Skip to content

Approvals

Propose-and-approve is Seylo’s default. When approvals are enabled, which is the default once you sign in through an identity provider, a change made in the portal does not touch the cluster. It becomes a request, and it waits until someone with approver or admin decides.

This page covers the request lifecycle, the self-approval rule, what the approver sees, and what is checked again after approval.

Every request has a kind, shown as a badge in the queue:

Kind Comes from
Release A new image tag, optionally with env and resources (your first deployment), a promotion (environments), or Rollback to this version on a component’s History page (rollback)
Manifest edit The manifest editor
Add component Adding a component through New namespace (a button on Overview)
Env edit, ConfigMap edit Configure → Environment, including copying keys between environments
Scheduled deploy Scheduled deployments. Approving one arms it and deploys nothing yet
Namespace copy Copy a namespace. The approver sees Secrets as key names only, and any Azure step the requester acknowledged
Rollout restart, Resource patch, Optimisation proposal The pages that make those changes
Grant New grant on Automation, or Allow this kind on a waiting change here. An approver signs it with Sign grant; signing records you as the grantor on every change Seylo later makes under it. See grants

The queue is Operate → Approvals for approvers and admins, and Operate → My requests for everyone else. Both show the same requests. Approvers and admins also get Approve and Reject buttons. A releaser sees requires approver instead.

When an admin has scoped roles to namespaces, a person who holds grants sees only the requests in the namespaces their grants cover, and may approve or reject only where a grant gives them approver. A request with no single namespace, such as a cluster-wide grant for Seylo, needs a role grant over every namespace of its cluster.

Status Meaning
pending Waiting for an approver. Nothing has touched the cluster
approved Approved and about to be applied
rejected An approver rejected it, or Seylo withdrew it: it expired, it was out of date, or a Nightshift fix was no longer needed. The note says which. Nothing was applied
applied Every part applied and the read-back matched
partial Some parts applied and some failed. Deliberately not reported as applied
drifted The API server accepted the write, but reading it back shows a different value. A GitOps controller, Helm or a mutating webhook owns that field
failed The cluster rejected the change, or the read-back failed

Approving runs the change immediately. Seylo applies what was submitted and keeps the original requester as the initiator, so the history says who asked and who approved. An optimisation proposal is the one kind that is compared with the live cluster first, item by item. See what is checked again.

There is no withdraw button. To drop a request you submitted, ask an approver to reject it.

A request still pending 72 hours after it was submitted expires. It was worked out from the cluster as it was then, and three days is long enough for that to have changed. It leaves the queue as rejected, with a note that starts Expired: and gives the time it was filed, and nothing is changed. If the change is still needed, submit it again: it is worked out afresh from what is running now.

Scheduled deployments do not expire this way: a schedule’s own run time decides when it stops meaning anything.

  • When a request is submitted, approvers and admins other than the requester are emailed Approval needed: <summary>, subject to each person’s notification preferences.
  • When a request is rejected, the requester is emailed Rejected: <summary> with the approver’s reason. The same email goes out when a request expires or is withdrawn as out of date, with that reason instead.

Each person chooses which of these emails they get under Your notification email.

Each request is a card in the queue. The summary line is written not to understate the change. A release that also rewrites environment variables says so (release(shop): checkout-api -> 1.4.3 (+env x2)), and a configuration-only change does not read like a version bump.

Next to it are the kind, the status, who asked and when, and, if it applies, a red foreign owner chip. Detail opens the rest, ownership first:

  • Ownership. If Argo CD, Flux, Helm or another controller owns anything the request would write, the approver sees the owner, the evidence, the consequence, and whether the requester acknowledged it (with who and when). Approving does not create an acknowledgement. Only the requester’s does.
  • What changes, rendered per kind rather than as raw JSON:
    • Release: each component with its tag change (current → 1.4.3), which aspects it carries (image, env, resources), every environment variable set or removed and whether the edit is a merge or a replace, and any resource changes. The current cluster value of a resource field is not in the request, so it shows as current. A rollback from History replaces env rather than merging it, so it also shows what it removes, by name, under Rollback to commit …: what it does to env, and its summary ends [env REPLACED: removes …]. See rollback.
    • Manifest edit: the complete proposed YAML for each kind it touches. It is not shown as a diff against the current version, so ask submitters to say what changed in the notes.
    • Add component: the form fields, and if the submitter edited the generated YAML by hand, a warning and a diff against what the form would have generated. The edited bytes are what get applied. See manifests.
    • Scheduled deploy: when it runs, in which timezone, the first run time, the catch-up grace and the target tags.
    • Optimisation proposal: each item with its current and proposed value. A request filed before 0.4.9 by Seylo’s retired autoscale engine is requested by Seylo autoscale and says why the change wasn’t made unattended. Decide it as any other request.
  • Notes the requester wrote.

No request shows a Secret value. Seylo does not read them.

What is checked again between approval and apply

Section titled “What is checked again between approval and apply”

A request can sit in the queue for hours, so Seylo does not rely on what was true at submission.

