Steve Niemitz created BEAM-7745:
-----------------------------------
Summary: StreamingSideInputDoFnRunner/StreamingSideInputFetcher
have suboptimal state access pattern during normal operation
Key: BEAM-7745
URL: https://issues.apache.org/jira/browse/BEAM-7745
Project: Beam
Issue Type: Improvement
Components: runner-dataflow
Reporter: Steve Niemitz
I spent some time tracking down sources of uncached state fetches in my job,
and one large category was the interaction of StreamingSideInputDoFnRunner +
StreamingSideInputFetcher.
Basically, during standard operations, when the main input is NOT blocked by
the side input, the side input fetcher will perform an uncached state read for
every input element. Changing it to cache the blockedMap state gave me a
~30-40% increase in throughput in my job.
The interaction is a little complicated, and there's a couple optimizations
here I can see.
Primarily, the blockedMap is only persisted if it is non-empty. Because the
WindmillStateCache won't cache a null value, this means that the "nothing is
blocked" signal is never actually cached, and will issue a state read to
windmill for each input element. The solution here seems like it is to persist
an empty map rather than a null when there are no blocked elements.
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
