Aliaksandr Valialkin
cfd6aa28e1
lib/promscrape/discovery/kubernetes: refresh endpoints and endpointslices scrape targets every 5 seconds, since they may depend on changed service and pod objects
...
This should make endpoints and endpointslices scrape targets eventually consistent with the maximum delay of 5 seconds after the related service or pod object changes.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1240
2021-05-12 14:10:34 +03:00
Aliaksandr Valialkin
904bbffc7f
lib/promscrape/discovery/kubernetes: start watchers for pods and services before starting watchers for endpoints
...
This should eliminate possible race when an update on endpoints depends on pods and/or services, which are missing in the cache yet.
This could result in missing targets based on endpoints or endpointslices.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1240
2021-05-05 12:23:50 +03:00
Aliaksandr Valialkin
2ab1266593
lib/promscrape/discovery/kubernetes: remove a mutex at urlWatcher - use groupWatcher mutex for accessing all the urlWatcher children
...
This simplifies the code a bit and reduces the probability of improper mutex handling and deadlocks.
2021-04-29 10:14:26 +03:00
Aliaksandr Valialkin
2ec7d8b384
lib/promscrape/discovery/kubernetes: fix a deadlock introduced in eddba29664
...
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1240
Thanks to @f41gh7 for providing the initial idea for deadlock fix at https://github.com/VictoriaMetrics/VictoriaMetrics/pull/1248
2021-04-27 14:57:51 +03:00
Aliaksandr Valialkin
eddba29664
lib/promscrape/discovery/kubernetes: refresh role: endpoints targets on service object removal as Prometheus does
...
This is a follow-up for ae37cfd528https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1240
2021-04-23 20:26:57 +03:00
Aliaksandr Valialkin
ae37cfd528
lib/promscrape/discovery/kubernetes: refresh endpoints and endpointslices targets on service object update like Prometheus does
...
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1240
2021-04-23 20:11:40 +03:00
Aliaksandr Valialkin
59f9960992
lib/promscrape/discovery: remove superflouos check in registerPendingAPIWatchers
...
The check `_, ok := uw.aws[aw]; !ok` isn't needed, since aw cannot exist in uw.aws
because of the check inside subscribeAPIWatcher
2021-04-07 13:07:39 +03:00
Aliaksandr Valialkin
3ec6639bbb
lib/promscrape/discovery/kubernetes: register pending apiWatchers in uw.aws
2021-04-06 11:12:13 +03:00
Lu Jiajing
b59164cf33
fix access to nil *url.URL ( #1180 )
...
* fix access to nil *url.URL
Signed-off-by: Megrez Lu <lujiajing1126@gmail.com>
* Update lib/promscrape/discovery/kubernetes/api_watcher.go
Co-authored-by: Aliaksandr Valialkin <valyala@gmail.com>
2021-04-05 22:25:31 +03:00
Aliaksandr Valialkin
b46194472f
lib/promscrape/discovery/kubernetes: reduce CPU time spent on registering big number of Kubernetes objects shared among big number of scrape jobs
...
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1182
2021-04-05 22:04:30 +03:00
Aliaksandr Valialkin
a51d0ec6ec
lib/promscrape/discovery/kubernetes: load objects missing in local cache from api seriver in getObjectByRole()
...
This should fix possible race for `role: endpoints` and `role: endpointslices` service discovery,
when the referred `pod` and `service` objects aren't propagated to urlWatcher cache yet.
See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1182#issuecomment-813353359 for details.
2021-04-05 20:31:17 +03:00
Aliaksandr Valialkin
f010d773d6
lib/promscrape/discovery/kubernetes: synchronously load Kubernetes objects on first access
...
Remove async registration of apiWatchers, since it breaks discovering `role: endpoints` and `role: endpointslices` targets,
which depend on pod and service objects.
There is no need in reloading `endpoints` and `endpointslices` targets if the referenced `pod` or `service` objects change,
since in this case the corresponding `endpoints` and `endpointslices` objects should also change because they contain
ResourceVersion of the referenced `pod` or `service` objects, which is modified on object update.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1182
2021-04-05 14:20:12 +03:00
Aliaksandr Valialkin
5b08e6fb16
lib/promscrape/discovery/kubernetes: properly track objects with the same names in multiple namespaces
...
