Event-time skew

Event-time skew - event time skew is the gap between how far along two streams are in event time: the click stream is at 12:04, the payments feed from the batchy upstream is still at 11:58. Each stream is internally fine. The trouble starts when you combine them, because a join or a shared window is only complete when the slower side has caught up, so the combined watermark moves at the pace of the laggard and the fast side's rows wait.

The two failure modes are both self-inflicted. Ignore the skew and you emit joins with half the data, then meet the other half as late data. Or buffer the fast side without bounds and discover the ops version of the problem, watermark alignment.

Do you need it? As a fact about your sources, yes: measure it, it is usually stable, and it sets your honest latency floor. Nothing to deploy. Ubik's lookup joins are watermark-gated: a row from the fast side waits in a buffer until the slow side's watermark proves its match is known, and that buffer is part of the checkpointed state, so a crash does not lose it.