That's an interesting thought. My code runs in FCGI and although the cassandra connection is used to serve multiple requests, those requests are supposedly processed sequentially, in a while ($request->Accept() >= 0) loop. However, we do call FCGI::finish to close the request (so the HTTP request would finish) while continue writing to cassandra. I wonder if that may cause the next request to be accepted and trump the current one? I don't think it should but it's worth testing.
Thanks, - Jianing On Thu, Jul 29, 2010 at 6:23 PM, Aaron Morton <aa...@thelastpickle.com> wrote: > I was accidentally sharing connections between threads, and getting strange > results. Is your client multi threaded? > > Can you provide some more information, such as the client library, how the > data is written and how you're deciding that the returned results are the > wrong ones. > > Is the read inconsistency against data that is frequently changing? Could it > be a problem with the way the data is being stored? > > Aaron > > > > > On 30 Jul, 2010,at 11:43 AM, Jianing Hu <jian...@gmail.com> wrote: > > Hi Aaron, > > Thanks for the reply. Can you explain what you mean by "sharing > connections around"? > > I'm just calling a simple "get", and the data returned is for a > completely different key. It's intermittent and hard to produce in my > test environment, but can be observed in our production environment > couple hundred times a day. > > Thanks, > - Jianing > > On Thu, Jul 29, 2010 at 2:58 PM, Aaron Morton <aa...@thelastpickle.com> > wrote: >> I noticed this once when accidentally sharing connections around. Could >> that >> be the case ? >> >> What sort of commands are you running ? Could you be seeing this problem ? >> http://www.mail-archive.com/user@cassandra.apache.org/msg04831.html >> >> Aaron >> >> >> On 29 Jul, 2010,at 12:47 PM, Jianing Hu <jian...@gmail.com> wrote: >> >> We recently migrated part of our MySQL database to a 3-node Cassandra >> cluster with a replication factor of 3. Couple of days ago we noticed >> that Cassandra sometimes returns the wrong data. Not corrupted data, >> but data for a different key than the one being asked for. This error >> appears to be random and intermittent, and happens for maybe every 10K >> reads. I'm working on a test suite that can reproduce this. Meanwhile >> just wanted to ask if anyone has seen this problem before? I've tried >> consistency levels of both QUORUM and ONE, and see the issue with >> both. >> >> Thanks, >> - Jianing >> >
