Skip to content

Managed savings

The Savings tab shows what a cluster could save, and by default every saving waits for a person’s approval. Managed savings lets Seylo apply the levers you choose on its own, inside limits you agree to, until a date. It is off until an admin switches it on, per cluster, and each lever has its own switch.

  1. Open FinOps → Cost & savings → Savings and pick the cluster. The badge beside Available savings reads Read only.
  2. As an admin, choose Let Seylo manage savings.
  3. The dialog, Let Seylo manage this cluster’s savings, lists each lever with what Seylo Writes, Where, and its Bounds, and the date it ends: 90 days from today. Tick the levers you want. Rightsizing and Fewer, better nodes start ticked; Spot for safe workloads, Better machine sizes and Schedule start unticked.
  4. For spot, choose per namespace how many replicas may run on spot, and tick the production namespaces you allow on spot. For the schedule, name any namespaces to keep awake at night.
  5. Choose Let Seylo manage.

The badge then reads Managed, and a Managed savings panel appears under the headline. What you agreed to is recorded as a grant, signed by you, and shown read-only on Operate → Automation beside every other grant.

The grant is used only for savings. It never lets Nightshift or anything else change a workload, and no other grant lets Seylo make a managed saving.

Every lever has the same rules around it:

  • one change per workload per day, and at most 50% in one step;
  • every change is checked after it is made, and put back if it does not hold, with an email to admins (Managed savings put back);
  • a workload Argo CD or Flux owns gets a pull request, or the exact change to make in Git, never an in-place write. One Helm owns, and one whose owner cannot be read, is left for a person;
  • Pause all automation on Automation stops all of it.

Writes: lowers CPU and memory requests on Deployments to their measured need, and raises a CPU or memory limit a container is pressing on, never both for one resource. It is the same judgement as Rightsizing on the Savings tab.

Bounds: at most 50% down or 50% up per step, one change per workload per 24 hours, never below measured use. Per run, at most 5 workloads, largest saving first, and the one container with the largest saving in each. A step that is cut to 50% does the rest a day later.

Held for a person, and shown in the panel:

  • a lower when the evidence behind it is low-confidence;
  • a lower of memory after an out-of-memory kill in the window;
  • a lower of the CPU request of a workload whose HPA scales on CPU utilisation: the HPA reads utilisation against the request, so lowering it changes when the HPA scales;
  • any change to a workload a VerticalPodAutoscaler manages (one not set to Off).

Checked for 30 minutes: its replicas are available, no container was killed for running out of memory, CPU is not throttled in more than 25% of periods (unless it already was), and, where traces exist, the error ratio has not risen by more than 2 points. Any of these failing puts the value back at once. A change whose pods cannot be read for 90 minutes is put back too: an unchecked change is not kept.

Writes: lowers replica counts and HPA floors that the node reduction plan finds spare, so your node autoscaler can remove the nodes that empty.

Bounds: one change at a time, the next only after the last one is checked; at most 50% per step; never below 1 replica.

Never: node pools and VM sizes. The panel gives you the command for those instead, under For a person.

Checked: any pod left Pending because it cannot be scheduled puts the change back at once, and so does a workload still short of replicas after 30 minutes.

Writes: creates two NodePools in your cluster, seylo-spot-first (spot, falling back to on-demand) and seylo-on-demand-spread. Then, one workload at a time, gives each safe Deployment a toleration for them, a preferred (never required) spot affinity, a spread across the two, and a PodDisruptionBudget where it has none.

Bounds:

  • stateless Deployments with 2 or more replicas only;
  • at most one in three replicas on spot by default. Per namespace you can choose one in two or two in three instead; every choice keeps some replicas on on-demand;
  • production namespaces stay on on-demand until you tick them. A namespace counts as production when its name has prod, production or prd as a dash-separated part, when it or a Deployment in it carries the production labels your policies use, or when its labels could not be read;
  • one workload moved per run, each checked for 30 minutes before the next;
  • nothing is removed before its replacement is Ready.