  1. The request is still pending. If two approvers act at once, one wins and the other is told Request "<id>" was decided by someone else a moment ago.
  2. The namespace still exists. If the cluster says the request’s namespace is gone, approval is refused with Not applied - Out of date: namespace <name> no longer exists …, the request is withdrawn, and nothing is changed. Only the cluster’s own answer that the namespace does not exist counts: if Seylo cannot reach the cluster, this check does not refuse, and the checks below decide.
  3. A Nightshift fix is re-checked against the live cluster. Seylo diagnoses the problem again and, for a fix that changes an object, compares the object the fix was worked out from with the object as it is now. If anything else about it changed in the meantime - it was edited, scaled, or deleted and created again - the fix is refused as no longer needed and withdrawn, rather than writing the old version back. See When a fix is no longer needed.
  4. Ownership, at the moment of decision. If a controller has taken over a target since the request was submitted, and the requester did not acknowledge it, approval is refused with the same Refusing to change … message the requester would have seen. If Seylo cannot reach the cluster to perform this check, approval is refused rather than the check skipped: Could not verify resource ownership before approving: … Retry when the cluster is reachable.
  5. Ownership and access again, at the write itself. The code that writes to the cluster checks ownership and namespace access once more, so nothing that reaches the write path can skip the check.
  6. An optimisation proposal’s items are compared with the live values. Before each item is written, Seylo reads its live value and compares it with the current value the approver reviewed. An item that changed since review is refused, naming the field and both values, and ends Generate a new proposal from Cost & proposals (it starts from the live value) and submit that. The other items still run. The request then ends partial, or failed if no item applied. “Cost & proposals” in that message is the node reduction plan under the Fewer, better nodes lever on the Savings tab.
  7. The result is read back. Every write is verified against the live object. That is how a request ends as drifted rather than applied.

What is not re-checked: an image tag is not looked up in the registry when a release is approved, and a release is not dry-run. A tag that does not exist fails at rollout as ImagePullBackOff. Scheduled deploys are the exception: see re-checks at fire time.

If the change did not fully land, the queue says so immediately:

Approved, but the change did NOT fully apply (partial): …

Detail then shows the apply message, a Read-back verification list with a badge per check (verified, unknown, drift or failed), and a Live rollout panel driven by Seylo’s own rollout watcher. The same operation is in Operate → Activity.

On a resource or replica change that is waiting, an admin also sees Allow this kind. It opens the grant form pre-filled so that the same kind of change on that workload can happen without an approval next time: the one resource, and the waiting change’s own size as the limit. It files a grant request, which an approver signs like any other; the waiting change stays here for its own decision. See allow this kind of change.

Reject opens a dialog that requires a reason. Nothing has touched the cluster and nothing will. The reason is shown to the requester on their request and recorded in the audit trail.

If your organisation runs with approvals disabled, changes apply as soon as they are submitted. Each one still writes audit rows, so there is a trail either way. The role checks and the ownership gate still apply, but no second person looks at the change before it runs.

Approvers and admins can read the audit trail under Operate → Activity → Audit log.

You see What it means What to do
requires approver instead of buttons You do not hold approver or admin Someone who does must decide
Request "<id>" is for namespace <name> on cluster <cluster>; your grants as approver there cover ... Your namespace grants do not include the request’s namespace, or the request is cluster-wide and your grants do not cover every namespace Someone whose grants cover it decides, or an admin widens your grant
Request "<id>" is already applied and cannot be approved. Someone decided it first Refresh the queue
Request "<id>" was decided by someone else a moment ago. Two approvers acted at the same moment Refresh. The other decision stands
Refusing to change Deployment shop/checkout-api: … on approve A controller took over the target after submission, and the requester did not acknowledge it The approver cannot acknowledge on the requester’s behalf. Reject it, and the requester resubmits with the acknowledgement, or changes it at the source
Could not verify resource ownership before approving: … Seylo could not reach the cluster to check ownership Check the agent is connected, then approve again
Not applied - Out of date: namespace <name> no longer exists … The namespace was deleted after the request was submitted. The request was withdrawn Nothing to do. If the change is still needed, submit it again where the workload runs now
Not applied - this fix is no longer needed: … changed after this fix was filed … The object a Nightshift fix was worked out from changed before anyone approved it. The request was withdrawn Nothing to do. If the problem is still there, Nightshift proposes a fresh fix from the live object
A request you submitted shows rejected with a note starting Expired: Nobody decided it within 72 hours Submit it again if it is still needed
Approved, but the change did NOT fully apply (failed): … The cluster rejected the write Open Detail. The apply message names the cause
Approved, but the change did NOT fully apply (drifted): … The write landed and was rewritten straight away Change the value where the owning controller reads it
NOT APPLIED: the … of … is … on the cluster now, but this proposal was reviewed against … on an optimisation proposal That item’s live value changed after the proposal was submitted, so it was refused and nothing was written to it Generate a new proposal from Cost & proposals and submit that