Hello all, I used to work in adtech. Adtech was great. CPM for ads is 1-$5 per thousand ad impression. If numbers are 5% off you can blame javascript click trackers.
Now, I work in a non addtech industry and they are really, really serious about exactly once. So there is this blog: https://www.confluent.io/blog/transactions-apache-kafka/ Great little snippet of code. I think I can copy it and implement it correctly. You know, but if you read that section about the zombie fencing, you learn that you either need to manually assign partitions or use the rebalance listener and have N producers. Hunting around github is not super helpful, some code snippets are less complete then even the snipped in the blog. I looked at what spring-kafka does. It does get the zombie fencing correct with respect to fencing id, other bits other bits of the code seem plausible. Notice I said "plausible", because I do not count a few end to end tests running single VM as a solid enough evidence that this works in the face of failures. I have been contemplating how one stress tests this exactly once concept, with something Jepsen like or something brute force that I can run for 5 hours in a row If I faithfully implemented the code in the transactional read-write loop and I feed it into my jepsen like black box tester it should: Create a topic with 10 partitions, Start launching read-write transaction code, start feeding input data, Maybe strings like 1 -1000, now start randomly killing vms with kill -9 kill graceful exits, maybe even killing kafka, and make sure 1-1000, pop out on the other end. I thought of some other "crazy" ideas. One such idea: If I make a transactional "echo", read x; write x back to the same topic. RunN instances of that and kill them randomly. If I am loosing messages (and not duplicating messages) then the topic would eventually have no data.. Or should I make a program with some math formula like receive x write xx. If duplication is happening I would start seeing multiple xx's Or send 1,000,000,000 messages through and consumer logs them to a file. Then use an etl tool to validate messages come out on the other side. Or should I use a nosql with increments and count up and ensure no key has been incremented twice. note: I realize I can just use kafka streams or storm, which has its own systems to guarantee "at most once" but Iooking for a way to prove what can be done with pure kafka. (and not just prove it adtech work (good enough 5% here or there) ) I imagine someone somewhere must be doing this. How where tips? Is it part of some kafka release stress test? I'm down to write it if it does not exist. Thanks, Edward, Thanks, Edward