[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [portablility] metrics interrogations

I think there is a confusion of terminology here. Let me attempt to clarify as a (mostly) outsider.

I think that Etienne is correct, in that the SDK harness only reports the difference associated with a bundle. So, if we have a metric that measures execution time, the SDK harness reports the time spent executing *that bundle*.

It does not need to report the total execution time which would require knowing the total execution time without that bundle. This would in fact be nearly impossible -- since multiple bundles may be executed simultaneously there is no "total execution time without this bundle" available within the system.

Robert is also correct, in that the SDK harness may send a partial diff based on the current state, when asked to do so. This is important -- if a bundle were to take 1 hour to complete, do we want the execution time to be 5 hours (for the entire hour) and then jump to 6 hours, or do we want it to be able to increase somewhat smoothly during the hour? Supporting a smooth increase requires being able to send a partial result back -- "if this bundle were to complete now, here is the diff that would be commited".

Of course, if a runner uses these partial updates it may need logic to deal with the case of a partial update from a bundle that has failed, but that I think that is outside the scope of the original question. Needless to say, it requires something like a committed value and "tentative" value reflecting bundles that are in progress.W

Does that match what both sides have in mind?

On Wed, Sep 12, 2018 at 8:05 AM Robert Bradshaw <robertwb@xxxxxxxxxx> wrote:
On Wed, Sep 12, 2018 at 4:20 PM Etienne Chauchot <echauchot@xxxxxxxxxx> wrote:
Thanks Robert for the details,

I did not know that the runner harness periodically asked the SDK harness for updates. I thought it was only communicating at the beginning of the bundle and at the end. Something like the simplified sequence diagram bellow

but if the metrics are not a regular diff but more the not-yet-committed dirty value, that means that the runner sends the metrics value to the sdk before the bundle is started processing.
So the sequence diagram becomes something more like:


The runner never sends any metrics to the SDK, it just listens. It may aggregate them or send them upstream. An SDK can choose to publish all, none, or some strategic subset of metrics in its progress responses, as "latest values." On work output, it publishes everything. (Regarding your diagram, there may not even be any "metric cells" in the Java Runner Harness.)
Le mardi 11 septembre 2018 à 17:53 +0200, Robert Bradshaw a écrit :
On Mon, Sep 10, 2018 at 11:07 AM Etienne Chauchot <echauchot@xxxxxxxxxx> wrote:

Hi all,

@Luke, @Alex I have a general question related to metrics in the Fn API: as the communication between runner harness and SDK harness is done on a bundle basis. When the runner harness sends data to the sdk harness to execute a transform that contains metrics, does it:

  1. send metrics values (for the ones defined in the transform) alongside with data and receive an updated value of the metrics from the sdk harness when the bundle is finished processing?
  2. or does it send only the data and the sdk harness responds with a diff value of the metrics so that the runner can update them in its side?
My bet is option 2. But can you confirm?

The runner harness periodically asks for the status of a bundle to which the runner harness may respond with a current snapshot of metrics. These metrics are deltas in the sense that only "dirty" metrics need to be reported (i.e. unreported metrics can be assumed to have their previous values) but are *not* deltas with respect to values, i.e. the full value is reported each time. As an example, suppose one were counting red and blue marbles. The first update may be something like

{ red: 5, blue: 7}

and if two more blue ones were found, a valid update would be

{ blue: 9 }

On bundle completion, the full set of metrics is reported as part of the same message that declares the bundle complete. 

On Tue, Sep 11, 2018 at 11:43 AM Etienne Chauchot <echauchot@xxxxxxxxxx> wrote:
Le lundi 10 septembre 2018 à 09:42 -0700, Lukasz Cwik a écrit :
Alex is out on vacation for the next 3 weeks.

Alex had proposed the types of metrics[1] but not the exact protocol as to what the SDK and runner do. I could envision Alex proposing that the SDK harness only sends diffs or dirty metrics in intermediate updates and all metrics values in the final update.
Robert is referring to an integration that happened to an older set of messages[2] that preceeded Alex's proposal and that integration with Dataflow which is still incomplete works as you described in #2.

Thanks Luke and Robert for the confirmation.

Robin had recently been considering adding an accessor to DoFns that would allow you to get access to the job information from within the pipeline (current state, poll for metrics, invoke actions like cancel / drain, ...). He wanted it so he could poll for attempted metrics to be able to test @RequiresStableInput.
Yes, I remember, I voted +1 to his proposal.

Integrating the MetricsPusher or something like that on the SDK side to be able to poll metrics over the job information accessor could be useful.

Well, in the design discussion, we decided to host Metrics Pusher as close as possible of the actual engine (inside the runner code chosen over the sdk code) to allow the runner to send system metrics in the future.

+1. The runner harness can then do whatever it wants (e.g. reporting back to its master, or pushing to another service, or simply dropping them), but the SDKs only have to follow the FnAPI contract. 

Besides, in his PR Alex talks about deprecated metrics. As he is off, can you tell me a little more about them ? What metrics will be deprecated when the portability framework is 100% operational on all the runners?