Build1 publisher3 min readPublished
Kubernetes v1.37 turns scale-to-zero autoscaling on by default
Sixteen of the 67 enhancements in the Garhwal release are stable, and the report names exactly one change that takes effect without anyone enabling it. HPE, meanwhile, consolidated its Terraform providers.
The Engineer · Build desk
What happened
- Kubernetes v1.37, nicknamed Garhwal after the Himalayan range, ships 67 enhancements split into 16 stable, 23 beta, 27 alpha and one deprecation.
- KYAML is now stable, and because every KYAML file remains valid YAML, no existing manifest has to be rewritten to keep working.
- HPE converged its Morpheus Terraform provider into the HPE Terraform provider and released tfmigrator to automate moving existing configurations across.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Anyone running HPAs on object or external metrics has to decide before upgrading whether zero replicas is acceptable for each of them, because that behavior needs no opt-in.
- cost Morpheus provider users inherit a migration they did not schedule; HPE automates the config rewrite, and verifying the resulting plan against production is still their job.
- capability A platform team can introduce KYAML one file at a time with no coordinated cutover, since existing YAML parsers keep reading the output.
- precedent Graduated status gives Kubeflow and Karmada a completed security review and vendor-neutral governance to cite when a dependency has to clear internal review.
"Enabled by default" is the phrase that decides whether a release costs you anything on upgrade day. The v1.37 summary from thenewstack.io applies it once, to HorizontalPodAutoscaler scale-to-zero at beta [3]. The scope is narrow. It covers workloads whose autoscalers read object or external metrics [3]. An HPA driven by CPU or memory sits outside what the report describes. So the pre-upgrade task is a query: list the HPAs with object or external metric targets, then decide for each one whether zero replicas is acceptable.
The full count is 16 stable, 23 beta, 27 alpha and one deprecation, which sums to exactly 67 [2][1]. Fifty of those entries are beta or alpha [2]. The report gives no feature gate defaults for them, so nothing in it supports the conclusion that the other 66 items are inert on upgrade [3]. It names one default change and leaves the gate list to you.
Two items arrive without asking anything. Resilient watch cache initialization is complete, and thenewstack.io ties it to resilience in large clusters and to avoiding control plane outages [6]. Manifest-based admission control is at beta, and pod-level checkpoint and restore at alpha [7].
KYAML is stable, and every KYAML file is still valid YAML [4]. Nobody has to rewrite anything, adoption can go one file at a time, and a tool that already parses YAML parses KYAML. The report names whitespace sensitivity and the "Norway Problem" as the headaches it targets [5].
The migration work in this week's news is HPE's. Morpheus Terraform Provider functionality has been converged into the HPE Terraform provider, and HPE released tfmigrator to automate the move off the standalone provider [8][9]. A vendor that ships the migration tool alongside the consolidation has decided the config change is mechanical enough to script. What the report does not say is what tfmigrator does with existing Terraform state, which is the part a platform team has to test before it trusts the automation. HPE technologists Colin Taylor, Don Wake and Eamonn O'Toole walked through the updates in an HPE Developer Community Meetup session [10]. HPE is also a presenting sponsor of the series the item ran in [11].
CNCF's three recent graduations are a governance milestone. Kubeflow, Karmada and Cloud Native Buildpacks all reached graduated status, out of a catalogue of 228 projects [12][14]. Graduation means the project is highly mature, has completed security reviews, and has a vendor-neutral governance model [13]. Kubeflow's headline number is 260 million PyPI downloads to date, reported with no breakdown of who or what did the fetching [15][4]. "Graduation marks a critical milestone, cementing Kubeflow as a mature option for enterprise AI workloads on Kubernetes," said CNCF CTO Chris Aniszczyk in the graduation announcement [16]. Karmada's v1.19 release advances multi-component scheduling for distributed AI training jobs [17].
What to watch
- Whether upgrade reports from large clusters push the scale-to-zero default back off before v1.38.
- Whether HPE documents what tfmigrator does to existing Terraform state, and whether it sets an end date for the standalone Morpheus provider.
- Whether KYAML starts appearing as generated output from kubectl and manifest tooling now that it is stable.