flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stefan Richter <s.rich...@data-artisans.com>
Subject Re: large sliding window perf question
Date Tue, 23 May 2017 19:35:18 GMT

Which state backend and Flink version are you using? There was a problem with large merging
states on RocksDB, caused by some inefficiencies in the merge operator of RocksDB. We provide
a custom patch for this with all newer versions of Flink.


> Am 23.05.2017 um 21:24 schrieb Chen Qin <qinnchen@gmail.com>:
> Hi there,
> I have seen some weird perf issue while running event time based job with large sliding
window (24 hours offset every 10s) 
> pipeline looks simple, 
> tail kafka topic and assign timestamp and watermark, forward to large sliding window
(30days) and fire every 10 seconds and print out.
> what I have seen first hand was checkpointing stuck, took longer than timeout despite
traffic volume is low ~300 TPS. Looking deeper, it seems back pressure kick in and window
operator consumes message really slowly and throttle sources.
> I also tried to limit window time to mins and all issues are gone.
> Any suggestion on this. My work around is I implemented processFunction and keep big
value state, periodically evaluate and emit downstream (emulate what sliding window does)
> Thanks,
> Chen

View raw message