Loki High CPU Usage

[cf-sewe]

Describe the bug
After some time, loki is running on 100% CPU.

To Reproduce
Steps to reproduce the behavior:

• unable to reproduce by algorithm, happens after some time
• happend 2 times , in different environments
• Originally reported in Slack, some weeks ago

Expected behavior

Environment:

• Kubernetes 1.18 (AWS)
• loki monolithic deployment via Helm
  • Loki v2.0.0 is running with CPU limit of 1 (1 core). Loki is running quite idle normally.
  • Promtail 2.0.0
• boltdb-shipper
• AWS S3

Screenshots, Promtail config, or terminal output

• pprof summary:

Showing nodes accounting for 5400ms, 63.83% of 8460ms total
Dropped 88 nodes (cum <= 42.30ms)
Showing top 10 nodes out of 111
      flat  flat%   sum%        cum   cum%
     760ms  8.98%  8.98%      760ms  8.98%  runtime.duffcopy
     760ms  8.98% 17.97%      770ms  9.10%  unicode/utf8.DecodeRuneInString
     720ms  8.51% 26.48%     1510ms 17.85%  github.com/prometheus/prometheus/promql/parser.(*Lexer).next
     690ms  8.16% 34.63%      800ms  9.46%  runtime.step
     610ms  7.21% 41.84%     1440ms 17.02%  runtime.pcvalue
     450ms  5.32% 47.16%     7920ms 93.62%  github.com/prometheus/prometheus/promql/parser.(*yyParserImpl).Parse
     440ms  5.20% 52.36%     2230ms 26.36%  runtime.gentraceback
     370ms  4.37% 56.74%      870ms 10.28%  github.com/prometheus/prometheus/promql/parser.lexInsideBraces
     330ms  3.90% 60.64%     3100ms 36.64%  github.com/prometheus/prometheus/promql/parser.(*Lexer).NextItem
     270ms  3.19% 63.83%      270ms  3.19%  runtime.futex

• pprof trace (10s)
  pprof_trace.zip

• loki 100 % in k9s
  

• grafana dashboard
  

• kubernetes dashboard for loki-stack
  

[cyriltovena]

Nothing shocking in the trace, you have someone doing a metric query, not sure why you would expect loki to not use CPU. Do you know what the query looks like ?

[cyriltovena]

It should be noted that in the same second you receive a push and you are querying.

[LucaDev]

I'm having the same issues.

After a restart the cpu usage goes back down

[cyriltovena]

Could I get a cpu profile instead of a trace please? Le sam. 5 déc. 2020 à 14:31, Luca Kröger <notifications@github.com> a écrit :

…

I'm having the same issues. After a restart the cpu usage goes back down [image: image] <https://user-images.githubusercontent.com/5812715/101244344-444a8d00-3706-11eb-9c76-46441371efa8.png> — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#3020 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAIBF3JJUJOEWFGCC5LNPK3STIY2FANCNFSM4UKKW5KA> .

[cf-sewe]

Actually the initial profile was already a CPU profile - just forgot to upload it ;)
I retested now with 20s window, all other pprof options are default.

[ec2-user@ip-192-168-100-136 ~]$ go tool pprof http://localhost:3100/debug/pprof/profile?seconds=20
Fetching profile over HTTP from http://localhost:3100/debug/pprof/profile?seconds=20
Saved profile in /home/ec2-user/pprof/pprof.loki.samples.cpu.002.pb.gz
File: loki
Type: cpu
Time: Dec 16, 2020 at 6:29am (UTC)
Duration: 20.12s, Total samples = 15.51s (77.10%)

pprof.loki.samples.cpu.002.pb.gz

[LucaDev]

Thank you @cf-sewe, @cyriltovena I wasn't able to reproduce the bug again until now. It happened two times so far.

[stale]

This issue has been automatically marked as stale because it has not had any activity in the past 30 days. It will be closed in 7 days if no further activity occurs. Thank you for your contributions.

[LucaDev]

@cf-sewe has the error occurred again recently?

[cf-sewe]

Yes that CPU hog occurs reliably after some time in my environment - however only on old Loki version 2.0.0.
Have not updated Loki to v2.1 yet, and won't soon.

[xal3xhx]

issue is occuring for me as well, running a fresh install of everything. will be back tomorrow with more info and what my current setup is

[LucaDev]

This should be reopened.

[xal3xhx]

I agree this needs to be reopened, I’m not home so I can’t test anything but I’m running the most up to date version of truenas and I have grafana inside a jail.

Loki is running on the same jail and is also the most up to date version with default configs.

promtail is running inside another jail on the same system, that jail is running nginx and promtail is looking at the logs with a standard config with 2 replace’s

the dashboard on grafana is the Loki V2.0 nginx dashboard

after some time the cpu on the system gets pinned to 100% I believe it was around 1-2 hours but I could be completely wrong there.

Stoping promtail has no effect proving it’s a Loki issue, restarting Loki fixes the problem for a little while.

as I final note both promtail and Loki are started with the system through an rc.d file using the Daemon command

will post configs and other reports when I get have later today

[xal3xhx]

After looking through the other issues posted it looks like this might be related to issue #3275