-
Notifications
You must be signed in to change notification settings - Fork 39.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add a metric to track usage of inflight request limit. #58342
Conversation
) | ||
currentMutatingInflightRequests = prometheus.NewGauge( | ||
prometheus.GaugeOpts{ | ||
Name: "apiserver_current_mutating_inflight_requests", |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can mutating
be a metric label?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
#58340 (comment) - It might for cleanness, but I'm not sure if we're not paying for that with a bit of lock contention.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
SG, thanks! Maybe add a comment here to avoid further questions?
523a30a
to
4205590
Compare
@@ -40,6 +50,40 @@ func handleError(w http.ResponseWriter, r *http.Request, err error) { | |||
glog.Errorf(err.Error()) | |||
} | |||
|
|||
// requestWatermark is used to trak maximal usage of inflight requests. | |||
type requestWatermark struct { | |||
sync.Mutex |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think we have a convention to not inheirt directly from sync.Mutex. So let's change to:
lock sync.Mutex
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hmm... I thought that at some point we had one that said that we should do it for simple things like this.
Done.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hmm, maybe I'm not aware of something..
But in any case, I think what you have now is definitely good too.
/approve no-issue |
/lgtm |
@@ -93,6 +138,7 @@ func WithMaxInFlightLimit( | |||
select { | |||
case c <- true: | |||
defer func() { <-c }() | |||
watermark.record(len(nonMutatingChan), len(mutatingChan)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We really want to add a blocking lock here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Fair point. I update a PR. Take a look if you like it better.
@@ -92,7 +137,12 @@ func WithMaxInFlightLimit( | |||
|
|||
select { | |||
case c <- true: | |||
defer func() { <-c }() | |||
nonMutatingLen := len(nonMutatingChan) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
should probably only update the length of the channel we're using in this request
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Done.
|
||
func (w *requestWatermark) record(readOnlyVal, mutatingVal int) { | ||
w.lock.Lock() | ||
defer w.lock.Unlock() |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this is still a blocking lock. might be a good use for non-locking updates (would want to benchmark with contention to be sure)
oldVal := atomic.LoadInt64(&w.readOnlyWatermark)
while oldVal < newVal && !atomic.CompareAndSwapInt64(&w.readOnlyWatermark, oldVal, newVal) {
oldVal = atomic.LoadInt64(&w.readOnlyWatermark)
}
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
My feeling is that 3 atomic operations are worse than lock, but I agree benchmark would be needed to be sure, but even that may change in the future versions of go. Which is why I'd want to proceed with the current version. Locking is now done after processing the request and releasing the "inflight" token, which should prevent this impacting the critical path in the apiserver. Does that sound good to you?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
My feeling is that 3 atomic operations are worse than lock
that would only happen on requests that bumped the high water mark... the majority would be a single atomic read.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Which is why I'd want to proceed with the current version. Locking is now done after processing the request and releasing the "inflight" token, which should prevent this impacting the critical path in the apiserver.
I don't know the http stack well enough to know if this would still block request completion. It seems bad to contend on a single write lock in the handler path for all requests
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Will moving this to separate go routine fix that?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
that would only happen on requests that bumped the high water mark... the majority would be a single atomic read.
missed that this gets reset when the metric is reported. yeah... this approach wouldn't be great.
wait.Forever(func() { | ||
watermark.lock.Lock() | ||
defer watermark.lock.Unlock() | ||
metrics.UpdateInflightRequestMetrics(watermark.readOnlyWatermark, watermark.mutatingWatermark) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
we're holding a lock hanging all requests at this point... either read into local vars or use atomic.ReadInt64/atomic.StoreInt64 to avoid a critical section that has callouts to prometheus here
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'd need to both atomically read and write those values. I'm not sure it's worth it.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Just read them locally to keep the critical section small before calling Prometheus
lock
read to local vars
reset
unlock
pass local vars to Prometheus
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yup, I was going to do this yesterday, but life attacked me before I was able to.
f6c82c9
to
6be50a5
Compare
Addressed all comments. @liggitt PTAL. |
/retest |
} | ||
} | ||
|
||
var watermark requestWatermark |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
since sync.Mutex is not supposed to be copied, passing this to anything would result in a copy of the lock. would be safer as var watermark = &requestWatermark{}
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Done.
/lgtm |
[APPROVALNOTIFIER] This PR is APPROVED This pull-request has been approved by: gmarek, liggitt, wojtek-t Associated issue requirement bypassed by: wojtek-t The full list of commands accepted by this bot can be found here.
Needs approval from an approver in each of these OWNERS Files:
You can indicate your approval by writing |
/test all [submit-queue is verifying that this PR is safe to merge] |
Automatic merge from submit-queue (batch tested with PRs 55792, 58342). If you want to cherry-pick this change to another branch, please follow the instructions here. |
prometheus.MustRegister(currentInflightRequests) | ||
} | ||
|
||
func UpdateInflightRequestMetrics(nonmutating, mutating int) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This metric seems like it belongs in the filter package, not here. We shouldn't have to register metrics in this package. This is a weird coupling.
…-#58340-#58342-upstream-release-1.8 Automatic merge from submit-queue. Automated cherry pick of #58340: Add apiserver metric for number of requests dropped by #58342: Add a metric to track usage of inflight request limit. Cherry pick of #58340 #58342 on release-1.8. #58340: Add apiserver metric for number of requests dropped by #58342: Add a metric to track usage of inflight request limit. ```release-note Add apiserver metric for current inflight-request usage and number of requests dropped because of inflight limit. ```
…58342-upstream-release-1.9 Automatic merge from submit-queue. Automated cherry pick of #58340: Add apiserver metric for number of requests dropped by #58342: Add a metric to track usage of inflight request limit. Cherry pick of #58340 #58342 on release-1.9. #58340: Add apiserver metric for number of requests dropped by #58342: Add a metric to track usage of inflight request limit. ```release-note Add apiserver metric for current inflight-request usage and number of requests dropped because of inflight limit. ```
This one is tricky. The goal is to know how 'loaded' given apiserver is before we start dropping the load, to so we need to somehow expose 'fullness' of channels.
Sadly this metric is pretty volatile so it's not clear how to do this correctly. I decided to do pre-aggregation to smoothen the metric a bit. In the current implementation the metric publishes maximum "usage" of the inflight is previous second.
If you have any ideas please share.
@smarterclayton @lavalamp @wojtek-t @liggitt @deads2k @caesarxuchao @sttts @crassirostris @hulkholden