Small and unbalanced downsampled chunks produced about thanos HOT 2 CLOSED

Two unrelated things are likely at play here.

The series in question contains a very large gap, which throws off our heuristic of how many result samples we will generate. I'll see whether this can be improved but it doesn't seem unreasonable to consider this an edge case for which we don't have to reach the target optimum.

The final chunk with the negative timestamps is a small bug.

from thanos.

The culprit is here: https://github.com/improbable-eng/thanos/blob/b4dbb272888930961d1418a8eb655410eb89de46/pkg/compact/downsample/downsample.go#L415-L428

The series in question is highly irregular and has stale markers set almost every 4th sample (something must've been sketchy with the target's reachability at the time).

Now we basically had the edge case of the last batch only containing a single sample and on top that sample being a staleness marker, which is skipped. Thus we add the nonsensical sample with timestamp -1.

Of course we should avoid having single sample batches as well. Will prepare a fix for both.

from thanos.