I would not recommend using RequiresTimeSortedInput in this way. I also would not use ingestion time, as in a distributed environment, time skew between workers might mess up the order.
I will ping the discussion on the sorted state API and add you. My hope is that this can be implemented efficiently soon, though efficient implementation will definitely be runner dependent. If you are using Flink, we'll have to figure how to implement this efficiently using Flink. Dataflow is planning on providing native support for sorted state. I don't know if Flink has native support for this, so it might have to be emulated using its existing state primitives. In the meanwhile, I would suggest using bagstate along with timers. The timer can periodically pull sorted messages out of the bag (you can use watermark timers here) and write back the messages that still have gaps in them. Reuven On Wed, Jun 10, 2020 at 6:34 PM Catlyn Kong <[email protected]> wrote: > Thank y’all for the input! > > > About the RequiresTimeSortedInput, we were thinking of the following 2 > potential approaches: > > 1. > > Assign kafka offset as the timestamp while doing a GroupByKey on > partition_id in a GlobalWindow > 2. > > Rely on the fact that Flink consumes from kafka partitions in offset > order and assign ingestion time as the timestamp. (We're using our own > non-KafkaIO based Kafka consumer extended from FlinkKafkaConsumer011 and > thus have direct control over timestamp and watermark assignment) > > We find it non-trivial to reason about watermark assignment especially > when taking into consideration that: > > 1. > > there might be restarts at any given time and > 2. > > advancing watermark in one kafka partition might result in: > 1. > > dropping elements from other kafka partitions (if we’re not > following native flink approach where we take the lowest watermark when > merging streams) or > 2. > > delay output from other kafka partitions since they’ll be buffered. > > Is there any recommendation on how this should be handled? > > In the direction of using a StatefulDoFn to buffer and reorder, we’re > concerned about performance since we need to serialize and deserialize the > entire BagState (with all the messages) everytime we process a message. And > potentially insert this StatefulDoFn in multiple places in the pipeline. Is > there any benchmark result of a pipeline that does something similar for us > to reference? > > The proposal for a sorted state API sounds promising, is there a > ticket/doc that we can follow? > > > On Wed, Jun 10, 2020 at 1:28 PM Reuven Lax <[email protected]> wrote: > >> I don't know how well RequiresTimeSortedInput will work for any late data. >> >> I think you will want to include the Kafka offset in your records (unless >> the records have their own sequence number) and then use state to buffer >> and sort. There is a proposal (and work in progress) for a sorted state >> API, which will make this easier and more efficient. >> >> Reuven >> >> On Wed, Jun 10, 2020 at 1:25 PM Luke Cwik <[email protected]> wrote: >> >>> For runners that support @RequiresTimeSortedInput, all your input will >>> come time sorted (as long as your element's timestamp tracks the order that >>> you want). >>> For runners that don't support this, you need to build a StatefulDoFn >>> that buffers out of order events and reorders them to the order that you >>> need. >>> >>> @Pablo Estrada <[email protected]> Any other suggestions for >>> supporting CDC type pipelines? >>> >>> On Tue, Jun 9, 2020 at 6:59 PM Catlyn Kong <[email protected]> wrote: >>> >>>> Thanks a lot for the response! >>>> >>>> We have several business use cases that rely strongly on ordering by >>>> Kafka offset: >>>> 1) streaming unwindowed inner join: say we want to join users with >>>> reviews on user_id. Here are the schemas for two streams: >>>> user: >>>> >>>> - user_id >>>> - name >>>> - timestamp >>>> >>>> reviews: >>>> >>>> - review_id >>>> - user_id >>>> - timestamp >>>> >>>> Here are the messages in each stream ordered by kafka offset: >>>> user: >>>> (1, name_a, 60), (2, name_b, 120), (1, name_c, 240) >>>> reviews: >>>> (ABC, 1, 90), (DEF, 2, 360) >>>> I would expect to receive following output messages: >>>> (1, name_a, ABC) at timestamp 90 >>>> (1, name_c, ABC) at timestamp 240 >>>> (2, name_b, DEF) at timestamp 360 >>>> This can be done in native Flink since Flink kafka consumer reads from >>>> each partition sequentially. But without an ordering guarantee, we can end >>>> up with arbitrary results. So how would we implement this in Beam? >>>> 2) unwindowed aggregation: aggregate all the employees for every >>>> organization. Say we have a new employee stream with the following schema: >>>> new_employee: >>>> >>>> - organization_id >>>> - employee_name >>>> >>>> And here are messaged ordered by kafka offset: >>>> (1, name_a), (2, name_b), (2, name_c), (1, name_d) >>>> I would expect the output to be: >>>> (1, [name_a]), (2, [name_b]), (2, [name_b, name_c]), (1, [name_a, >>>> name_d]) >>>> Again without an ordering guarantee, the result is non deterministic. >>>> >>>> Change data capture (CDC) streams are a very common use case for our >>>> data pipeline. As in the examples above we rely on Kafka offsets to make >>>> sure we process data mutations in the proper order. While in some cases we >>>> have Flink native solutions to these problems (Flink provides ordering >>>> guarantees within the chosen key), we are now building some new Beam >>>> applications that would require ordering guarantees. What is the >>>> recommended approach in Beam for such use cases? If this isn’t currently >>>> supported, do we have any near plan to add native ordering support in Beam? >>>> >>>> >>>> On 2020/06/09 20:37:22, Luke Cwik <[email protected]> wrote: >>>> > This will likely break due to:> >>>> > * workers can have more then one thread and hence process the source >>>> in> >>>> > parallel> >>>> > * splitting a source allows for the source to be broken up into >>>> multiple> >>>> > restrictions and hence the runner can process those restrictions in >>>> any> >>>> > order they want. (lets say your kafka partition has unconsumed >>>> commit> >>>> > offset range [20, 100), this could be split into [20, 60), [60, 100) >>>> and> >>>> > the [60, 100) offset range could be processed first)> >>>> > >>>> > You're right that you need to sort the output however you want within >>>> your> >>>> > DoFn before you make external calls to Kafka (this prevents you from >>>> using> >>>> > the KafkaIO sink implementation as a transform). There is an >>>> annotation> >>>> > @RequiresTimeSortedInput which is a special case for this sorting if >>>> you> >>>> > want it to be sorted by the elements timestamp but still you'll need >>>> to> >>>> > write to Kafka directly yourself from your DoFn.> >>>> > >>>> > On Mon, Jun 8, 2020 at 4:24 PM Hadi Zhang <[email protected]> wrote:> >>>> > >>>> > > We are using the Beam 2.20 Python SDK on a Flink 1.9 runner. Our> >>>> > > messages originate from a custom source that consumes messages from >>>> a> >>>> > > Kafka topic and emits them in the order of their Kafka offsets to >>>> a> >>>> > > DoFn. After this DoFn processes the messages, they are emitted to >>>> a> >>>> > > custom sink that sends messages to a Kafka topic.> >>>> > >> >>>> > > We want to process those messages in the order in which we receive> >>>> > > them from Kafka and then emit them to the Kafka sink in the same> >>>> > > order, but based on our understanding Beam does not provide an> >>>> > > in-order transport. However, in practice we noticed that with a >>>> Python> >>>> > > SDK worker on Flink and a parallelism setting of 1 and one >>>> sdk_worker> >>>> > > instance, messages seem to be both processed and emitted in order. >>>> Is> >>>> > > that implementation-specific in-order behavior something that we >>>> can> >>>> > > rely on, or is it very likely that this will break at some future> >>>> > > point?> >>>> > >> >>>> > > In case it's not recommended to depend on that behavior what is >>>> the> >>>> > > best approach for in-order processing?> >>>> > >> >>>> > > >>>> https://stackoverflow.com/questions/45888719/processing-total-ordering-of-events-by-key-using-apache-beam> >>>> >>>> > > recommends to order events in a heap, but according to our> >>>> > > understanding this approach will only work when directly writing to >>>> an> >>>> > > external system.> >>>> > >> >>>> > >>>> >>>
