[FG:InPlacePodVerticalScaling] PLEG watch conditions: rapid polling for expected changes #128518
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
/triage accepted |
bd966a6 to
f460c5b
Compare
f460c5b to
826044c
Compare
|
[APPROVALNOTIFIER] This PR is APPROVED This pull-request has been approved by: tallclair 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 |
|
I pulled in the change to block on the ResizeStatus being cleared from #128377, and ran the e2e tests locally. Without this PR, the tests timed out (at 27/35 resize tests). With this PR, the tests complete in about 10 minutes. |
|
/assign @yujuhong |
de4e5f0 to
7fce6f2
Compare
yujuhong
reviewed
Nov 6, 2024
| "pod", format.Pod(pod), "resourceName", resourceName) | ||
| return err | ||
| } | ||
| resizeKey := fmt.Sprintf("%s:resize:%s", container.Name, resourceName) |
There was a problem hiding this comment.
Do we only run this function updatePodContainerResources when there''s actual resizing to perform, i.e., subsequent sync will skip this function even if the resources haven't converged yet?
I'm thinking about the corner cases such as kubelet restart, whether the watch condition will be set again. And also how often we may see the watch condition being set.
There was a problem hiding this comment.
This function will get called on every SyncPod iteration until the resources converge. I think that's OK here. Updating the resources should take much less than the 1 minute resync period, so I'd expect this to be a very small number of times updating the watch condition. In the restart case, the watch condition will get added back if the resources still haven't converged.
|
Hey @yujuhong @tallclair |
|
/lgtm Can we also run the resize test to verify? |
|
LGTM label has been added. Git tree hash: 69f7ae5ac4a623c24beca49cf1a397b1eeb58481
|
71b3ea2 to
24443b6
Compare
|
Fixed lint error |
|
/test ? |
|
@tallclair: The following commands are available to trigger required jobs:
The following commands are available to trigger optional jobs:
Use
In response to this:
Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository. |
|
/test pull-e2e-gci-gce-alpha-enabled-default |
|
/retest |
|
/lgtm |
|
LGTM label has been added. Git tree hash: b63385b6bdf2f370f2fcb77d61649b9c29d9d0f3
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What type of PR is this?
/kind feature
What this PR does / why we need it:
Motivation & Background:
The Kubelet's PLEG only triggers events and fetches the latest PodStatus from the runtime when a container changes state. This is a problem for in-place pod resize, since resource updates are not considered a state change. To get around this,
pleg.UpdateCachewas called at the end of SyncPod if a pod was resizing to force a "reinspection" of the pod status. This violates some principles of the PLEG, and also means that resizes need to wait a full resync period (1 minute) to detect that the resize completed.What this PR does:
This PR adds a new concept of "Watch Conditions" to the PLEG. If a pod has a watch condition set, then the pod is reinspected (via
runtime.GetPodStatus) on everyRelistloop (every 2 seconds) until the watch condition function returns true, indicating that the condition was reached. The Kubelet sets a watch condition for each resource resize, so it no longer needs to manually do the reinspection via update cache.Not Yet Implemented: When a watch condition is completed, an event is emitted by the PLEG to trigger SyncPod. This should greatly speed up pod resize completion.
Which issue(s) this PR fixes:
Fixes #123940
Special notes for your reviewer:
This PR starts with some refactoring. I can pull the refactor into a separate PR if you prefer.
Does this PR introduce a user-facing change?
/sig node
/priority important-longterm