Troubleshooting cost and savings
Each table is keyed on the text Seylo shows, so you can search for the message
you are looking at. ... stands for a part that names your own workload,
number or reason.
Some messages still name the old Cost & proposals tab: that is the node reduction plan on Savings. Messages about chargeback and the billing integrations are on chargeback and make the figures match your invoice.
Usage and metrics
Section titled “Usage and metrics”| Message begins | Meaning | What to do |
|---|---|---|
Cannot list nodes, so the kubelet metrics source cannot be scraped |
The agent is namespace-scoped and the kubelet source needs Nodes |
Install the seylo-observability chart, use the prometheus or metrics-server source, or install the agent cluster-wide. See right-sizing and usage evidence |
Metrics collection is explicitly DISABLED for cluster |
The cluster’s metrics source is none |
An admin sets a source on the Clusters page. kubelet needs nothing installed |
... in-cluster agent is not connected, so nothing is being collected |
The agent is offline. Figures are from the last successful collection | Check the agent on Clusters |
No metrics have been collected for this cluster yet |
Collection works, but the first window has not closed | Wait for one collection window |
Collection is working (last run ...) but no aggregate falls inside the last ...h |
Nothing in the selected window or namespaces | Widen the window or the namespace filter |
Insufficient data: ... sample(s) collected, at least 30 are required |
Too few samples for a recommendation | Wait. Nothing is proposed from weak evidence |
Insufficient data: the observation window is ...h, need >= 6h |
The window is too short to see daily peaks | Wait, or widen the window |
Only ... day(s) of history: below the 7-day minimum |
No baseline yet, so no proposal item and no forecast | Wait until 7 days are collected |
Hourly data covers ... day(s), below the 14 needed for an hour-of-week profile |
No weekly rhythm yet; levels are still usable | Wait. Confidence stays medium meanwhile |
This workload is flat: peak/trough is ...x, below the 1.3x worth scaling for |
A real pattern, too small to act on | None needed. Unattended reductions on this workload will queue |
No hour-of-week profile is available, so there is nothing to project forward |
Forecast refused: no hourly history | Wait for hourly history to build |
Not enough history to forecast: |
Forecast refused: the baseline is insufficient |
Wait for 7 days of history |
Right-sizing guardrails
Section titled “Right-sizing guardrails”Shown when you send a right-sizing for approval from Savings.
| Message begins | Meaning | What to do |
|---|---|---|
A VerticalPodAutoscaler "..." manages this Deployment in updateMode: ... |
Blocked. The VPA would overwrite the change | Set the VPA to updateMode: Off, or remove it |
HorizontalPodAutoscaler "..." scales this Deployment on CPU UTILISATION |
Warning. Changing the request moves the HPA’s set-point | Re-check the HPA target after applying, or leave CPU out of the change |
Could not read HPA/VPA objects: |
Warning. Seylo could not check for an autoscaler | Treat the change as possibly conflicting with one |
The proposed memory limit ... is at or below the largest working set observed |
Blocked. It would guarantee an OOM kill | Use a higher limit |
... is outside the configured bounds |
Blocked. Below 10m or above 8000m CPU, or below 32 MiB or above 32 GiB memory, by default | Use a value inside the bounds |
No usage data is stored for ... |
Blocked. Nothing justifies a change | Wait for collection |
Cost and prices
Section titled “Cost and prices”| Message begins | Meaning | What to do |
|---|---|---|
no price on file for ... in ... |
No rate in the price book for this instance type and region | Seylo fetches public rates by itself on first use; reload in a minute. If it stays, Refresh prices says why, and you can upload a rate. See cost |
Automatic price fetching is off on this server |
Your Seylo server does not fetch public rates by itself, usually because its outbound access is restricted | An admin uses Refresh prices, or uploads rates |
no node.kubernetes.io/instance-type label |
The node cannot be looked up | Label the node, or accept it as unpriced |
no topology.kubernetes.io/region label |
Prices differ by region | Label the node, or accept it as unpriced |
... node(s) could not be priced, so this is a FLOOR for compute, not a total |
Some nodes are excluded from the figure | Price the nodes listed under Unpriced nodes |
No node on this cluster could be priced. Refresh or upload the price book. |
Nothing is priced | Refresh or upload |
The price book is EMPTY, so every cost and savings figure will be reported as unpriced |
Your organisation’s view of the price book has no rates: no public list price has been fetched and your organisation has uploaded none | Fetch Azure rates or Fetch AWS rates on the Price book card, or upload rates for another cloud. Uploaded rates are private to your organisation; public list prices are shared |
