http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/data-enrichment-with-SQL-use-case-tp19520p19770.html
Hi Michael,
Windowing works when you’re joining timestamped metadata and non-metadata.
The common case I’m referring to is where there’s some function state (e.g. rules to process data, machine learning models, or in my case clusters), where you want to process the non-metadata with the "current state”.
In that case, blindly applying whatever metadata has been collected to incoming non-metadata often doesn’t work well. That’s why Fabian was suggesting various approaches (below) to work around the problem. The general solution (his option #2, with buffering) will work, but can lead to OOME and feels like it breaks the basic Flink back-pressure mechanism, due to in-operator buffering.
If it was possible to essentially allow Flink to block (or not pull, for sources) from the non-metadata stream when appropriate, then no buffering would be needed. Then it would be straightforward to do things like…
- drain all metadata from a Kafka topic before applying that to the other stream.
- defer processing data from the other stream if there was newer metadata.
As an aside, what I’m seeing with Flink 1.5 and using a connected keyed & broadcast stream is that the CoFlatMapFunction seems to be giving priority to data going to the flatMap1() method, though this could be an odd side effect of how iterations impact the two streams.
— Ken
I agree in the general case you need to operate on the stream data based on the metadata you have. The side input feature coming some day may help you, in that it would give you a means to receive inputs out of band. But, given changing metadata and changing stream data I am not sure this is any different from dual stream data inputs. Either you use windowing to do small batches of data to allow coordination of stream and metadata, or you use the metadata you have collected to date on receipt of the stream data. Given flink will do record by record processing you have the option of controlling the timing as needed for your use case.
Michael
Hi Michael,
I agree there are cases where it’s possible to implement a solution via buffering.
But this case of using broadcast state to update a function operating on streaming data seems common enough that it would be useful for Flink to provide some help.
Additionally, even with buffering there are currently challenges...
1. For the case I’m dealing with (iterative KMeans clustering) you don’t have a time when "metadata is aggregated", as it’s constantly evolving.
2. It’s sometimes not possible to know when you’ve received all of the metadata (e.g. if you’re reading from a Kafka topic).
3. Buffering the non-metadata can create an unbounded memory issue.
Regards,
— Ken
Using a flat map function, you can always buffer the non-meta data stream in the operator state until the metadata is aggregated, and then process any collected data. It would require a RichFlatMap to hold data.
Michael
Hi Fabian,
Hi Alex,
An operator that has to join two input streams obviously requires two inputs. In case of an enrichment join, the operator should first read the meta-data stream and build up a data structure as state against which the other input is joined. If the meta data is (infrequently) updated, these updates should be integrated into the state.
The problem is that it is currently not possible to implement such an operator with Flink because operators cannot decide from which input to read, i.e., they have to process whatever data is given to them.
Hence, it is not possible to build up a data structure from the meta data stream before consuming the other stream.
This seems like a common situation, and one where it might be relatively easy for Flink to help resolve.
Specifically, for a connected stream feeding a Co(Flat)MapFunction, it seems like we could let Flink know how to pick elements from the two network buffers - e.g. random, round robin, or by timestamp.
I don’t know how this works with chained operators, but it does seem a bit odd to have operators create buffers of elements when (network) buffers often already exist.
If there’s no network buffers in play (e.g. there’s a direct chain of operators from a source) then it could be something that’s not supported, though with the future source-pull architecture that would also be easy to resolve.
Anyway, I could take a whack at this if it seems reasonable.
— Ken
There are a few workarounds that work in special cases.
1) The meta data is rather small and never updated. You put the meta data as a file into a (distributed) file system an read it from each function instance when it is initialized, i.e., in open(), and put into a hash map. Each function instance will hold the complete meta data in memory (on the heap). Since the meta data is broadcasted, the other stream does not need to be partitioned to join against the meta data in the hash map. You can implement this function as a FlatMapFunction or ProcessFunction.
2) The meta data is too large and/or is updated. In this case, you need a function with two inputs. Both inputs are keyed (keyBy()) on a join attribute. Since you cannot hold back the non-meta data stream, you need to buffer it in (keyed) state until you've read the meta data stream up to a point when you can start processing the other stream. If the meta data is updated at some point, you can just add the new data to the state. The benefits of this approach is that the state is shared across all operators and can be updated. However, you might need to initially buffer quite a bit of data in state if the non-meta data stream has a high volume.
Hope that one of these approaches works for your use case.
Best, Fabian
--------------------------------------------
+1 530-210-6378
--------------------------------------------
+1 530-210-6378
--------------------------------------------
+1 530-210-6378