[
https://issues.apache.org/jira/browse/BEAM-1283?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15957382#comment-15957382
]
Luke Cwik commented on BEAM-1283:
---------------------------------
I don't think the Fn API really is interested in whether an SDK exposes the
concept of finishBundle and whether it is allowed to perform output.
The issue is being able to implement something like PGBK without support for
finishBundle within an SDK.
> DoFn finishBundle should be required to specify the window
> ----------------------------------------------------------
>
> Key: BEAM-1283
> URL: https://issues.apache.org/jira/browse/BEAM-1283
> Project: Beam
> Issue Type: Bug
> Components: beam-model, sdk-java-core, sdk-py
> Reporter: Kenneth Knowles
> Labels: backward-incompatible
> Fix For: First stable release
>
>
> The spec is here in Javadoc:
> https://github.com/apache/beam/blob/master/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/DoFn.java#L128
> "If invoked from {{@StartBundle}} or {{@FinishBundle}}, this will attempt to
> use the {{WindowFn}} of the input {{PCollection}} to determine what windows
> the element should be in, throwing an exception if the {{WindowFn}} attempts
> to access any information about the input element. The output element will
> have a timestamp of negative infinity."
> This is a collection of caveats that make this method not always technically
> wrong, but quite a mess. Ideas that reasonable folks have suggested lately:
> - The {{WindowFn}} cannot actually be applied because {{WindowFn}} is
> allowed to see the element type. The spec just avoids this by limiting which
> {{WindowFn}} can be used.
> - There is no natural output timestamp, so it should always be provided. The
> spec avoids this by specifying an arbitrary and fairly useless timestamp.
> - If it is a merging {{WindowFn}} like sessions that has already been merged
> then you'll just have a bogus proto window regardless of explicit timestamp
> or not.
> The use cases for these methods are best addressed by state plus window
> expiry callback, so we should revisit this spec and probably just wipe it.
> There are some rare case where you might need to output from {{FinishBundle}}
> in a way that is not _actually_ sensitive to bundling (perhaps modulo some
> downstream notion of equivalence) in which case you had better know what
> window you are outputting to. Often it should be the global window.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
