Hi Ufuk, Yes, it does help with Rocksdb backend! After tune checkpoint frequency align with network throughput, task manager released and job get cancelled are gone.
Chen > On May 10, 2016, at 10:33 AM, Ufuk Celebi <u...@apache.org> wrote: > >> On Tue, May 10, 2016 at 5:07 PM, Chen Qin <qinnc...@gmail.com> wrote: >> Future, to keep large key/value space, wiki point out using rocksdb as >> backend. My understanding is using rocksdb will write to local file systems >> instead of sync to s3. Does flink support memory->rocksdb(local disk)->s3 >> checkpoint state split yet? Or would implement kvstate interface makes flink >> take care of large state problem? > > Hey Chen, > > when you use RocksDB, you only need to explicitly configure the file > system checkpoint directory, for which you can use S3: > > new RocksDBStateBackend(new URI("s3://...")) > > The local disk path are configured via the general Flink temp > directory configuration (see taskmanager.tmp.dirs in > https://ci.apache.org/projects/flink/flink-docs-release-1.0/setup/config.html, > default is /tmp). > > State is written to the local RocksDB instance and the RocksDB files > are copied to S3 on checkpoints. > > Does this help? > > – Ufuk