Autoscaler: Allow setting 'minimum headroom' for autoscaling

[yuvipanda]

I want to be able to say 'if the cluster is more than X% full, scale up until it is not'. This is important in super dynamic clusters that are very spiky - we run a Kubernetes cluster for a University, and a large spike of pods start up when classes start. If we waited for them to fail Scheduling before adding more nodes, this provides them with a suboptimal experience (since it might take several minutes for a new node to spin up).

One problem would be defining what 'full' is, in a way that doesn't duplicate what's in the scheduler.

[davidopp]

We also got this request from a GKE customer recently. So there are at least two people who want it. :)

[yuvipanda]

I also want this on GKE :D

[mwielgus]

We are working on this https://github.com/kubernetes/autoscaler/pull/77.

[yuvipanda]

oooo, awesome! Is it being planned to coincide with 1.8? Or later?

On Thu, Jun 29, 2017 at 5:20 PM, Marcin Wielgus [e-mail]
wrote:

We are working on this #77
https://github.com/kubernetes/autoscaler/pull/77.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/kubernetes/autoscaler/issues/148#issuecomment-312141886,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAB23pfNlnVU2u4j5q4IO9qxCDNN0T3Xks5sJD9PgaJpZM4OJ9Pi
.

--
Yuvi Panda T
http://yuvi.in/blog

[davidopp]

A nice complement to this feature would be a way to pre-pull images to the headroom nodes so that a pending pod pays neither the node creation overhead (headroom feature) nor the image pull overhead (pre-pull feature) and can start running right away. We'd need to figure out a way the cluster admin or user specify which images should be pre-pulled where.

[jonastl]

This. Together with making the scaling a parallel operation would solve our problem.
Waiting for 50-100 new machines to boot and install the backplane software takes _forever_, since scaling up is a serial operation (takes 160 seconds per node exactly on GKE)

[MaciekPytel]

@jonastl Which version of CA are you using? Scaling up shouldn't be serial - CA estimates how many nodes are required and adds them in a single request. And it only waits for request to come back, not for nodes to actually start.

[MaciekPytel]

Sorry, just realised you mentioned GKE - in this case I mean what cluster version are you using (as CA is bundled with cluster version on GKE).

[jonastl]

@MaciekPytel Version 1.6.7

[MaciekPytel]

@jonastl In that case it definitely shouldn't be serial. That being said your comment https://github.com/kubernetes/autoscaler/pull/77#pullrequestreview-54688096 suggest you're using a very unusual setup, so perhaps there is a bug somewhere that only manifests for your setup.

It may be worth creating a new issue for that with some information about your setup (cluster version, cluster size, number of pods and description of how they're scheduled). Alternatively we can have a chat on kubernetes slack and see if there is something we can figure out quickly.

[jonastl]

@MaciekPytel, it turned out that when we enabled resource constraints (CPU and memory) to a degree that filled a node group member, then scaling speed was much improved, so my remark above about serial scaling can be scratched with this new insight.

The solution was non-obvious to us, but now that we've found out about the scalers behavior with the expected (undocumented) knobs turned, we're happy with the scaling speed.

[davidopp]

Is there an ETA on this?

[mwielgus]

Next K8S release (1.9). In 1.8 we were busy improving the performance of the current functionality and this feature makes all the computations much more complex.

[yuvipanda]

Is someone working on this for 1.9?

[yuvipanda]

After some thinking, I've come up with a scheme (for GKE) involving two nodepools that'll satisfy our use cases, and have written it up at https://github.com/berkeley-dsep-infra/data8xhub/issues/7. If anyone with more knowledge of the autoscaler can take a look at that and lmk how terrible the idea is, I would highly appreciate it.

[choldgraf]

Any movement on this? It would be quite useful for ensuring we don't hit ceiling effects before new nodes are requested!

[fejta-bot]

Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale.
Stale issues rot after an additional 30d of inactivity and eventually close.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle stale

[yuvipanda]

Any idea how we can help get this moving? :)

[aleksandra-malinowska]

This can be achieved using pod priority and preemption, see (How can I configure overprovisioning with Cluster Autoscaler?)[https://github.com/kubernetes/autoscaler/blob/master/cluster-autoscaler/FAQ.md#how-can-i-configure-overprovisioning-with-cluster-autoscaler]