OK, I figured out the confusing unit test: the chunkSeriesIterator tests actually filter out values outside of the [mint, maxt] interval, which means that mint and maxt are hard limits (and that instances of Seek() or Next() returning values outside those boundaries are bugs).

But questions (2), (3) and (4) still stand. Thanks.

Hi, thanks for taking a look! You've made some really valid points!

1. Yes, mint and maxt are hard limits and if return values out of that range, it is a bug. I think we are lucky so far because prometheus also does some filtering, but we should definitely fix them in tsdb.

2. Seek should only seek forward and we should not allow it to seek backwards in the same chunk.

3. Seek need not forward the iterator. If there are two Seek(t0), then the second one is essentially a noop.

4. Yes, I think we should. By definition, Seek(t0) will give you the first point equal to or after t0. For the case you highlighted, it might seem a little odd, but you are assuming the delta between the points to be small. If the points are indeed spaced out one every 5 years, for example, if I seek something one year behind, I should see the first value.

Also, we should re-use the iterator if it's the same chunk, we are decompressing the chunk from beginning every single Seek, which should NOT happen :) https://github.com/prometheus/tsdb/blob/master/querier.go#L800

Finally, if you want to add SeekBefore to TSDB, I don't we would want that. I think the premise was that we have a streaming interface with only reading forward. And SeekBefore semantics would mean we might have to go back in time and decompress chunks from the past. I am actually not sure, what you are trying to achieve, but do you mind opening a new issue with the proposal?

Thanks for volunteering to do the work, I look forward to reviewing the PRs and let me know if you need help.

PR coming up soon (still have some documentation to go through).

Regarding question (4) though, my example might not have been the best. What I am worried about is that one might request a chunkedSeries that covers the past 10 minutes and then Seek() for the value 12 minutes ago: due to the mint limitation the result will be from the past 10 minutes, even though other samples might exist between 12 and 10 minutes ago and given a different mint the result would be different.

More importantly I believe that requesting the last 10 minutes and then seeking back beyond that is very likely to be a bug and it is more likely go unnoticed if a value is returned than if it consistently fails.

Oh, and about SeekBefore(), the idea is to find the last value at or before time t, not to seek backwards.

Currently the way the instant value of a series at time t is retrieved is to do a Seek(t - 5min) and then iterate forward, caching all values on the Prometheus side and use that cache to find the last value. This is inefficient (particularly with high resolution data) because t - 5min often falls within a previous chunk which must all be parsed and iterated over only to find the actual value in the next chunk. A SeekBefore() would first attempt to find the value at or before t in the chunk that contains t (where it will be found in 119/120 cases) and only if that fails go back a chunk. But it would all happen within the iterator, so from the client's point of view the iterator itself never goes backwards.

It is simply an optimization for retrieving the instant value of a series without always iterating over the previous 5 minutes. The chunkedSeries still gets created with mint = t - 5min, maxt = t, but instead of blindly iterating over all chunks covering that range, it will only iterate over the last chunk 99%+ of the time. I remember seeing a TODO from @fabxc somewhere in the code regarding this particular optimization.

Anyway, created PR prometheus-junkyard/tsdb#329 which attempts to fix (1), (2) and (3).

I also see (4) as an issue, as well as what At() should do when called either before Next()/Seek() or after they have returned false. But I leave that decision up to you guys.

I added more info to Seek interface, refactored tests to check the mentioned behaviors and adjusted all implementations in prometheus-junkyard/tsdb#665

I belive this PR fixes this issue.

@bwplotka @free Was this resolved in the end?

Is there anything left to do here?

Hello from the bug scrub.

Is there anything left to do here?

We take three years of silence as "no". Closing.