After a ZK crash, or client timeout sometimes it's hard to determine from the logs what happened. Knowing if ZK was responsive at the time would help a lot. For example, ZK might spend a lot of time waiting on GC (there is still some misconception that ZK is a storage).
To help detect this, HADOOP already has a great tool called JVM Pause Monitor. (As the name suggest, it can be also used for monitoring, but it also helps post-mortem in a lot of cases). Basically it has a daemon that sleeps for one second, and if the sleep time exceeds the 1s by more than the threshold (1s: INFO, 10s: WARN by default - this can be configurable in our case, see below), it will alert/make a log entry. It can also monitor the time GC took.
The class implementing this is in HADOOP-common, but ZK should not depend on this package. Since this is a straightforward implementation, and in the past five years the few commits it had is nothing really serious, I think we could just copy this class in ZooKeeper, and introduce it as a configurable feature, by default it can be off.
- Create a class in ZK (under zookeeper/server/util/) called JvmPauseMonitor.
- Make feature configurable, by default: OFF
- Make sleep time and threshold time configurable
- Update documentation
- Add [current size of the heap OR % of heap used] in the log entry whenever sleep threshold had exceeded by a lot (10s)