Aljoscha, Thank you for your quick response. Yes, I am using FsStateBackend, so I will try RocksDB backend.
Regards, Hironori 2016-04-05 21:23 GMT+09:00 Aljoscha Krettek <aljos...@apache.org>: > Hi, > I guess you are using the FsStateBackend, is that correct? You could try > using the RocksDB state backend: > https://ci.apache.org/projects/flink/flink-docs-master/apis/streaming/state_backends.html#the-rocksdbstatebackend > > With this, throughput will be lower but the overhead per checkpoint could be > lower. Also, with this most of the file copying necessary for the checkpoint > will be done while data processing keeps running (asynchronous snapshot). > > As to incremental snapshots. I'm afraid this feature is not yet implemented > but we're working on it. > > Cheers, > Aljoscha > > On Tue, 5 Apr 2016 at 14:06 Hironori Ogibayashi <ogibaya...@gmail.com> > wrote: >> >> Hello, >> >> I am trying to implement windowed distinct count on a stream. In this >> case, the state >> have to hold all distinct value in the window, so can be large. >> >> In my test, if the state size become about 400MB, checkpointing takes >> 40sec and spends most of Taskmanager's CPU. >> Are there any good way to handle this situation? >> >> Flink document mentions about incremental snapshot, and I am interested in >> it, >> but could not find how to enable it. (not implemented yet?) >> >> https://ci.apache.org/projects/flink/flink-docs-release-1.0/internals/stream_checkpointing.html >> >> Regards, >> Hironori