Hi Shaosu, Do you have an estimate on the total size of state you are keeping for the windows? How many messages/sec, how large a window, message size, etc would be good details to include.
Also, which state backend are you using? Have you considered using the RocksDB state backend. This backend will spill Flink state to disk if it's larger than available RAM. You'll also probably want to use "fully async" mode for the RocksDBStateBackend. -Jamie On Fri, Sep 2, 2016 at 1:45 PM, Shaosu Liu <s...@uber.com> wrote: > Hi, > > I have had issues when I processed large amount of data (large windows > where I could not do incremental updates), flink slowed down significantly. > It did help when I increased the amount of memory and used off heap > allocation. But it only delayed the onset of the probelm without solving > it. > > Could some one give me some hints on how Flink manage window buffer and > how streaming manages its memory. I see this page on batch api memory > management and wonder what is the equivalent for streaming? > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=53741525 > > -- > Cheers, > Shaosu > -- Jamie Grier data Artisans, Director of Applications Engineering @jamiegrier <https://twitter.com/jamiegrier> ja...@data-artisans.com
