Great catch, Reza. Dataflow uses a customized C++ based Pubsub source that
calculates the watermark differently. So the Java-based PubsubIO may have
different behavior with respect the watermark (expected) or may simply have
bugs (not desirable).

How are you feeding test data into Pubsub? Are you using TestPubsub [1] or
your own wrapper?

Kenn

[1]
https://github.com/apache/beam/blob/master/sdks/java/io/google-cloud-platform/src/main/java/org/apache/beam/sdk/io/gcp/pubsub/TestPubsub.java

On Sat, Mar 7, 2020 at 1:46 AM Reza Rokni <[email protected]> wrote:

> Hi,
>
> There are some known issues with the DirectRunner PubSubIO, which are not
> present with Dataflow Runner. One of them is around watermarks:
>
> https://issues.apache.org/jira/browse/BEAM-7322
>
> Not sure if this is part of the issue here, but worth exploring..
>
> When testing are you sending a small volume of information and then
> stopping or are you sending continuous output?
>
> Cheers
>
> Reza
>
> On Sat, Mar 7, 2020 at 1:29 PM Kenneth Knowles <[email protected]> wrote:
>
>> Can you reproduce it if you replace your Pubsub source with a TestStream
>> and verify with PAssert [1]? This would enable you to easily build a unit
>> test. You could even open a pull request adding that to the test suite for
>> GroupIntoBatches [2]. That would be an excellent contribution to Beam.
>>
>> Kenn
>>
>> [1] https://beam.apache.org/blog/2016/10/20/test-stream.html
>> [2]
>> https://github.com/apache/beam/blob/master/sdks/java/core/src/test/java/org/apache/beam/sdk/transforms/GroupIntoBatchesTest.java
>>
>> On Mon, Mar 2, 2020 at 9:25 AM Vasu Gupta <[email protected]>
>> wrote:
>>
>>> Input : a-1, Timestamp : 1582994620366
>>> Input : c-2, Timestamp : 1582994620367
>>> Input : e-3, Timestamp : 1582994620367
>>> Input : d-4, Timestamp : 1582994620367
>>> Input : e-5, Timestamp : 1582994620367
>>> Input : b-6, Timestamp : 1582994620368
>>> Input : a-7, Timestamp : 1582994620368
>>>
>>> Output : Timestamp : 1582994620367, Key : e-3,5
>>> Output : Timestamp : 1582994620368, Key : a-1,7
>>>
>>> As you can see c-2 and d-4 are missing and I never received these
>>> packets.
>>>
>>> On 2020/02/28 18:15:03, Kenneth Knowles <[email protected]> wrote:
>>> > What are the timestamps on the elements?
>>> >
>>> > On Fri, Feb 28, 2020 at 8:36 AM Vasu Gupta <[email protected]>
>>> wrote:
>>> >
>>> > > Edit: Issue is on Direct Runner(Not Direction Runner - mistyped)
>>> > > Issue Details:
>>> > > Input data: 7 key-value Packets like: a-1, a-4, b-3, c-5, d-1, e-4,
>>> e-5
>>> > > Batch Size: 5
>>> > > Expected output: a-1,4, b-3, c-5, d-1, e-4,5
>>> > > Getting Packets with irregular size like a-1, b-5, e-4,5 OR a-1,4,
>>> c-5 etc
>>> > > But i always got correct number of packets with BATCH_SIZE = 1
>>> > >
>>> > > On 2020/02/27 20:40:16, Kenneth Knowles <[email protected]> wrote:
>>> > > > Can you share some more details? What is the expected output and
>>> what
>>> > > > output are you seeing?
>>> > > >
>>> > > > On Thu, Feb 27, 2020 at 9:39 AM Vasu Gupta <
>>> [email protected]>
>>> > > wrote:
>>> > > >
>>> > > > > Hey folks, I am using Apache beam Framework in Java with
>>> Direction
>>> > > Runner
>>> > > > > for local testing purposes. When using GroupIntoBatches with
>>> batch
>>> > > size 1
>>> > > > > it works perfectly fine i.e. the output of the transform is
>>> consistent
>>> > > and
>>> > > > > as expected. But when using with batch size > 1 the output
>>> Pcollection
>>> > > has
>>> > > > > less data than it should be.
>>> > > > >
>>> > > > > Pipeline flow:
>>> > > > > 1. A Transform for reading from pubsub
>>> > > > > 2. Transform for making a KV out of the data
>>> > > > > 3. A Fixed Window transform of 1 second
>>> > > > > 4. Applying GroupIntoBatches transform
>>> > > > > 5. And last, Logging the resulting Iterables.
>>> > > > >
>>> > > > > Weird thing is that it batch_size > 1 works great when running on
>>> > > > > DataflowRunner but not with DirectRunner. I think the issue
>>> might be
>>> > > with
>>> > > > > Timer Expiry since GroupIntoBatches uses BagState internally.
>>> > > > >
>>> > > > > Any help will be much appreciated.
>>> > > > >
>>> > > >
>>> > >
>>> >
>>>
>>

Reply via email to