Checked: the pool reads back Ready; the workload has every replica Ready, nothing Pending, no more replicas on spot than its share, and a spot node actually started. Still no spot node after 90 minutes means no saving, so it is put back.

A spot node taken back. If a spot node of seylo-spot-first disappears and a workload is short of replicas for 10 minutes, that workload alone goes back to on-demand. A workload short for 30 minutes without that is put back too.

Seylo never deletes a NodePool. To stop one, it sets its CPU limit to zero, after its workloads are back home, so Seylo never removes a running node.

Where it runs: only through a node autoscaler whose objects Seylo can write: AKS node auto-provisioning, Karpenter on EKS and EKS Auto Mode. On a cluster with the classic cluster autoscaler or none, Seylo writes nothing and shows the command that turns node auto-provisioning on; it never runs it. On GKE, spot and better sizes are not run, and the panel says so.

It needs agent 0.2.21 or later with the node autoscaler write permission, which is off by default. If it is missing, the panel says so and gives the command. With Helm, upgrading to 0.2.21 and turning it on at once:

Terminal window
helm upgrade seylo-agent oci://ghcr.io/seylo-inc/charts/seylo-agent \
--version 0.2.22 --namespace seylo-agent --reset-then-reuse-values \
--set image.tag=0.2.22 \
--set image.digest=sha256:ed9222f8b6e0beb0fb12873388439ce270304bc5b54dc41a80a96707c552661f \
--set rbac.write.autoscaler.enabled=true

That permission lets the agent create and update NodePools and node classes, never delete them. See what 0.2.21 adds.

Writes: adds the cheaper machine size the Savings tab found to the list of sizes your existing NodePool may use. The sizes it runs today stay allowed, so the list only ever widens. Your autoscaler replaces nodes only by its own consolidation, new node first.

Bounds: only a size your cloud account says is available here and within quota. A size that is kept on the Savings tab with a note about a zone, or one that could not be checked, is never used here.

Checked for 30 minutes: pods Ready and nothing more Pending than before. Otherwise the previous list is put back. If the cloud refuses to start a machine of the new size (no capacity, quota, or not offered), the step is taken back at once, and that size is not used again until a newer read says it is available.

It runs where spot runs, and needs the same agent permission.

Only for a cluster an admin marked This cluster is not production on the Schedule lever, in the hours chosen there. Until then the lever cannot be ticked, and says to mark the cluster first.

Writes: outside your hours, scales each Deployment and StatefulSet to 0 replicas, recording the count on the workload (seylo.in/schedule-replicas) and in Seylo’s own record. Your node autoscaler then removes the empty nodes. At the start of your hours it puts back exactly those counts.

Around the workloads:

  • An HPA pauses itself when its workload is at 0, and is never changed; the wake starts it again.
  • A KEDA ScaledObject is paused with autoscaling.keda.sh/paused-replicas: "0", and the pause removed at the wake.
  • A Flux Kustomization or HelmRelease that owns a workload is suspended for the night, before its workloads sleep, and resumed after they wake.
  • Argo CD is never written. A workload sleeps only when its Application ignores /spec/replicas and either syncs with RespectIgnoreDifferences=true or has no automated sync. Otherwise it stays awake, and the panel gives the exact YAML to add in Git.
  • Helm-installed workloads are scaled too. A helm upgrade at night sets the count back, and Seylo leaves it.

Never touched: kube-*, *-system and openshift-* namespaces, Argo CD, Flux, Karpenter, KEDA and cert-manager, Seylo’s agent and observability, DaemonSets, Jobs and CronJobs, a workload another controller owns, a StatefulSet that deletes its claims when scaled down, the namespaces you listed, and anything annotated seylo.in/schedule: "always-on" (on the workload or its namespace). At most 300 workloads sleep.

A night is held, and tried again every 15 minutes, while there is a change waiting for approval on the cluster, an open incident with a critical alert firing, a Job running in a namespace that would sleep, or a change of Seylo’s still being checked. Anything that cannot be read counts as not clear. A night never starts within 60 minutes of the wake. A cluster with no node autoscaler does not sleep: no node would go.

