[ 
https://issues.apache.org/jira/browse/BEAM-146?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Daniel Halperin updated BEAM-146:
---------------------------------
    Fix Version/s: 0.2.0-incubating

> WindowFn.AssignContext leaks implementation details about compressed 
> WindowedValue representation
> -------------------------------------------------------------------------------------------------
>
>                 Key: BEAM-146
>                 URL: https://issues.apache.org/jira/browse/BEAM-146
>             Project: Beam
>          Issue Type: Bug
>          Components: beam-model
>            Reporter: Kenneth Knowles
>            Assignee: Kenneth Knowles
>            Priority: Minor
>             Fix For: 0.2.0-incubating
>
>
> Today, {{WindowFn.AssignContext}} provides simultaneous access to all of the 
> windows that a value has been placed in.
> Providing access to the current window for a value is convenient for, e.g. 
> converting day windows to hour windows for each hour of the assign day. But 
> providing access to all the assigned windows allows spooky action across 
> windows, and is generally not intended to be observable - elements are 
> semantically considered to be "duplicated" into each of the assigned windows.
> This ticket proposes that the {{AssignContext}} should provide only a single 
> window, and that windows should be "exploded" prior to window re-assignment 
> so that elements are only observed within one window at a time. This can be 
> accomplished trivially today via surgical insertion of 
> {{RequiresWindowAccess}} but the {{AssignContext}} should have its API 
> adjusted to be explicit about it, too.
> This will affect only pipelines for which _all_ of the following hold:
>  - assigns to sliding windows (or custom {{WindowFn}} that places each 
> element in multiple windows)
>  - re-assigns to different windows without a {{GroupByKey}} between.
>  - the new window assignment actually does depend on the full set of windows 
> assigned
> I hypothesize the number of such pipelines is zero.
> I expect to address this during the Beam Runner API design.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to