing the same write path using CQL writes instead demonstrates > >>>> similar behavior. Low p99s except for periodic full timeouts. We > enabled > >>>> tracing for several operations but were unable to get a trace that > completed > >>>> successfully -- Cassandra started logging many messages as: > >>>> > >>>> INFO [ScheduledTasks:1] - MessagingService.java:946 - _TRACE messages > >>>> were dropped in last 5000 ms: 52499 for internal timeout and 0 for > cross > >>>> node timeout > >>>> > >>>> And all the traces contained rows with a "null" source_elapsed row: > >>>> > https://gist.githubusercontent.com/mheffner/1d68a70449bd6688a010/raw/0327d7d3d94c3a93af02b64212e3b7e7d8f2911b/trace.out > >>>> > >>>> > >>>> We've exhausted as many configuration option permutations that we can > >>>> think of. This cluster does not appear to be under any significant > load and > >>>> latencies seem to largely fall in two bands: low normal or max > timeout. This > >>>> seems to imply that something is getting stuck and timing out at the > max > >>>> write timeout. > >>>> > >>>> Any suggestions on what to look for? We had debug enabled for awhile > but > >>>> we didn't see any msg that pointed to something obvious. Happy to > provide > >>>> any more information that may help. > >>>> > >>>> We are pretty much at the point of sprinkling debug around the code to > >>>> track down what could be blocking. > >>>> > >>>> > >>>> Thanks, > >>>> > >>>> Mike > >>>> > >>>> -- > >>>> > >>>> Mike Heffner <m...@librato.com> > >>>> Librato, Inc. > >>>> > >>> > >> > >> > >> > >> -- > >> > >> Mike Heffner <m...@librato.com> > >> Librato, Inc. > >> > > > > > > > > -- > > > > Mike Heffner <m...@librato.com> > > Librato, Inc. > > > > > > -- > Close the World, Open the Net > http://www.linux-wizard.net >
