Hi guys, I’m using flink on production in Mapp. We recently swapped from storm. Before I have put this live I was doing performance tests and I found something that “feels” a bit off. I have a simple streaming job reading from kafka, doing window for 3 seconds and then storing into hbase.
Initially we had this second step written with a fold function, since I thought performance and resource wise it’s a better idea. But I couldn’t reach more than 120k writes per second to HBase and I thought hbase sink is a bottlenck here. But then I tried doing the same with window function and my performance jumped to 2 millions writes per second. Just wow :) Comparing to storm where I had max 320k per second it is amazing. Both fold and window functions were doing the same thing, taking together all the records for the same tenant and user (key by is used for that) and putting it in one batched object with arraylists for the mutations on user profile. After that passing this object to the sink. I can post the code if its needed. In case of fold I was just adding profile mutation to the list and in case of window function iterating over all of it and returning this batched entity in one go. I’m wondering if this is expected to have 20 times slower performance just by using fold function. I would like to know what is so costly about this, as intuitively I would expect fold function being a better choice here since I assume that window function is using more memory for buffering. Also my colleagues when they were doing PoC on flink evaluation they were seeing very similar results to what I am seeing now. But they were still using fold function. This was on flink version 1.0.3 and now I am using 1.2.0. So perhaps there is some regression? Please let me know what you think. Cheers, Kamil.
