Kafka Streams, Flink and StoatFlow answer "is this window done?" with three different clocks. StoatFlow runs two of them: it accepts a late record the way Kafka Streams does, on the operator's stream time, and closes the window the way Flink does, on a single global watermark. Part 1 of a four-part series on event time — including two corrections to a claim our own compatibility matrix made about grace.
A record can fail on the way in, in the middle, or on the way out. StoatFlow gives every gate the same verdicts — continue, fail, or dead-letter — and settles them all inside the exactly-once transaction: DLQ records commit on the same barrier as your output, and what cannot be handled kills the epoch, never the guarantee. The full model, from one bad record to a Kubernetes restart.
Money can only be moved, never created — so a stream that tracks balances should read total = 0 at every consistent cut. The Flink Table API gets it right 0.035% of the time; our Kafka Streams twin sends total to −1,619 … +1,792. StoatFlow holds it at exactly 0, at every one of its committed cuts. Here is why, and the measured proof.
A drop-in Kafka Streams DSL has to stay byte-stable while the engine beneath it stays hidden. Those two goals pull apart — and every interesting bug we hit while obfuscating StoatFlow lived on the seam between them.
KIP-1271/1285 let a state store keep a record's headers next to its value — shipped in Apache Kafka 4.3.0. What the KIP is, where Kafka Streams has got to (and what is still in flight), how StoatFlow 1.0.0 compares — and why a one-varint format change becomes an engine-wide one.
Kafka Streams kept each task's changelog offsets in a separate checkpoint file that could drift from the state it described. KIP-1035, in Kafka 4.3, moves them inside the state store — atomic with the data, and the keystone for transactional state stores.
