feat: support --scale-down-delay-after-* per nodegroup
#5729
Conversation
k8s-ci-robot
added
do-not-merge/work-in-progress
|
/assign |
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_delay_processor.go
Outdated
Show resolved
Hide resolved
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_delay_processor.go
Outdated
Show resolved
Hide resolved
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_delay_processor.go
Outdated
Show resolved
Hide resolved
vadasambar
changed the title
--scale-down-delay-after-add per nodegroup--scale-down-delay-after-* per nodegroup
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_processor.go
Outdated
Show resolved
Hide resolved
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_delay_processor.go
Outdated
Show resolved
Hide resolved
|
|
||
| if lastScaledUps[nodeGroup.Id()] != nil && | ||
| lastScaledUps[nodeGroup.Id()].ExpectedAddTime.After(currentTime) { | ||
| klog.V(4).Infof("Skipping scale down on node group %s because it was scaled up recently at %v", nodeGroup.Id(), lastScaledUps[nodeGroup.Id()].Time) |
There was a problem hiding this comment.
I have set the log level at v4 but that might not be the best thing to do. It might help to have this at v2 so that the users know what is happening around scale down. Any feedback on this is welcome 🙏
vadasambar
force-pushed
the
feat/3071/scale-down-after-add-per-ng-poc
branch
2 times, most recently
from
4a22c5e
to
24f850b
Compare
|
Took longer than I thought. Sorry everyone who are interested in this PR 🙏 @x13n I think the PR is ready to review. I didn't think it would grow so big. Might be difficult review as whole. I'm open to splitting it if you want (haven't thought too deeply about the how so I am not sure of the effort involved). If it helps, I am happy to schedule a call or go over the PR in the sig meeting. |
|
@x13n wanted to remind you about this PR 🙏 |
|
Apologies for lack of activity here, I didn't have a longer time block to review a large PR like this one recently. This is in my queue, I'll try to get to this in the next couple of days. |
Noted. If it decreases your load, I can do first pass reviews on some of the PRs that you haven't reviewed yet and you can do the second pass. I recently started reviewing PRs hoping to become a reviewer (maybe not as good as you :D). |
|
Sadly my load doesn't only come from PR reviews :) I do appreciate the offer to review other PRs though, this repo would certainly benefit from having more reviewers! |
| limitations under the License. | ||
| */ | ||
|
|
||
| package observers |
There was a problem hiding this comment.
I don't particularly like the idea of having a generic observers package. If we keep on adding different observers here, it will become a mix of completely unrelated objects, except for the fact that they implement the observer pattern. Let's have a specific observer for specific use case, in a specific package.
There was a problem hiding this comment.
I think there is value in having observers under the same package because it makes them easy to identify (just like processors directory). Maybe something like this makes more sense?
├── observers
│ └── nodegroupchange
│ └── scale_state_observer.go
This makes the import path as ../cluster-autoscaler/observers/nodegroupchange. We'd still have a specific package called nodegroupchange while letting the user know that it's an observer.
If we go with specific observer -> specific use case -> specific package approach, the name of the package is going to be nodegroupchangeobserver which IMO is a bit too long. We can shorten it to ngchangeobserver or ngcobserver but that means you need to know ng stands for nodegroup or (nodegroupchange if we shorten it to ngc).
Open to other better approaches as well (I am leaning more towards first). Let me know what you think.
There was a problem hiding this comment.
Grouping under the same observers directory, but having separate packages sounds acceptable to me. I disagree we need to have observer as part of the package name though: referring to the object externally as nodegroupchange.Observer should be well readable, there's no need to repeat the same word in the package name and in the type names within the package.
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_processor.go
Outdated
Show resolved
Hide resolved
| // RegisterFailedScaleUp records failed scale-up for a nodegroup. | ||
| RegisterFailedScaleUp(nodeGroup cloudprovider.NodeGroup, reason string, gpuResourceName, gpuType string, currentTime time.Time) error | ||
| // RegisterFailedScaleDown records failed scale-down for a nodegroup. | ||
| RegisterFailedScaleDown(nodeGroup cloudprovider.NodeGroup, reason string, currentTime time.Time) error |
There was a problem hiding this comment.