This cluster runs on ..., and its rates cannot be fetched here yet. |
The message goes on to say why: for Google Cloud, your Seylo server has no Google pricing key; for other providers, there is no public list | Upload rates on the Price book card |
This cluster's nodes report more than one cloud |
No single price list applies | The cluster stays unpriced |
None of the ... node(s) reports a spec.providerID |
Bare metal, kind, k3s or similar: no public price list | The cluster stays unpriced |
This cluster's nodes report provider "..." |
A provider Seylo has no price list for | The cluster stays unpriced |
... rows are for SKUs that were not requested ... which means the $filter was ignored |
Azure returned the wrong rows; the whole response was discarded | Refresh again later. Nothing was stored |
AWS publishes no EC2 price file for region |
AWS has no price file for that region | Those nodes stay unpriced |
Ran out of time before reading |
The AWS refresh hit its time budget | Refresh again |
... still have NO price and will be reported as unpriced rather than estimated |
After an Azure refresh, the cloud lists that SKU or region under another name | Those nodes stay unpriced and are left out of every total. Check the nodes’ instance-type and region labels |
hourlyUsd must be a positive number |
Upload row rejected | Fix that row; the valid rows were stored |
region is required - prices differ by region |
Upload row rejected | Add the region |
Node is a CLUSTER-SCOPED kind, but this agent is scoped to namespaces |
A namespace-scoped agent cannot read Nodes, so there is no cost figure | Install the agent cluster-wide to see cost |
No item could be attributed a currency saving |
The proposal has no priced item, typically an empty price book | Add rates. The capacity findings stand |
Proposal rejections
Section titled “Proposal rejections”These appear in the list of rejected candidates under a proposal.
| Message begins | Rule | What to do |
|---|---|---|
Baseline confidence is insufficient: |
Fewer than 7 days of history | Wait |
The container declares no CPU request, so there is nothing to reduce. |
Adding a request is a different change | Set a request deliberately, outside the proposal |
Already right-sized or under-sized: |
The safe value is not below the current request | None |
The proposed ... is below the observed maximum of ... |
Never below the observed maximum | None. The workload needed it once |
Only ...% headroom over the observed p95, below the required 30%. |
Headroom over p95 | None |
At the busiest observed hour (... |
Less than 15% margin at the busiest hour of the week | None. The weekly peak decides |
Usage is growing ...%/week, so this value would be outgrown |
Inside the 8-week minimum | Revisit when growth slows |
This container was OOM-killed ... time(s) during the window |
Memory data is truncated by the limit | Raise memory first |
Already at the floor: |
1 replica | None. Seylo never proposes fewer than 1 |
HPA unknown: |
The HPA list could not be read | Check the agent can list HPAs in that namespace |
PDB unknown: |
The PodDisruptionBudget list could not be read, or was truncated | Check the agent can list PDBs. Absence of evidence is not “no PDB” |
... (No topology was collected for this workload.) |
Replicas, HPA and PDB were not read for it | Regenerate. If it persists, check the agent’s permissions |
PDB ... sets maxUnavailable ...; at ... replica(s) that permits 0 disruptions |
A lower count would make the workload undrainable | Relax the PDB, or keep the count |
PDB ... requires minAvailable ...; at ... replica(s) that permits 0 disruptions |
As above | Relax the PDB, or keep the count |
... (multiple PDBs select these pods, so Kubernetes permits no voluntary disruption) |
Two budgets select the same pods | Remove the overlap |
Container ... declares no CPU request, so "per-replica usage relative to request" has no denominator. |
No CPU request | Set one deliberately |
Not over-provisioned horizontally: |
The workload needs its replicas | None |
HPA ... already allows scaling down to ... |
The HPA can take the saving itself | None |
After the 40% single-step cap the smallest permitted count is ... |
The cap leaves nothing to take this step | None |
Submitting and approving proposals
Section titled “Submitting and approving proposals”| Message begins | Meaning | What to do |
|---|---|---|
Applied together, these changes would leave ...m of allocatable CPU |
Refused as a set: the plan breaches the 20% headroom over forecast peak | Decide which items to drop, then regenerate after they settle |
This proposal is refused as a set and cannot be submitted. |
As above, at submission | As above |
The proposal has changed since it was generated, so nothing was submitted. |
An item you selected no longer exists | Re-read and re-select. The tab regenerates for you |
itemIds[] is required. |
Nothing was selected | Select at least one automated item |
