Hello, As for use cases, in my old job at Ericsson we were building a streaming system that was processing data from telephone networks, and it was using key-value stores a LOT. For example, keeping track of various state info of the users (which cell are they currently connected to, what bearers do they have, ...); mapping from IDs of users in one subsystem of the telephone network to the IDs of the same users in an other subsystem; mapping from IDs of phone calls to lists of IDs of participating users; etc. So I imagine they would like this a lot. (At least, if they were considering moving to Flink :))
Best, Gabor 2015-09-08 13:35 GMT+02:00 Gyula Fóra <gyf...@apache.org>: > Hey All, > > The last couple of days I have been playing around with the idea of > building a streaming key-value store abstraction using stateful streaming > operators that can be used within Flink Streaming programs seamlessly. > > Operations executed on this KV store would be fault tolerant as it > integrates with the checkpointing mechanism, and if we add timestamps to > each put/get/... operation we can use the watermarks to create fully > deterministic results. This functionality is very useful for many > applications, and is very hard to implement properly with some dedicates kv > store. > > The KVStore abstraction could look as follows: > > KVStore<K,V> store = new KVStore<>; > > Operations: > > store.put(DataStream<Tuple2<K,V>>) > store.get(DataStream<K>) -> DataStream<KV<K,V>> > store.remove(DataStream<K>) -> DataStream<KV<K,V>> > store.multiGet(DataStream<K[]>) -> DataStream<KV<K,V>[]> > store.getWithKeySelector(DataStream<X>, KeySelector<X,K>) -> > DataStream<KV<X,V>[]> > > For the resulting streams I used a special KV abstraction which let's us > return null values. > > The implementation uses a simple streaming operator for executing most of > the operations (for multi get there is an additional merge operator) with > either local or partitioned states for storing the kev-value pairs (my > current prototype uses local states). And it can either execute operations > eagerly (which would not provide deterministic results), or buffer up > operations and execute them in order upon watermarks. > > As for use cases you can probably come up with many I will save that for > now :D > > I have a prototype implementation here that can execute the operations > described above (does not handle watermarks and time yet): > > https://github.com/gyfora/flink/tree/KVStore > And also an example job: > > https://github.com/gyfora/flink/blob/KVStore/flink-staging/flink-streaming/flink-streaming-core/src/test/java/org/apache/flink/streaming/api/KVStreamExample.java > > What do you think? > If you like it I will work on writing tests and it still needs a lot of > tweaking and refactoring. This might be something we want to include with > the standard streaming libraries at one point. > > Cheers, > Gyula