No implementation of RegisterFailedScaleDown is actually returning non-nil errors, so let's drop the return value from interface. Also, observers are supposed to observe, not interfere with execution, so it actually makes sense for all these functions to not return anything.
There was a problem hiding this comment.
I have removed the code to return errors for all the functions except RegisterScaleUp because I want to use it as a replacement for RegisterOrUpdateScaleUp and return an error in case a negative delta is passed to RegisterScaleUp . More info in #5729 (comment)
Open to any other options/ideas.
There was a problem hiding this comment.
Checking if an int is negative is simple enough to be done before making a call to RegisterScaleUp. It would then be possible to remove the error return type.
| @@ -344,6 +347,250 @@ func TestStaticAutoscalerRunOnce(t *testing.T) { | |||
| mock.AssertExpectationsForObjects(t, onScaleUpMock) | |||
| } | |||
|
|
|||
| func TestStaticAutoscalerRunOnceWithScaleDownDelayPerNG(t *testing.T) { | |||
There was a problem hiding this comment.
Could you make this test table based? Looks like it is sharing a lot of code with other tests in this file.
There was a problem hiding this comment.
Updated the code. Let me know what you think.
| time.Now()) | ||
| currentTime); err != nil { | ||
| return errors.ToAutoscalerError(errors.InternalError, err) | ||
| } |
There was a problem hiding this comment.
Potential future follow-up comment (this PR is already huge so I wouldn't do it here): metrics and events exporting done below look like great candidates for becoming observers of the same notification as well.
|
|
||
| currentTime := time.Now() | ||
|
|
||
| if !p.scaleUps[nodeGroup.Id()].IsZero() && |
There was a problem hiding this comment.
The following 3 blocks of code are almost identical, looks like opportunity for sharing a bit of code.
There was a problem hiding this comment.
I have attempted to make it more DRY. Let me know what you think.
There was a problem hiding this comment.
Thanks, looks better. I'm not sure if recent is the best name here, but I don't have a better idea :)
cluster-autoscaler/processors/scaledowncandidates/scale_down_candidates_delay_processor_test.go
Outdated
Show resolved
Hide resolved
k8s-ci-robot
added
size/XL
|
Thank you for the review @x13n. A little bogged down with multiple things. I will mention you here once I get through addressing all the comments 🙏 |
k8s-ci-robot
added
the
needs-rebase
| recent := func(m map[string]time.Time, d time.Duration, msg string) bool { | ||
| if !m[nodeGroup.Id()].IsZero() && m[nodeGroup.Id()].Add(d).After(currentTime) { | ||
| klog.V(4).Infof("Skipping scale down on node group %s because it %s recently at %v", | ||
| msg, nodeGroup.Id(), m[nodeGroup.Id()]) |
There was a problem hiding this comment.
msg and nodeGroup.Id() arguments should be swapped to fit the format string.
There was a problem hiding this comment.
Good catch. Swapped them now.
|
|
||
| currentTime := time.Now() | ||
|
|
||
| if !p.scaleUps[nodeGroup.Id()].IsZero() && |
There was a problem hiding this comment.
