"api/series" uses too much memory

[vitkovskii]

Bug Report

What did you do?
Send series query like: /api/v1/series?match[]={le!=""}

What did you expect to see?
Well, this is a very heavy query, so I expect Prometheus to hit timeout or other limits.

What did you see instead? Under which circumstances?
The query is executing for ~30 mins until Prometheus is OOM killed.
Normally Prometheus use ~25Gb of memory. This query forces to use more than 90Gb.

Here is memory usage graph:

Disk read:

Environment
We deploy Prometheus to Kubernetes using Prometheus Operator

@brian-brazil

[brian-brazil]

What version of Prometheus is this?

[d-ulyanov]

2.9.1 build with go1.12.4

[brian-brazil]

How many series does that match, and how many do you have in total? Have you tried 2.9.2?

[vitkovskii]

@brian-brazil We don't know how many metrics match the source query (because it can't finish). But if pass start and end(100 seconds range): /api/v1/series?match[]={le!=""}&start=1557303607&end=1557303707. The query will:

1. Run for ~20 sec for the first byte;
2. Additionally, run for ~2 mins 20 sec to output 497Mb of data;
3. Return the result with 1357676 metrics.

That's for only 100 seconds range. If start/end is omitted query will run for all range (2 weeks). I assume that the source query won't return much more metrics, but it scans many metrics(le!="") for a long time range. And where is no memory/time limits for the query.

[brian-brazil]

Sounds like what I fixed in prometheus-junkyard/tsdb#595, which will be in 2.10

[d-ulyanov]

Hey, @brian-brazil
It seems that /series endpoint implementation has no contexts propagation, maybe better to add ctx there and add timeouts? What do you think?

[brian-brazil]

Given how the TSDB is currently implemented, it wouldn't make a difference.

[zemek]

we're also experiencing this issue in 2.15.2
(i think this is actually what was causing our issues in #6595 )

[mateuszlitwin]

We are hitting this issue on production. It seems like like /series endpoint can run for a very long time for certain queries, eventually OOMing the database (most of the memory is allocated for labels that would be returned by the endpoint). Based on the current implementation (v2.16) it looks like there is no timeout mechanism for the /series endpoint. What is the recommended solution for this problem?

Top 25 cumulative allocations https://gist.github.com/mateuszlitwin/523be876fb8c568df6f0472829a3ec1d

This is the loop that is doing the allocations (func (api *API) series(r *http.Request) apiFuncResult).

metrics := []labels.Labels{}
for set.Next() {
  metrics = append(metrics, set.At().Labels())
}

[brian-brazil]

I looked into this recently, and as it stands metadata queries are handled the same as normal queries which results in creating a chunk iterator which will page in the first chunk - taking a long time and potentially trashing the page cache. https://github.com/prometheus/prometheus/tree/series-hint has a potential fix.