3G of memory should not lead to long GC pauses unless you're running very close to the edge of available memory. Paradoxically, running with 6G of memory may lead to _fewer_ noticeable pauses since the background threads can do the work, well, in the background.
Best, Erick On Mon, May 8, 2017 at 7:29 AM, Satya Marivada <satya.chaita...@gmail.com> wrote: > Hi Piyush and Shawn, > > May I ask what is the solution for it, if it is the long gc pauses? I am > skeptical about the same problem in our case too. We have started with 3G > of memory for the heap. > Did you have to adjust some of the memory allotted? Very much appreciated. > > Thanks, > Satya > > On Sat, May 6, 2017 at 12:36 PM Piyush Kunal <piyush.ku...@myntra.com> > wrote: > >> We already faced this issue and found out the issue to be long GC pauses >> itself on either client side or server side. >> Regards, >> Piyush >> >> On Sat, May 6, 2017 at 6:10 PM, Shawn Heisey <apa...@elyograg.org> wrote: >> >> > On 5/3/2017 7:32 AM, Satya Marivada wrote: >> > > I see below exceptions in my logs sometimes. What could be causing it? >> > > >> > > org.apache.zookeeper.KeeperException$SessionExpiredException: >> > >> > Based on my limited research, this would tend to indicate that the >> > heartbeats ZK uses to detect when sessions have gone inactive are not >> > occurring in a timely fashion. >> > >> > Common causes seem to be: >> > >> > JVM Garbage collections. These can cause the entire JVM to pause for an >> > extended period of time, and this time may exceed the configured >> timeouts. >> > >> > Excess client connections to ZK. ZK limits the number of connections >> > from each client address, with the idea of preventing denial of service >> > attacks. If a client is misbehaving, it may make more connections than >> > it should. You can try increasing the limit in the ZK config, but if >> > this is the reason for the exception, then something's probably wrong, >> > and you may be just hiding the real problem. >> > >> > Although we might have bugs causing the second situation, the first >> > situation seems more likely. >> > >> > Thanks, >> > Shawn >> > >> > >>