Thanks, looks better. I'm not sure if recent is the best name here, but I don't have a better idea :)
Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: update scale down status after every scale up - move scaledown delay status to cluster state/registry - enable scale down if `ScaleDownDelayTypeLocal` is enabled - add new funcs on cluster state to get and update scale down delay status - use timestamp instead of booleans to track scale down delay status Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: use existing fields on clusterstate - uses `scaleUpRequests`, `scaleDownRequests` and `scaleUpFailures` instead of `ScaleUpDelayStatus` - changed the above existing fields a little to make them more convenient for use - moved initializing scale down delay processor to static autoscaler (because clusterstate is not available in main.go) Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: remove note saying only `scale-down-after-add` is supported - because we are supporting all the flags Signed-off-by: vadasambar <surajrbanakar@gmail.com> fix: evaluate `scaleDownInCooldown` the old way only if `ScaleDownDelayTypeLocal` is set to `false` Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: remove line saying `--scale-down-delay-type-local` is only supported for `--scale-down-delay-after-add` - because it is not true anymore - we are supporting all `--scale-down-delay-after-*` flags per nodegroup Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: fix clusterstate tests failing Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: move back initializing processors logic to from static autoscaler to main - we don't want to initialize processors in static autoscaler because anyone implementing an alternative to static_autoscaler has to initialize the processors - and initializing specific processors is making static autoscaler aware of an implementation detail which might not be the best practice Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: revert changes related to `clusterstate` - since I am going with observer pattern Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: add observer interface for state of scaling - to implement observer pattern for tracking state of scale up/downs (as opposed to using clusterstate to do the same) - refactor `ScaleDownCandidatesDelayProcessor` to use fields from the new observer Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: remove params passed to `clearScaleUpFailures` - not needed anymore Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: revert clusterstate tests - approach has changed - I am not making any changes in clusterstate now Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: add accidentally deleted lines for clusterstate test Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: implement `Add` fn for scale state observer - to easily add new observers - re-word comments - remove redundant params from `NewDefaultScaleDownCandidatesProcessor` Signed-off-by: vadasambar <surajrbanakar@gmail.com> fix: CI complaining because no comments on fn definitions Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: initialize parent `ScaleDownCandidatesProcessor` - instead of `ScaleDownCandidatesSortingProcessor` and `ScaleDownCandidatesDelayProcessor` separately Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: add scale state notifier to list of default processors - initialize processors for `NewDefaultScaleDownCandidatesProcessor` outside and pass them to the fn - this allows more flexibility Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: add observer interface - create a separate observer directory - implement `RegisterScaleUp` function in the clusterstate - TODO: resolve syntax errors Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: use `scaleStateNotifier` in place of `clusterstate` - delete leftover `scale_stateA_observer.go` (new one is already present in `observers` directory) - register `clustertstate` with `scaleStateNotifier` - use `Register` instead of `Add` function in `scaleStateNotifier` - fix `go build` - wip: fixing tests Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: fix syntax errors - add utils package `pointers` for converting `time` to pointer (without having to initialize a new variable) Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: wip track scale down failures along with scale up failures - I was tracking scale up failures but not scale down failures - fix copyright year 2017 -> 2023 for the new `pointers` package Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: register failed scale down with scale state notifier - wip writing tests for `scale_down_candidates_delay_processor` - fix CI lint errors - remove test file for `scale_down_candidates_processor` (there is not much to test as of now) Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: wip tests for `ScaleDownCandidatesDelayProcessor` Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: add unit tests for `ScaleDownCandidatesDelayProcessor` Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: don't track scale up failures in `ScaleDownCandidatesDelayProcessor` - not needed Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: better doc comments for `TestGetScaleDownCandidates` Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: don't ignore error in `NGChangeObserver` - return it instead and let the caller decide what to do with it Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: change pointers to values in `NGChangeObserver` interface - easier to work with - remove `expectedAddTime` param from `RegisterScaleUp` (not needed for now) - add tests for clusterstate's `RegisterScaleUp` Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: conditions in `GetScaleDownCandidates` - set scale down in cool down