Google Cloud Dataflow won't override your setting. The dynamic sharding occurs if you don't explicitly set a numShard value.
On Wed, May 24, 2017 at 9:14 AM, Josh <jof...@gmail.com> wrote: > Hi Lukasz, > > Thanks for the example. That sounds like a nice solution - > I am running on Dataflow though, which dynamically sets numShards - so if > I set numShards to 1 on each of those AvroIO writers, I can't be sure that > Dataflow isn't going to override my setting right? I guess this should work > fine as long as I partition my stream into a large enough number of > partitions so that Dataflow won't override numShards. > > Josh > > > On Wed, May 24, 2017 at 4:10 PM, Lukasz Cwik <lc...@google.com> wrote: > >> Since your using a small number of shards, add a Partition transform >> which uses a deterministic hash of the key to choose one of 4 partitions. >> Write each partition with a single shard. >> >> (Fixed width diagram below) >> Pipeline -> AvroIO(numShards = 4) >> Becomes: >> Pipeline -> Partition --> AvroIO(numShards = 1) >> |-> AvroIO(numShards = 1) >> |-> AvroIO(numShards = 1) >> \-> AvroIO(numShards = 1) >> >> On Wed, May 24, 2017 at 1:05 AM, Josh <jof...@gmail.com> wrote: >> >>> Hi, >>> >>> I am using a FileBasedSink (AvroIO.write) on an unbounded stream >>> (withWindowedWrites, hourly windows, numShards=4). >>> >>> I would like to partition the stream by some key in the element, so that >>> all elements with the same key will get processed by the same shard writer, >>> and therefore written to the same file. Is there a way to do this? Note >>> that in my stream the number of keys is very large (most elements have a >>> unique key, while a few elements share a key). >>> >>> Thanks, >>> Josh >>> >> >> >