beam-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Kenneth Knowles (JIRA)" <>
Subject [jira] [Reopened] (BEAM-145) OutputTimeFn#assignOutputTime overrides WindowFn#getOutputTime in unfortunate ways
Date Tue, 19 Jul 2016 18:45:20 GMT


Kenneth Knowles reopened BEAM-145:

There's an unfortunate mutual default-invocation loop left over from the fix to this. It will
go away when OutputTime becomes an AST instead of a Fn, but keeping this open to track that
there remains an issue.

> OutputTimeFn#assignOutputTime overrides WindowFn#getOutputTime in unfortunate ways
> ----------------------------------------------------------------------------------
>                 Key: BEAM-145
>                 URL:
>             Project: Beam
>          Issue Type: Bug
>          Components: runner-core
>            Reporter: Kenneth Knowles
>            Assignee: Kenneth Knowles
>            Priority: Minor
>              Labels: windowing
>             Fix For: 0.1.0-incubating
> Today the {{OutputTimeFn}} includes {{#assignOutputTime}}, {{#combine}}, and {{#merge}}.
Together these express the grouping of timestamps, analogous to the grouping of values in
a GBK / Combine, in a canonical way.
> The default {{OutputTimeFn}} is provided by the {{WindowFn}}. In particular, {{SlidingWindows}}
provides an {{OutputTimeFn}} that shifts input timestamps later to avoid watermark stuckness
and then takes the minimum to compute the output timestamp.
> The SDK additionally provides instance for "min", "max" and "end of window" output timestamps.
> Unfortunately,  if one overrides the {{OutputTimeFn}} to one of these, the shifting done
by {{SlidingWindows}} is lost.
> This is actually only a minor problem for now, since "min" is the default, "end of window"
is unaffected, and "max" has only esoteric uses.The fix is easy:
> This is interrelated with another suggested change:  Since there are only three common
{{OutputTimeFn}} instances, and it is a high bandwidth API, it does not seem worthwhile to
leave it in userland. So it is proposed to reduce it to an enum, which would leave only the
{{WindowFn}} as a userland place for timestamp adjustments. (requiring special casing for
end-of-window, since it cannot be implemented without owning {{#assignOutputTime}})

This message was sent by Atlassian JIRA

View raw message