One other nice enhancement around this would be if a transform could
indicate that it was executing a "slow" operation.
A good example is writing in BigQueryIO, it's very reasonable/normal for a
load job to run for more than 5 minutes, and the "stuck" message can be
confusing to users. The rewording to "operation ongoing" in the PR seems
like a good improvement here as well though.
On Thu, Jan 9, 2020 at 8:26 PM Pablo Estrada wrote:
> Hello Beam users and community,
>
> The Beam Python SDK, and Java workers have a utility where they will print
> a log message whenever there's an execution thread where no state
> transitions happen for over five minutes.
>
> These messages are common in two scenarios:
> 1. A deadlock happening in the worker (very uncommon, but possible)
> 2. An operation simply takes over 5 minutes (e.g. a slow RPC, waiting for
> an external event, etc).
>
> The old phrasing of these logs has often been a bit confusing, and
> led users to think that there was actual stuckness in the pipeline, when
> reality was more harmless: an operation was just slow.
>
> I am introducing a change[1] for the Apache Beam SDK to rephrase these
> logs, and make them less confusing.
>
> If you ever used these logs for your debugging, know that the string will
> change, but the logs will remain : ).
> If you didn't know about these, now you do, and hopefully they will be
> useful to you! : )
>
> Thanks!
> -P.
>
> [1] https://github.com/apache/beam/pull/10446/files
>