This is a follow-up for 12e4785fe8https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1170
2021-04-02 14:45:32 +03:00
Aliaksandr Valialkin
12e4785fe8
lib/promscrape/discovery/kubernetes: properly discover targets in multiple namespaces
...
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1170
2021-04-02 14:28:30 +03:00
Aliaksandr Valialkin
f39c84b21f
lib/promscrape/discovery/kubernetes: typo fix in error message
2021-03-26 12:46:14 +02:00
Aliaksandr Valialkin
9761ffd161
lib/promscrape/discovery/kubernetes: properly handle too old resource version error message from Kubernetes watch API
2021-03-26 12:28:10 +02:00
Aliaksandr Valialkin
b88806ecbf
lib/promscrape/discovery/kubernetes: do not start object watcher until initial objects are loaded
2021-03-14 21:55:00 +02:00
Aliaksandr Valialkin
d409898515
lib/promscrape/discovery/kubernetes: further optimize kubernetes service discovery for the case with many scrape jobs
...
Do not re-calculate labels per each scrape job - reuse them instead for scrape jobs with identical Kubernetes role
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1113
2021-03-14 21:14:53 +02:00
Aliaksandr Valialkin
7a16e8e3a2
lib/promscrape/discovery: fixes after 133b288681
...
- Removed a deadlock in addAPIWatcher
- Do not create unused ScrapeWork objects
- Do not spend CPU resources on creating objectByKey map in addAPIWatcher
This work is based on https://github.com/VictoriaMetrics/VictoriaMetrics/pull/1125
2021-03-13 15:18:51 +02:00
Aliaksandr Valialkin
def014eb75
lib/promscrape/discovery/kubernetes: remove debug lines left after the commit 133b288681
2021-03-12 11:22:33 +02:00
Aliaksandr Valialkin
133b288681
lib/promscrape/discovery/kubernetes: use a single watcher per apiURL
...
Previously multiple scrape jobs could create multiple watchers for the same apiURL. Now only a single watcher is used.
This should reduce load on Kubernetes API server when many scrape job configs use Kubernetes service discovery.
2021-03-11 16:43:04 +02:00
Aliaksandr Valialkin
bebcb8130c
lib/promscrape/discovery/kubernetes: localize Bookmark parsing code
...
This is a follow-up for e772d1c920
2021-03-11 13:08:08 +02:00
Aliaksandr Valialkin
e772d1c920
lib/promscrape/discovery/kubernetes: reduce load on Kubernetes API server by using watch bookmarks
...
This allows continuing object watch from the last bookbark instead of reloading all the objects
on watch errors or timeouts.
See https://kubernetes.io/docs/reference/using-api/api-concepts/#watch-bookmarks
2021-03-10 15:06:35 +02:00
Aliaksandr Valialkin
4b18a4f026
lib/promscrape/discovery/kubernetes: remove too verbose logs about starting and stopping the watchers
...
Log the number of objects loaded per each watch url
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1113
2021-03-09 15:05:14 +02:00
Aliaksandr Valialkin
ab4f090c63
lib/promscrape/discovery/kubernetes: reduce memory usage further when big number of scrape jobs are configured for the same kubernetes_sd_config role
...
Serialize reloading per-role objects, so they don't occupy too much memory when objects for many scrape jobs are simultaneously refreshed.
Do not reload per-role objects if they were already refreshed by concurrent goroutines. This should reduce load on Kubernetes API server
when big number of scrape jobs are configured for the same Kubernetes role.
This is a follow-up for 17b87725edhttps://github.com/VictoriaMetrics/VictoriaMetrics/issues/1113
2021-03-07 19:51:03 +02:00
Aliaksandr Valialkin
17b87725ed
lib/promscrape/discovery/kubernetes: reduce memory usage when Kubernetes service discovery is configured on a big number of scrape jobs
...
Previously vmagent was creating a separate Kubernetes object cache per each scrape job.
This could result in increased memory usage when monitoring a Kubernetes cluster with big number of objects (pods / nodes / services, etc.)
as seen at https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1113
Now it uses a shared map of scrape objects across multiple scrape jobs.
2021-03-05 17:29:55 +02:00
Aliaksandr Valialkin
c9a25e931b
lib/promscrape/discovery/kubernetes: move apiWatcher code to a separate file
2021-03-05 12:36:05 +02:00
