[
https://issues.apache.org/jira/browse/BEAM-1314?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16048581#comment-16048581
]
Luke Cwik edited comment on BEAM-1314 at 6/14/17 1:35 AM:
----------------------------------------------------------
I believe this is more of an implementation detail of how the Java SDK works.
Nothing in the model says that a DoFn which uses user state can't be re-used
across different keys within the same bundle. I believe the flaws are within
the usage of a KeyedWorkItem and how the current key propagates through all the
context objects within Java SDK runners core.
was (Author: lcwik):
I believe this is more of an implementation detail of how the Java SDK works.
Nothing in the model says that a DoFn which uses user state can't be re-used
across different keys within the same bundle. I believe the flaws are within
the usage of a KeyedWorkItem and how the current key propagates through all the
context objects within the Java SDK runners core.
> DoFn per-key lifecycle (Setup/Teardown)
> ---------------------------------------
>
> Key: BEAM-1314
> URL: https://issues.apache.org/jira/browse/BEAM-1314
> Project: Beam
> Issue Type: Wish
> Components: beam-model
> Reporter: Eugene Kirpichov
>
> DoFn's that use state and timers are implicitly per-key. Setup/Teardown
> methods are usually used to establish expensive resources - long-standing
> connections and such.
> For per-key DoFn's, we'd often want to use these per-key, so it'd be good to
> have ability in the model to ask that there be 1 instance of the DoFn per
> key, reused between e.g. different timer or trigger firings for this key, but
> not used for other keys.
> E.g. this would be particularly useful for Splittable DoFn - for its ability
> to reuse expensive resources between checkpoints.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
