Re: [DISCUSS] Semantic and implementation of per-job mode

2019-10-30 Thread SHI Xiaogang
Hi Thanks for bringing this. The design looks very nice to me in that 1. In the new per-job mode, we don't need to compile user programs in the client and can directly run user programs with user jars. That way, it's easier for resource isolation in multi-tenant platforms and is much safer. 2. Th

Re: [ANNOUNCE] Zili Chen becomes a Flink committer

2019-09-11 Thread SHI Xiaogang
Congratulations! Regards, Xiaogang Guowei Ma 于2019年9月11日周三 下午7:07写道: > Congratulations Zili ! > > Best, > Guowei > > > Fabian Hueske 于2019年9月11日周三 下午7:02写道: > >> Congrats Zili Chen :-) >> >> Cheers, Fabian >> >> Am Mi., 11. Sept. 2019 um 12:48 Uhr schrieb Biao Liu > >: >> >>> Congrats Zili! >>

Re: How to handle Flink Job with 400MB+ Uberjar with 800+ containers ?

2019-08-30 Thread SHI Xiaogang
ready started TaskManagers in P2P fashion, not to have a > blocker on HDFS replication. > > Spark job without any tuning exact same size jar with 800 executors, can > start without any issue at the same cluster in less than a minute. > > *Further questions:* > > *@ S

Re: How to handle Flink Job with 400MB+ Uberjar with 800+ containers ?

2019-08-29 Thread SHI Xiaogang
Hi Datashov, We faced similar problems in our production clusters. Now both lauching and stopping of containers are performed in the main thread of YarnResourceManager. As containers are launched and stopped one after another, it usually takes long time to boostrap large jobs. Things get worse wh

Re: S3 recovery and checkpoint directories exhibit explosive growth

2017-07-16 Thread SHI Xiaogang
Hi Prashantnayak Thanks a lot for reporting this problem. Can you provide more details to address it? I am guessing master has to delete too many files when a checkpoint is subsumed, which is very common in our cases. The number of files in the recovery directory will increase if the master canno

Re: Checkpointing with RocksDB as statebackend

2017-06-28 Thread SHI Xiaogang
Hi Vinay, We observed a similar problem before. We found that RocksDB keeps a lot of index and filter blocks in memory. With the growth in state size (in our cases, most states are only cleared in windowed streams), these blocks will occupy much more memory. We now let RocksDB put these blocks in

Re: Checkpoints very slow with high backpressure

2017-05-31 Thread SHI Xiaogang
Hi rhashmi We are also experiencing slow checkpoints when there exist back pressure. It seems there is no good method to handle back pressure now. We work around it by setting a larger number of checkpoint timeout. The default value is 10min. But checkpoints usually take more time to complete whe

Re: trying to externalize checkpoint to s3

2017-05-22 Thread SHI Xiaogang
Hi Sathi, According to the format specification of URI, "abc-checkpoint" is the host name in the given uri and the path is null. Therefore, FsStateBackend are complaining about the usage of the root directory. Maybe "s3:///abc-checkpoint" ("///" instead of "//") is the uri that you want to use. I

Re: Question about start with checkpoint.

2017-05-21 Thread SHI Xiaogang
Hi Yunfan, Jobs are supposed to correctly restart from both savepoints and checkpoints with different parallelisms if only operator states and keyed states are used. In the cases where there exist unpartitionable states (e.g., those are produced by the Checkpointed interface), the job will fail to

Re: Appropriate State to use to buffer events in ProcessFunction

2017-03-16 Thread SHI Xiaogang
Hi Yassine, If I understand correctly, you are needing sorted states which unfortunately are not supported in Flink now. We have some ideas to provide such sorted states to facilitate the development of user applications. But it is still under discussion due to the concerns on back compatibility.