Based on the benchmarking we did to compare boom to Kowalski, it's clear that Kowalski is limited by how we read from kafka before submitting alerts for processing. At the moment, in one thread we read from kafka + decode the message + check if alerts are already processed, all sequentially before we submit an alert for processing by one of the workers in the cluster.

This means that we do a lot of work single threaded before submitting anything to the workers for processing, when that work could be done by the workers in order to parallelize it.

In this PR, we edit the kafka reading logic to only poll messages and directly submit them for processing. Then, workers themselves are in charge of decoding the avro data and checking if an alert is already processed (i.e. already in the DB).