I see where you are coming from. It is truly a marginal feature of Beam at
the moment, but really *really* useful in debugging, when a runner takes
advantage of it. More inline - FWIW it may seem like I'm defending the
tests, but I have been in the same situation and felt the same way. I'm
just trying to ground things in specifics.

On Thu, Oct 26, 2017 at 2:24 PM, Eugene Kirpichov <
[email protected]> wrote:

> Hello,
>
> DisplayData is currently in an odd position:
> - Dataflow is the only runner that supports it at all [1]
> - Dataflow supports it only for primitive transforms, but discards it for
> composite transforms [2](which is quite problematic for very complex but
> commonly used composite transforms, such as TextIO.write() and
> BigQueryIO.write())
>

As of today, the Java and Python Dataflow runners actually transmit the
entire original pipeline, which has slots for display data on composites.
So the future in which this is actually used - at least for Dataflow -
might be very near.


- There are a lot of tests testing display data of composite transforms -
> these tests are currently testing something that is a no-op in production;
> and these tests often are a maintenance burden when refactoring a
> transform.


Which composites' tests in particular? Are they legitimate tests?


What should we do about this? I see two options:
> - Publish guidance for library authors that Beam is not currently ready for
> supporting HasDisplayData on composite transforms, and you shouldn't bother
> implementing or testing it (though you can implement it on primitive
> transforms of course - which is basically just Read.from(Source) and ParDo)
> - Say that people should be implementing and testing this feature
> nevertheless, and promise that Beam runners are going to make it "worth it"
> (not be a no-op) in the foreseeable future.
>

Option 3: Don't say anyone _should_ do any particular thing, but promise
that [some] Beam runners are [probably] going to make it "worth it" pretty
soon. I expect third-party transform authors are already not using this,
are they?

Kenn



> [1]
> https://lists.apache.org/thread.html/53b33f17d981054ed198af03511969
> 07bf147b6acf16160dbf395b62@%3Cdev.beam.apache.org%3E
> [2] https://issues.apache.org/jira/browse/BEAM-366
>

Reply via email to