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

Re: Multimap PCollectionViews' values udpated rather than appended

But I think what I'm experiencing is quite different. Basically the side input is updated, but only one element is found on the Iterable that is the value of any key of the multimap.

I mean, no concatenation seems to be happening. On the linked thread, Kenn suggests that every firing will add the new value to the set of values for the emitted key, but what I'm experiencing is that the new value is there, but just itself (i.e: is the only element in the set).

@Robert, I'm using Repeatedly.forever(AfterProcessingTime.pastFirstElementInPane())

On Wed, May 30, 2018 at 7:46 PM Lukasz Cwik <lcwik@xxxxxxxxxx> wrote:
An alternative to the thread that Kenn linked (adding support for retractions) is to add explicit support for combiners into side inputs. The system currently works by using a hardcoded concatenating combiner, so maps, lists, iterables, singletons, multimaps all work by concatenating the set of values emitted and then turning it into a view which is why it is an error for a singleton and map view if the trigger fires multiple times.

On Wed, May 30, 2018 at 10:01 AM Kenneth Knowles <klk@xxxxxxxxxx> wrote:
Yes, this is a known issue. Here's a prior discussion:

It is actually long-standing and the solution is known but hard.

On Wed, May 30, 2018 at 9:48 AM Carlos Alonso <carlos@xxxxxxxxxxxxx> wrote:
Hi everyone!!

Working with multimap based side inputs on the global window I'm experiencing something unexpected (at least to me) that I'd like to share with you to clarify.

The way I understand multimaps is that when one emits two values for the same key for the same window (obvious thing here as I'm working on the Global one), the newly emitted values are appended to the Iterable collection that is the value for that particular key on the map.

Testing it in this job (it is using scio, but side inputs are implemented with PCollectionViews):

The steps to reproduce are:
1. Create one table on the target BQ
2. Run the job
3. Patch the table on BQ (add one field), this should generate a new TableSchema for the corresponding TableReference
4. An updated value of the fields number appear on the logs, but there is only one element within the iterable, as if it had been updated instead of appended!!

Is that the expected behaviour? Is a bug? Am I missing something?