No cluster-level forecast peak was available, so the aggregate headroom check could not run. |
A caveat: the set was not checked | Read the items with that in mind |
NOT APPLIED: the ... of ... is ... on the cluster now, but this proposal was reviewed against ... |
Returned after approval: the item’s live value changed between submission and approval, so it was refused and nothing was written to it. The other items were handled on their own | Generate a new proposal, which starts from the live value, and submit that |
NOT APPLIED: the live ... could not be read |
Returned after approval: Seylo couldn’t read the item’s live value to compare it, so it refused the item | Check the agent is connected, then generate and submit a new proposal |
Node reduction has no executor in this tool yet; it is reported as outstanding, not done. |
Returned after approval for a node-reduction item | Remove the node through your autoscaler or node pool |
Requires a human: this tool does not make availability decisions on your behalf. |
Returned for a manual item | Do the task yourself |
Scale-down blockers
Section titled “Scale-down blockers”| Badge | Code | What to do |
|---|---|---|
| PDB | pdb-would-be-violated |
Raise replicas, or relax minAvailable or maxUnavailable |
| safe-to-evict=false | safe-to-evict-false |
Remove the annotation only if the workload truly tolerates eviction |
| local storage | local-storage |
List scratch volumes in cluster-autoscaler.kubernetes.io/safe-to-evict-local-volumes, or move real data to a PVC |
| no controller | no-controller |
Give the pod an owner, or annotate it safe-to-evict=true if it is disposable |
| system pod | system-pod |
Give the system workload a PDB |
| do-not-disrupt | do-not-disrupt |
Remove karpenter.sh/do-not-disrupt when it is no longer needed |
| scale-down disabled | node-scale-down-disabled |
Remove cluster-autoscaler.kubernetes.io/scale-down-disabled if its reason no longer holds |
| no room elsewhere | insufficient-space-elsewhere |
Right-size first, or accept the node is needed |
| capacity unknown | capacity-unknown |
Upgrade the agent and refresh. An agent before 0.1.9 does not report node allocatable |
Savings: sizes and spot prices
Section titled “Savings: sizes and spot prices”These appear in the Fewer, better nodes and Spot details on Savings. Why a lever reads $0 is on when a lever reads $0.
| Message begins | Meaning | What to do |
|---|---|---|
Availability and quota not checked - connect the cloud account (Integrations > Cloud accounts). |
No cloud account covers this cluster, so the sizes were not checked against what your account can get | Connect the cloud account. Until then, check a size yourself before you use it |
Quota: ... vCPUs left for ... in ...; one ... needs ... |
The family’s vCPU quota has no room for one more machine of that size, so it is left out | Ask your cloud for more quota, or use another size |
Quota: ... regional vCPUs left in ... |
The region’s total vCPU quota has no room | As above |
Spot quota: ... low-priority vCPUs left in ... |
Azure’s spot vCPU quota, which is separate from on-demand, has no room | Ask Azure for more spot quota |
Not offered to this subscription in zone(s) ... of ...; this pool's nodes are not zonal, so that does not apply. |
Kept: the size is restricted only in zones, and this pool is not zonal | None |
Not offered to this subscription in zone(s) ... of ...; this pool runs in zone(s) ... |
Kept: the size is restricted only in zones this pool does not run in | None |
Not listed in zone(s) ... of ...; where this pool's nodes run is not known. |
Google Cloud does not list the type in every zone, and the pool’s zones were not read | Check the nodes carry topology.kubernetes.io/zone |
The vCPU quota could not be read: this account's Seylo role has no servicequotas:GetServiceQuota |
The AWS account was connected before 10 November 2026 | Save the new template for the account on Integrations → Cloud accounts, deploy it, and press Test |
The cloud refused ... here on ... |
A machine of that size was refused when your autoscaler tried to start it. It is not proposed again until a newer read says it is available | None, or ask your cloud about capacity in that region |
Not priced: connect the AWS account (Integrations > Cloud accounts). |
AWS spot prices are read through your connected account | Connect the AWS account |
Not priced: ... connected AWS accounts list ... |
More than one connected AWS account covers the region, and a node does not say which account it is in | Keep one connected account per region for this cluster |
Not priced yet: Seylo is reading Google's Cloud Billing Catalog |
The first read of Google’s prices takes several pages | Reload in a minute |
Not priced: Seylo has no spot price list for ... |
No spot prices are read for this cloud | None. The lever is left out of the total, which says “at least” |
Managed savings
Section titled “Managed savings”These are in the managed savings panel and in the activity log on Automation.