The wake is checked: every workload back to its count and Ready within 20 minutes. Whatever is not is reported to admins, once a night, and kept awake for 7 days. A workload someone or something else scaled back up during the night is left as it was set, and also kept awake for 7 days.

In the panel, the schedule reads Asleep, Awake or Waking, with Next sleep at ... or Wakes at ... in your time zone, and what happened last night. releaser can choose Wake now, Skip tonight, and Sleep tonight as scheduled to undo a skip. Kept awake lists what stays awake and why.

Stopping the whole cluster is a person’s step. The panel gives the command for your cloud, with its caveats: for example, AKS cannot stop a cluster that uses node auto-provisioning, and EKS has no cluster stop.

Once a cluster is managed, the Managed savings panel under the headline shows:

Part What it holds
The sentence For example Seylo manages the switched-on levers until ..., with the date, or why it is stopped, expired or paused
Saved this month The saving Seylo measured, per lever, with How this is worked out
Levers A switch per lever (admin), with its bounds
What Seylo manages For spot and sizes: which node autoscaler runs, the NodePools Seylo created or widened, what is held and why
Schedule Tonight’s state and the night’s controls
What Seylo changed Every change with its state: Checking, Held, Put back, Needs a person, Restored, Pull request open, Pull request closed or Did not take
For a person What Seylo will not do itself, each with the exact command and a link to the cloud’s documentation

Every change is also a line in What Seylo did on its own on Automation, beside Nightshift’s.

Button Who What it does
Stop managing releaser No new change from now. What Seylo already changed stays. A change still being checked is still checked, and put back if it fails
Restore original admin Stops managing first, then puts back, per workload, the value before Seylo’s first change, each read back. Spot placements come off before their pools stop. A value someone changed since, one that cannot be read, and one made through a pull request are left alone, each with the reason
Renew admin Signs a new grant for another 90 days, through the same dialog. An expired cluster shows Renew managed savings

Stop, Restore original, switching the Schedule lever off, expiry and Pause all automation all wake a sleeping cluster first: nothing that ends management leaves a cluster asleep.

Expiry. The grant ends 90 days after it was signed. Seven days before, admins get Managed savings expiring. After it, Seylo makes no new change until an admin renews; nothing is rolled back.

Grants signed before 11 November 2026 do not include raising limits. Their raises are held for a person with ... Renew managed savings to include raises.

Managed savings costs 5% of the saving Seylo measures, per lever. No saving, no fee. It is on top of the platform fee, and only for clusters you manage: a saving a person approved from the Read only page is covered by the platform fee.

The saving is measured, not promised:

  • Before minus what actually ran. For each hour, Seylo prices what the cluster would have run with the values before its changes, and what it actually ran, from the same prices the Savings tab uses: your bill, the cloud’s list price, or your price file. Both sides use the same prices, so discounts you buy yourself, such as reservations and savings plans, are never counted as Seylo’s saving.
  • Only changes in effect. A change put back, restored, failed or not merged earns nothing. Spot earns only for hours actually on spot. A night’s sleep earns only when the wake is checked.
  • Unpriced is not charged. An hour with a machine that has no price, or that could not be read, earns nothing.
  • A raise earns nothing. It is a risk taken away, not a saving.
  • Each change earns for at most 12 months, and only while the cluster is managed: stop or let it expire, and you stop paying, even for changes that keep saving.

The monthly invoice has one line per lever, for example Managed savings (rightsizing): 5% of $X measured, the same figure as Saved this month. A lever with no saving that month is a free line, not on the invoice. See plans and billing.

Role Can
viewer See the badge, the panel and what Seylo changed
releaser Stop managing; Wake now, Skip tonight, Sleep tonight as scheduled; Pause all automation on Automation
admin Let Seylo manage savings, the lever switches, spot shares and production namespaces, Restore original, Renew