add namespace to unseal error metric #463
Conversation
| @@ -64,11 +64,6 @@ func init() { | |||
| prometheus.MustRegister(unsealRequestsTotal) | |||
| prometheus.MustRegister(unsealErrorsTotal) | |||
| prometheus.MustRegister(conditionInfo) | |||
|
|
|||
There was a problem hiding this comment.
Why did you have to remove this?
There was a problem hiding this comment.
becaue we can't populate it any more, as the label combination is status abd namespace, and you can't init labels with just reason field, you need both.
There was a problem hiding this comment.
I don't know why it had to be initialized in the first place.
@kskewes I see you added that in #375; is it really necessary these counters are set to 0?
There was a problem hiding this comment.
The backstory is here: https://www.robustperception.io/existential-issues-with-metrics
How it relates:
Sealed Secrets Controller behaviour in an error case as I understand it.
- Startup
- Attempt unseal but there's error. Increase the associated counter metric (with label).
- Retry unseal attempt 5 times and then give up.
Prometheus behaviour:
- Startup
- Detect Sealed Secrets Controller pod (service discovery)
- Scrape
/metricsendpoint on retry interval (10s or 60s or as configured)
These scrapes can happen at any point after the controller has started up.
Situation:
- Prometheus client libraries initialise counters with no labels to 0 by default.
- Prometheus client libraries don't initialise counters with labels by default as it doesn't know what labels will be set.
So we do this manually. - Above enables easy alerting rules as you can check for an increase in the counter. An increase from null/non-existing metric to X value is not easily alert-able on. There's an
absent()function in Prometheus but there are a lot of edge cases we have tried to handle in our other applications with varying degrees of success.
Options (in no specific order):
- Add the namespace label per this PR as is - this is likely to result in missed alerts as a error metric for a namespace is likely to go from
NaNto5within the scrape period interval due to fast retry loop and Prometheus won't catch it. The second increase (or unseal error) will alert via the existing alert. - Add the namespace label per this PR and have a go at rewriting the alert with namespace label and trying to catch
absent()metrics withoffsetetc per the above link. - Add the namespace label and have the controller register the metrics per error reason per namespace. Controller would need to watch Namespaces via Kubernetes API and initialise metric for new namespaces as added. I think cardinality wise this if fine.
- Do nothing (We'd like the label, but little bit to do per above).
There was a problem hiding this comment.
Other options:
-
List+watch for namespaces via k8s api and initialize metrics based in existing namespaces.
-
Advise users on how to craft alerts correctly in case of missing values (using
orand other trickeries as Brian hinted at in the linked doc)
There was a problem hiding this comment.
Cool. The retry interval must have changed since I last tested. :)
With 10s scrape interval provided we have retries over at least 2 it might be ok.
It would be good to update the alert unit test in tests.yaml and the alert itself in jsonnet.
Would you like to do in separate PR? Or I can once we pull next release and can validate.
There was a problem hiding this comment.
Thanks Karl, I don't have bandwidth to take care of this, it would be really great if I could delegate that to you
There was a problem hiding this comment.
Sure. Happy to pickup after current tasks. :)
Can validate retry versus scrape intervals in the unit test.
There was a problem hiding this comment.
Can we initialize at least "fetch" (which doesn't have a namespace) so the counter exists. See #543
|
thank you for taking a look and let me know if you have any problems with it, I might be able to help ;) |
Add part of the initialization from bitnami-labs#463 See bitnami-labs#543
I've been running sealed secrets and I love it thanks!
One caveat is that it's impossible to direct
SealedSecretsUnsealErrorRateHighalert to specific teams. So it ends up that one team, gets all of the failures.I think to solve it we can add the namespace to the metric so that alerts can be directed to teams that own approperiate namespaces.
One typical example is when a secret already exists and then you add a new sealed secret and it doesn't override the existing secret. In this case I would like to alert the appropriate namespace owner ;)