if the number of scale down candidates is 0 Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: use `ng1` instead of `ng2` in existing test Signed-off-by: vadasambar <surajrbanakar@gmail.com> feat: wip static autoscaler tests Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: assign directly instead of using `sdProcessor` variable - variable is not needed Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: first working test for static autoscaler Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: continue working on static autoscaler tests Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: wip second static autoscaler test Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: remove `Println` used for debugging Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: add static_autoscaler tests for scale down delay per nodegroup flags Signed-off-by: vadasambar <surajrbanakar@gmail.com> chore: rebase off the latest `master` - change scale state observer interface's `RegisterFailedScaleup` to reflect latest changes around clusterstate's `RegisterFailedScaleup` in `master` Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: fix clusterstate test failing Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: fix failing orchestrator test Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: rename `defaultScaleDownCandidatesProcessor` -> `combinedScaleDownCandidatesProcessor` - describes the processor better Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: replace `NGChangeObserver` -> `NodeGroupChangeObserver` - makes it easier to understand for someone not familiar with the codebase Signed-off-by: vadasambar <surajrbanakar@gmail.com> docs: reword code comment `after` -> `for which` Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: don't return error from `RegisterScaleDown` - not needed as of now (no implementer function returns a non-nil error for this function) Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: address review comments around ng change observer interface - change dir structure of nodegroup change observer package - stop returning errors wherever it is not needed in the nodegroup change observer interface - rename `NGChangeObserver` -> `NodeGroupChangeObserver` interface (makes it easier to understand) Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: make nodegroupchange observer thread-safe Signed-off-by: vadasambar <surajrbanakar@gmail.com> docs: add TODO to consider using multiple mutexes in nodegroupchange observer Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: use `time.Now()` directly instead of assigning a variable to it Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: share code for checking if there was a recent scale-up/down/failure Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: convert `ScaleDownCandidatesDelayProcessor` into table tests Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: change scale state notifier's `Register()` -> `RegisterForNotifications()` - makes it easier to understand what the function does Signed-off-by: vadasambar <surajrbanakar@gmail.com> test: replace scale state notifier `Register` -> `RegisterForNotifications` in test - to fix syntax errors since it is already renamed in the actual code Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: remove `clusterStateRegistry` from `delete_in_batch` tests - not needed anymore since we have `scaleStateNotifier` Signed-off-by: vadasambar <surajrbanakar@gmail.com> refactor: address PR review comments Signed-off-by: vadasambar <surajrbanakar@gmail.com> fix: add empty `RegisterFailedScaleDown` for clusterstate - fix syntax error in static autoscaler test Signed-off-by: vadasambar <surajrbanakar@gmail.com>
vadasambar
force-pushed
the
feat/3071/scale-down-after-add-per-ng-poc
branch
from
6aa4922
to
5de49a1
Compare
k8s-ci-robot
removed
the
needs-rebase
|
@x13n I have addressed the review comments. Can I get a review on this again 🙏 |
|
Thanks! This was a lot of work, but I think it is now good to go. /lgtm |
k8s-ci-robot
added
the
lgtm
|
[APPROVALNOTIFIER] This PR is APPROVED This pull-request has been approved by: vadasambar, x13n The full list of commands accepted by this bot can be found here. The pull request process is described here
Needs approval from an approver in each of these files:
Approvers can indicate their approval by writing |
k8s-ci-robot
added
the
approved
|
Big thanks to everyone who put the work in on this! Will this be back ported into older versions of the autoscaler? We'd love to take advantage of this after the next release. |
cherry-pick to 1.29 is clean: #6484 There's a lot of merge conflicts in 1.28 (and I am assuming 1.27 as well). |
|
Can we document this in e.g. https://github.com/kubernetes/autoscaler/blob/master/cluster-autoscaler/FAQ.md? |
Signed-off-by: vadasambar surajrbanakar@gmail.com
What type of PR is this?
/kind feature
What this PR does / why we need it:
We want to support
--scale-down-delay-after-*per nodegroup so that if nodegroup A scales up, it doesn't block nodegroup B from scaling down (which is the case now). For details, check #3071Which issue(s) this PR fixes:
Fixes #3071
Special notes for your reviewer:
This PR is still in WIP and PoC phase. The plan is to get an okay from the community before proceeding with the approach in this PR. If you have any feedback on the PR, I would love to have it.#5729 (comment)Does this PR introduce a user-facing change?
Additional documentation e.g., KEPs (Kubernetes Enhancement Proposals), usage docs, etc.:
P.S.
I had originally intended to support only
--scale-down-delay-after-addand hence the name of the PR branchvadasambar:feat/3071/scale-down-after-add-per-ng-poc. But I have increased the scope of the PR now to support all--scale-down-delay-after-*flags per nodegroup.