Series fetch issue about thanos HOT 9 CLOSED

This is on certain query: graph?g0.range_input=2w&g0.end_input=2017-11-23%2013%3A48&g0.expr=go_goroutines&g0.tab=0

(after some time and then 100% reproducible)

from thanos.

This started happening just after you started compaction again?

from thanos.

well not really. I was doing lot's of queries for different things and it started on certain query after some time.. so not sure if there is any correlation with compaction. (new compactor is running 1,5h)

from thanos.

Yes, did you ever see this happen before that though?

from thanos.

nope

from thanos.

I mean this ultimately just seems like some series entries get a good bit longer than the 4KB we expected.
It seems generally possible given we generated a lot of chunks initially with out 5s scrape interval. If you compact a week worth of those together, the series entry has to reference several hundred chunks.

13KB of chunk-indexing data for a single series in a single block actually makes me a bit worried :) But ultimately anything becomes worrisome if you keep the high resolution in long-term storage.

The answer here is basically to just bump the range we fetch per series.

from thanos.

I just hit it as well. It seems indeed to be non-deterministic for the same query over the same time range... that's rather odd though.

from thanos.

Got 'fixed' by bumping the maximum series size we are reading. Anything fully save will require upstream changes to the storage format.

from thanos.

Revisiting this as we saw some issues of violating this again e.g #552, plus investigate estimating postings entry size as well here: #1839

Given just one error, it might still not worth to change. Just leaving the calculation for max size of series entry:

• Max varint lengths: https://github.com/golang/go/blob/master/src/encoding/binary/varint.go#L19
• max series entry size in bytes: 10 (max uvarint64) + 2 * 5 (max uvarint32) * a number of labels pairs + 10 (max uvarint64) + 3 * 10 (max uvarint64) * chunks

from thanos.