`Add support for triggers.
Triggers are special runner side behavior indicating how to handle data WRT the watermark and window. Commonly configuring the processing for “late data” and similar.
These are not currently implemented for user use in the Go SDK. Reshuffle configures triggers, but it’s not accessible. A correct trigger implementation can at least re-implement Reshuffle in a user pipeline, rather than handled specially within the framework.
- Requires extending the window package to be able to configure the various triggers.
- Specifically being able to compose triggers as also permitted by the proto.
- Requires updating the graphx package translate.go to marshal (and unmarshal?) the triggers to and from Beam PipelineProto Windowing strategies.
- Requires supporting triggers with the beam.WindowInto transform for user pipeline use as well as complete documentation on its use from the user side.
- Panes need to be decoded, otherwise triggering will cause runtime errors: https://lists.apache.org/thread.html/r94c42d2d116f6464cd6b689543e5e578edf8310bf7c6e48a0958a56c%40%3Cdev.beam.apache.org%3E
- Handle pane propagation and observation in the exec package, and in user dofns.
- Panes indicate whether data was on time or not, and similar facets which may be relevant for processing.
- Might simply extend the existing window interface.
Similar to windowing, many of the same places as https://issues.apache.org/jira/browse/BEAM-11100 need to be modified.
At simplest though, it's mostly a runner side construction, with less concern on the exec side, and generally much simpler.
Appropriate integration tests against portable runners must be implemented:
And optionally add support for the configurable triggers to the the Go Direct Runner. However, the results must be compared and validated against a semantically correct runner like the python portable runner first. At minimum, the Go Direct Runner should be made aware of triggers and produce a coherent error whenever there's a trigger it can't deal with.