osdir.com


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: intentional back-pressure (or a poor man's side-input)


Maybe it could work with Flink’s 1.5 credit base flow control. But you would need a way to express state “block one input side of the CoProcessFunction”, pass this information up to the input gate and handle it probably similar to how `org.apache.flink.streaming.runtime.io.CachedBufferBlocker` blocks inputs in case of checkpoint barrier. You can not just block inside `processElement1` method.

However I haven’t thought it through and maybe there could be some issues regarding checkpointing (what should happen to checkpoint barriers if you are blocking one side of the input? Should this block checkpoint barrier as well? Should you cancel checkpoint?).

Piotrek

> On 2 May 2018, at 16:31, Derek VerLee <derekverlee@xxxxxxxxx> wrote:
> 
> 
> I was just thinking about about letting a coprocessfunction "block" or cause back pressure on one of it's streams?
> Has this been discussed as an option?
> Does anyone know a way to effectively accomplish this?
> 
> I think I could get a lot of mileage out of something like that without needing a full implementation of FLIP-17 (which I would eagerly await still). 
> 
> As mentioned on another thread, one could use a liststate to buffer the main input until the "side input" was sufficiently processed.  However the downside of this is that I have no way to control the size of those buffers, whereas with backpressure, the system will naturally take care of it.