Your Kubernetes Requests Are a Guess. groundcover Wants an Agent to Keep Fixing Them.

Your Kubernetes Requests Are a Guess. groundcover Wants an Agent to Keep Fixing Them.

BackerLeader ●60 ●376 ●556
calendar_today • schedule4 min read

Most Kubernetes CPU and memory requests get set once and never revisited. Someone picks a number at deploy time, before the workload has seen real traffic. Then everyone moves on.

Shahar Azulay, CEO and co-founder of groundcover, sees the result in nearly every cluster. "Requests get set once, at deploy time, by someone making their best guess before the workload has seen real load," he said.

The guess usually runs high, because over-provisioning feels safe. That becomes "spend nobody can attribute." Guess low and you get throttling, OOM kills, and incidents. "Either way, teams find out a week or a month late, from an out-of-memory error or the cloud bill."

groundcover's answer is Autonomous Infrastructure, now in beta. It's an AI agent that right-sizes Kubernetes workloads using telemetry groundcover's observability platform already collects. Azulay said it's built on Wand's optimization engine. I asked him how it works and, more important, how it avoids causing the outages it's supposed to prevent.

One workload at a time isn't enough

Most rightsizing tools evaluate each workload on its own. Azulay says that misses how clusters actually behave.

"A request that's right for one service can still starve its neighbors on a crowded node," he said.

Autonomous Infrastructure factors node contention into every recommendation, so growing one workload doesn't push the node out of resources. Stability is the first constraint. It shrinks what's over-provisioned and grows what's running short.

It also sizes injected sidecars, including Istio, Linkerd, and Vault agents, plus native sidecars. Azulay called sidecars "often where waste hides."

No new agent, no new SDK

The tool runs on data groundcover's eBPF sensor already collects at the kernel level. So the agent works from real production behavior without developers instrumenting each service.

"That means teams don't have to add another collection agent, retrofit their applications with new SDKs, or wait for a separate instrumentation project before getting value," Azulay said. Existing OpenTelemetry data adds context where it's available.

Off by default, one Deployment at a time

The feature is off by default. A team opts a cluster in. Then an admin turns it on one Deployment at a time, singly or in bulk. Only admins can make changes, and every action creates an audit event.

The beta scope is narrow on purpose. It manages Kubernetes Deployments only, so other workload kinds aren't touched. Deactivating it stops new changes, and rolling out the deployment restores the original values.

I asked what teams should keep away from it, such as a production database. Azulay didn't offer a do-not-touch list. He said groundcover plans to add application context to the engine so it knows when it's sizing something like Postgres. For now, the Deployment-only scope keeps most stateful databases out of reach anyway, since those typically run as StatefulSets.

Guardrails for the next traffic spike

Cutting requests saves money until traffic spikes. Azulay said that tension "is the whole reason we built it the way we did." Here's how the beta handles it:

  • Requests only, never limits. The customer's declared limit is the ceiling.
  • Node-aware sizing. One workload's increase doesn't squeeze its neighbors.
  • Hands off the autoscaler. Any resource an HPA or KEDA scales on by utilization is left alone.
  • Least-disruptive changes. It resizes in place with no restart where possible, and falls back to a rollout with a backoff gate where it can't.
  • No GitOps drift. It writes annotations only.

The autoscaler rule matters more than it looks. HPA utilization is measured as a percentage of the request. Change the request underneath it and you change when the HPA scales. Two controllers adjusting the same signal is a good way to get a 3 a.m. page.

The limits rule is about ownership. "Limits are the line a team has drawn for how much a workload is allowed to use, and we treat that as theirs," Azulay said. If a workload's right size would exceed its limit, the tool marks it as Bounded. The team sees "a stability signal rather than having us quietly move their guardrail."

What the savings look like

groundcover is still inviting design partners, so Azulay said it's early for customer results. Each Deployment shows an estimated cost impact, shown net so teams see both savings and what they're spending to add stability. Estimates use default list prices. Admins can enter their own prices and discounts.

He did share savings figures from four customers:

Workloads Monthly savings Reduction
1,626 $134,904 83%
10,017 $86,447 61%
603 $5,859 78%
420 $19,840 57%

Read those as directional, since groundcover didn't detail how each was measured. Still, they match what most platform engineers suspect about their own clusters.

Questions to ask any vendor

I asked Azulay what claim about AI-driven Kubernetes optimization engineers should doubt right now. His answer was direct.

"Be skeptical of anything that says 'set and forget' or fully hands-off," he said. Production changes need to be "visible, bounded and reversible," and a team should see all three before it trusts a tool.

He offered two questions to put to any vendor, including groundcover. Will it show me what it would change in my clusters before I turn anything on? Does it size each workload alone, or account for what else is on the node?

I'd add three more. What happens when it meets an HPA? Will it create drift in my GitOps repo? And how do I put everything back?

The bottom line

Rightsizing isn't a new problem. Kubernetes has had the Vertical Pod Autoscaler for years, and plenty of teams still leave it in recommendation mode because they don't trust it with production.

groundcover's bet is that trust comes from constraints, not confidence. Respect limits. Stay out of the autoscaler's lane. Make every change auditable and reversible. That's a sound design for a beta. Whether it holds up under real traffic is what design partners will find out next.

Either way, setting requests once and finding out from the cloud bill is getting harder to defend.

Want more #devops & #devops deep dives?

Join 5,976+ developers sharing real benchmarks, code breakdowns, and projects. Free forever.

1 Comment

0 votes

More Posts

Your AI Agent Doesn't Need Your Token: Edera's Ariadne Conill on Object Capabilities

Tom Smithverified - Nov 5

Kamal vs Kubernetes: An Honest Comparison for Teams Who Don’t Need 1,000 Services

Alexandre Vazquez - Jul 24

Your AI Doesn't Just Write Tests. It Runs Them Too.

Kevin Martinez - May 12

What Security Teams Want From AI Agents Before They Let Them Act Alone

Tom Smithverified - Nov 18

How a Kubernetes Cluster Works | Control Plane & Worker Nodes

Ijay - Sep 29
chevron_left
21.1k Points • 992 Badges
290Posts
175Comments
145Connections
LLM Training & Evaluation Specialist with hands-on experience building major AI models. As one of th... Show more

Related Jobs

Commenters (This Week)

5 comments
1 comment
1 comment

Contribute meaningful comments to climb the leaderboard and earn badges!