Not sure how it works in Dataflow or Flink, but I'm working on an implementation for Spark using the (almost) only stateful operator it has - "mapWithState" - and the State needs to correspond to a key. Each micro-batch, the Sources recreate the readers and "look-up" the latest checkpoint.
On Mon, Sep 12, 2016 at 9:15 PM Raghu Angadi <[email protected]> wrote: > On Wed, Sep 7, 2016 at 7:13 AM, Amit Sela <[email protected]> wrote: > > > @Raghu, hashing <topic, partition> is exactly what I mean, but I'm asking > > if it should be abstracted in the Source.. Otherwise, runners will have > to > > *is instance of* on every Source, and write their own hash > implementation. > > Since splits contain the "splitted" Source, and it contains it's own > > CheckpointMark, and hashing would probably be tied to that > CheckpointMark, > > why not abstract it in the UnboundedSource ? > > > > I don't quite follow how a runner should be concerned about hashing used by > a Source for its own splits. Can you give a concrete example? To me it > looks like source and checkpoint objects are completely opaque to the > runners. > > > > On Wed, Sep 7, 2016 at 3:02 AM Raghu Angadi <[email protected]> > > wrote: > > > > > > If splits (UnboundedSources) had an identifier, this could be > avoided, > > > and checkpoints could be persisted accordingly. > > > > > > The order of the splits that a source returns is preserved. So during > an > > > update, you can expect 5th split gets invoked with the same checkpoint > > mark > > > that 5th split saved before upgrade. You can hash <topic, partition> to > > one > > > of the indices. > > > > > > KafkaIO.java doe not support change in partitions. > > > > > > On Tue, Sep 6, 2016 at 4:04 PM, Eugene Kirpichov < > > > [email protected]> wrote: > > > > > > > Oh! Okay, looks like this is a part of the code I was unfamiliar > with. > > > I'd > > > > like to know the answer too, in this case. > > > > +Daniel Mills <[email protected]> can you comment ? > > > > > > > > On Tue, Sep 6, 2016 at 3:32 PM Amit Sela <[email protected]> > wrote: > > > > > > > > > That is correct, as long as non of the Kafka topics "grow" another > > > > > partition (which it could). > > > > > In that case, some bundle will have to start reading from this > > > partition > > > > as > > > > > well, and since all other bundles already have a "previous > > checkpoint" > > > it > > > > > matters which checkpoint to relate to. I don't know how it's > > > implemented > > > > in > > > > > Dataflow, but in Spark I'm testing using mapWithState to store the > > > > > checkpoints per split. > > > > > It also seems that order does matter to the KafkIO: > > > > > > > > > > https://github.com/apache/incubator-beam/blob/master/ > > > > sdks/java/io/kafka/src/main/java/org/apache/beam/sdk/io/ > > > > kafka/KafkaIO.java#L636 > > > > > > > > > > On Wed, Sep 7, 2016 at 1:24 AM Eugene Kirpichov > > > > > <[email protected]> wrote: > > > > > > > > > > > Hi Amit, > > > > > > Could you explain more about why you're saying the order of > splits > > > > > matters? > > > > > > AFAIK the semantics of Read.Unbounded is "read from all of the > > splits > > > > in > > > > > > parallel, checkpointing each of them independently", so their > order > > > > > > shouldn't matter. > > > > > > > > > > > > On Tue, Sep 6, 2016 at 3:17 PM Amit Sela <[email protected]> > > > wrote: > > > > > > > > > > > > > UnboundedSources generate initial splits, which are basically > > > splits > > > > of > > > > > > > them selves - for example, if an UnboundedKafkaSource is set to > > > read > > > > > from > > > > > > > topic T1 which is made of 2 partitions P1 and P2, it will > > > (optimally) > > > > > > split > > > > > > > into two UnboundedKafkaSource, one per partition. > > > > > > > During the lifecycle of the "reader" bundles, CheckpointMarks > are > > > > used > > > > > to > > > > > > > checkpoint "last-read" and so readers may restart/resume. I'm > > > > assuming > > > > > > this > > > > > > > is how newly created partitions will automatically be added to > > > > readers. > > > > > > > > > > > > > > The problem is that it's all fine while there is only one topic > > > > (Kafka > > > > > > > topic-partition pairs are ordered), but if reading from more > then > > > one > > > > > > topic > > > > > > > this may break: > > > > > > > T1,P1 > > > > > > > T1,P2 > > > > > > > T1,P3 > > > > > > > T2,P1 > > > > > > > The order is not maintained and T2,P1 is 4th now. > > > > > > > > > > > > > > If splits (UnboundedSources) had an identifier, this could be > > > > avoided, > > > > > > and > > > > > > > checkpoints could be persisted accordingly. > > > > > > > For example, an UnboundedKafkaSource could use the hash code of > > > it's > > > > > > > assigned topic-partition pairs. > > > > > > > > > > > > > > I don't know how relevant this is to other Sources, but I guess > > it > > > is > > > > > as > > > > > > > long as they may grow their partitions dynamically (though I > > might > > > be > > > > > > > completely wrong...) and I don't see much of a downside. > > > > > > > > > > > > > > Thoughts ? > > > > > > > > > > > > > > This is something that troubles me now while working on > > > > Read.Unbounded, > > > > > > and > > > > > > > from a quick look I saw that the FlinkRunner expects "stable" > > > > splitting > > > > > > as > > > > > > > well.. How does Dataflow handle this ? > > > > > > > > > > > > > > Thanks, > > > > > > > Amit > > > > > > > > > > > > > > > > > > > > > > > > > > > >
