feat(blog): Add new blog featuring adoption of in-place Pod resource updates#976
Conversation
… updates This commit features a blog post about the adoption of in-place Pod resource updates in Gardener.
… section This commit updates the Vertical Pod Autoscaler section with few wording improvements. Also: - Fix graph edges titles to reflect logic operation rather than technical endpoint/function calls. - Rewording in the in-place updates benefits section.
✅ Deploy Preview for gardener-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
Important Review skippedReview was skipped due to path filters ⛔ Files ignored due to path filters (2)
CodeRabbit blocks several paths by default. You can override this behavior by explicitly including those paths in the path filters. For example, including ⚙️ Run configurationConfiguration used: Path: .coderabbit.yml Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
Pod resource updates
n-boshnakov
left a comment
There was a problem hiding this comment.
Thank you for the blog, @vitanovs! I only have a minor suggestion to leave.
Co-authored-by: Nikolay Boshnakov <nikolay.boshnakov@sap.com>
|
/assign |
ialidzhikov
left a comment
There was a problem hiding this comment.
Thanks for taking the time write a blog pod about the biggest achievement in VPA since its introduction - in-place updates! 🎉
Suggestions inline ⬇️
|
|
||
|  | ||
|
|
||
| With sections covering `VerticalPodAutoscaler` resource overviews (segregated by _update mode_) and panels displaying success rates per resource, the new dashboard can be used for both monitoring and generating status reports on applied resource recommendations. |
There was a problem hiding this comment.
We could add a section with the achievements on our side. We together checked with you that in dev environments 90% of the Pod evictions were replaced with in-place updates.
We should share this success story with concrete numbers and screenshots from the monitoring dashboard. Maybe the number is even bigger now.
| - __Reduced scheduling overhead__: No need to re-schedule `Pod`s across the cluster | ||
| - __Reduced initialization overhead__: Applications __do not__ go through full initialization all over again | ||
| - __Preserved `Pod` identity__: `Pod` names, IPs, and volumes remain unchanged | ||
| - __Improved resource efficiency__: More granular and responsive resource optimization |
There was a problem hiding this comment.
One more strong benefit is the following. Not all workload can gracefully handle restarts/evictions due to bugs or limitations. IMO, we should link issues we were facing that were caused by such evictions.
One such issue is kubernetes/kubernetes#126921. We had VPA for the csi-node-plugin DaemonSet. VPA was evicting the csi-node-plugin Pods to update their resources. Deletion of the csi-node-plugin Pod removes the driver information (with the max allocatable disks) from the CSINode object. On startup, the csi-node-plugin adds the driver to the CSINode object. In the time window the csi-node-plugin is down, the CSINode object does not have information about the volume attachment limits. This makes kube-scheduler to schedule Pods with volumes to Nodes which already exhausted their volume attachment limit. This ends up in Pod PVC to stuck in attaching forever until an operator fixes the issue.
Another major bug that occurred in the past due to eviction/restart: kubernetes/autoscaler#7726.
We also had number of other bugs due to the restarts but we could only mention the above occurrences.
|
|
||
| Having the ability to bypass the rollout process when updating `Pod` resources drastically improves the scaling efficiency. Eliminating the overhead of `Pod` scheduling and application initialization is among the primary benefits of the new _update_ mechanism. The following points summarize the key factors when considering using _in-place_ updates: | ||
|
|
||
| - __Zero-downtime scaling__: Resources are adjusted without `Pod` recreation or service interruption |
There was a problem hiding this comment.
Here, we could make it stronger by explaining the components in Gardener which have downtime with Pod evictions:
- The non-HA Shoot's etcd - when the single replica etcd pod is evicted, the whole control place has downtime
- vali and prometheus run single replica no matter the Shoot is HA or not. Any eviction by VPA of vali/prometheus results in short downtime of the logging/monitoring backend.
| Like any cluster-wide configuration change, migrating the `vpa` resources' _update mode_ presented a unique challenge that solved this by leveraging the [Gardener Resource Manager](https://github.com/gardener/gardener/blob/master/docs/concepts/resource-manager.md) and its extensible architecture. | ||
|
|
||
| We developed a dedicated `MutatingWebhook` that automatically filters relevant `vpa` resources and applies the _update mode_ change, making the migration seamless. The webhook is deployed through the `VPAInPlaceUpdates` _feature gate_ (available in both `gardenlet` and [Gardener Operator](https://github.com/gardener/gardener/blob/master/docs/concepts/operator.md)). | ||
| We also built a _rollback_ mechanism to ensure safety—if the feature gate is disabled, migration scripts automatically revert the _update mode_ changes during `gardenlet` or `operator` initialization. These changes have been available since Gardener [v1.137](https://github.com/gardener/gardener/releases/tag/v1.137.0), making _in-place_ update mode adoption fully operational and ready for production use. |
There was a problem hiding this comment.
"Migration scripts automatically revert…during initialization" — calling them "scripts" undersells what's happening; this is reconciliation logic in gardenlet/operator startup. Consider: "the gardenlet / operator initialization logic automatically reverts the update mode."
Co-authored-by: Ismail Alidzhikov <9372594+ialidzhikov@users.noreply.github.com>
Co-authored-by: Ismail Alidzhikov <9372594+ialidzhikov@users.noreply.github.com>
Co-authored-by: Ismail Alidzhikov <9372594+ialidzhikov@users.noreply.github.com>
Co-authored-by: Ismail Alidzhikov <9372594+ialidzhikov@users.noreply.github.com>
Co-authored-by: Ismail Alidzhikov <9372594+ialidzhikov@users.noreply.github.com>
Co-authored-by: Ismail Alidzhikov <9372594+ialidzhikov@users.noreply.github.com>
How to categorize this PR?
/kind enhancement
What this PR does / why we need it:
This PR features a new brief blog post about the adoption of in-place Pod resource updates in Gardener.
Which issue(s) this PR fixes:
Part of gardener/gardener#12955