| Message begins | Meaning | What to do |
|---|---|---|
... Renew managed savings to include raises. |
The cluster’s managed grant was signed before 11 November 2026, when raising limits was not part of it, so a raise waits for a person | An admin renews managed savings, or a person approves the raise |
The grant expired on ...; Seylo makes no new change until an admin renews it. |
The 90 days are over | An admin chooses Renew managed savings |
Stopped by ... |
Someone chose Stop managing | An admin switches it on again, which signs a new grant |
The ... autoscaling kill switch is on, so Seylo makes no change until it is released. |
Pause all automation is on | Resume it on Automation when it is safe |
Paused: ... |
The managed grant was suspended, for example after three changes in a row did not hold | Find what fights the change, then renew |
Grant decisions
Section titled “Grant decisions”These are the reasons recorded when a grant decides a change to a dimension. A change that waits for a person changes nothing unattended. Nightshift files it for approval once per workload and dimension. See how each decision is made.
A change filed for approval
Section titled “A change filed for approval”| Reason contains | Meaning | What to do |
|---|---|---|
Already filed for approval (#...), still ...; no second request is filed while one is open. |
A request for the same workload and dimension is open, possibly one a person submitted | Decide that request. If the change has moved, the reason goes on to say so |
Not filed again: #... proposed the same change and was rejected. |
A person rejected this change. It is filed again only when the proposed value moves by at least 10% or the current value changes | None, or submit it by hand to ask again now |
Decision reasons
Section titled “Decision reasons”| Reason contains | Verdict | What to do |
|---|---|---|
The tenant-wide autoscaling kill switch is engaged |
refuse | Resume automation on Automation when it is safe |
The kill switch for cluster ... is engaged |
refuse | Ask whoever paused it - named on Automation - to resume it |
No grant covers ... |
a person decides | Approve it, or request a grant |
The grant is SUSPENDED |
a person decides | Resume the grant once the cause is fixed |
The grant expired at ... |
a person decides | Request a new grant |
... owns this workload and actively reconciles it |
a person decides | Change it in Git, or approve it once by hand with the owner acknowledged |
... owns this workload, so an unattended change would survive until that owner next writes |
a person decides | Change it in the Helm values, or approve once by hand |
ownership could not be determined |
a person decides | Check the agent. Unreadable ownership is treated as foreign |
The baseline behind this proposal is ... confidence, below the ... |
a person decides | Wait for more history, or approve by hand |
The grant does not permit unattended changes to ... |
a person decides | Request a grant that names the dimension |
The grant lists ... as permitted but sets no bounds for it |
a person decides | Request a grant with limits |
The change is ...% ..., beyond the ...% this grant permits in one step |
a person decides | Approve by hand; large moves are meant to be seen |
The target ... is below the grant's floor or above the grant's ceiling |
a person decides | Approve by hand, or request a new grant |
The current value is 0 |
a person decides | The first non-zero value needs a person |
Now is outside the grant's permitted window |
a person decides | None. It is considered again inside the window |
The grant's window could not be evaluated |
a person decides | Request a grant with a valid timezone |
... consecutive unattended changes on this target failed or were reverted. |
refuse, and the grant suspends | Find what is fighting the change, then resume |
... unattended change(s) already made to this target in the last ...h |
a person decides | None. The rate limit resets as the window moves |
The last unattended change was ... minute(s) ago, inside the grant's ...-minute cooldown. |
a person decides | None |
This is a reduction, and the usage pattern behind it was not judged actionable |
a person decides | Approve by hand, or wait for a stronger pattern |
The proposed value equals the current value; there is nothing to do. |
refuse | None |
Requesting a grant
Section titled “Requesting a grant”| Message begins | What to do |
|---|---|
A timezone is required. |
Choose one from the Timezone list, such as Europe/London |
"..." is not a timezone this runtime recognises. |
Choose a zone from the Timezone list |
... is permitted but has no bounds. |
Fill in Max increase per change % and Max decrease per change % |
... exceeds the 50% hard cap |
Set 50 or less |
Hour range ... is invalid. |
Use 0 to 24, start before end; split a window that crosses midnight into two |
minConfidence must be "medium" or "high". |
Choose medium or high |
The expiry is ... days away, beyond the 90-day maximum |
Choose 90 days or fewer |
Raise a CPU request ...: set the CPU request ceiling (millicores); without one it does nothing. |
Fill in CPU request ceiling |
Add an HPA ...: set Most replicas, ... or sreActions names "create-hpa", so send its replica ceiling too |
Fill in Most replicas |
This autoscale policy is no longer valid at approval time, so no grant was created: |
Request it again with a valid expiry |