Skip to content

KubeAPIErrorBudgetBurn

Šimon Let edited this page May 20, 2021 · 7 revisions

Impact

The overall availability of your Kubernetes cluster isn't guaranteed anymore.
There may be too many errors returned by the APIServer and/or responses take too long for guarantee proper reconciliation.

This is always important; the only deciding factor is how urgent it is at the current rate

Critical

First check the labels long and short.

  • long: 1h and short: 5m: less than ~2 days -- You should fix the problem as soon as possible!
  • long: 6h and short: 30m: less than ~5 days -- Track this down now but no immediate fix required.

Warning

First check the labels long and short.

  • long: 1d and short: 2h: less than ~10 days -- This is problematic in the long run. You should take a look in the next 24-48 hours.
  • long: 3d and short: 6h: less than ~30 days -- (the entire window of the error budget) at this rate. This means that at the end of the next 30 days there won't be any error budget left at this rate. It's fine to leave this over the weekend and have someone take a look in the coming days at working hours.

Example: If you have a 99% availability target this means that at the end of 30 days you're going to be below 99% at this rate.

Runbook

  1. Take a look at the APIServer Grafana dashboard.
    1. At the very top check your current availability and how much percent of error budget is left. This should indicate the severity too.
    2. Do you see an elevated error rate in reads or writes?
    3. Do you see too many slow requests in reads or writes?
  2. Run debugging queries in Prometheus or Grafana Explore to dig deeper.
    1. If you don't see anything obvious with the error rates, it might be too many slow requests. Check the queries below!
  3. Maybe it's some dependency of the APIServer? etcd?

Example Queries for slow requests:

Change the rate window according to your long label from the alert.
Make sure to update the alert threshold too, like > 0.01 to > 14.4 * 0.01 for exmaple.

Slow Read Requests:

If you don't get any results back then there aren't too many slow requests - that's good. If you get results than you know what type of requests are too slow.

Cluster scoped:

(
sum(rate(apiserver_request_duration_seconds_bucket{job="apiserver",le="40",scope="cluster",verb=~"LIST|GET"}[1d]))
- 
sum(rate(apiserver_request_duration_seconds_count{job="apiserver",verb=~"LIST|GET"}[3d]))
)
/
sum(rate(apiserver_request_total{job="apiserver",verb=~"LIST|GET"}[3d]))
> 0.01

Namespace scoped:

(
sum(rate(apiserver_request_duration_seconds_bucket{job="apiserver",le="5",scope="namespace",verb=~"LIST|GET"}[1d]))
- 
sum(rate(apiserver_request_duration_seconds_count{job="apiserver",verb=~"LIST|GET"}[3d]))
)
/
sum(rate(apiserver_request_total{job="apiserver",verb=~"LIST|GET"}[3d]))
> 0.01

Resource scoped:

(
sum(rate(apiserver_request_duration_seconds_bucket{job="apiserver",le="1",scope=~"resource|",verb=~"LIST|GET"}[1d])) or vector(0)
- 
sum(rate(apiserver_request_duration_seconds_count{job="apiserver",verb=~"LIST|GET"}[3d]))
)
/
sum(rate(apiserver_request_total{job="apiserver",verb=~"LIST|GET"}[3d]))
> 0.01

Slow Write Requests

(
sum(rate(apiserver_request_duration_seconds_count{job="apiserver",verb=~"POST|PUT|PATCH|DELETE"}[3d]))
-
sum(rate(apiserver_request_duration_seconds_bucket{job="apiserver",le="1",verb=~"POST|PUT|PATCH|DELETE"}[3d]))
)
/
sum(rate(apiserver_request_total{job="apiserver",verb=~"POST|PUT|PATCH|DELETE"}[3d]))
> 0.01

Learn more about Multiple Burn Rate Alerts in the SRE Workbook Chapter 5.