If you run the same app on more than one Kubernetes cluster, say production on EKS, staging on GKE and a dev cluster on AKS, a simple question turns into three terminals. Which pods aren't ready? Is the api Deployment on the same image everywhere? You switch kubeconfig contexts, run the same kubectl command three times and compare the output by eye.
Skyportal can run that one command on several clusters at the same time. You put the clusters and namespaces in a chat's scope, ask SARA, and she sends one kubectl or helm command to every target at once. Anything that needs your OK comes back as one card with a checkbox per cluster and namespace, so you can approve dev and staging and leave production alone.
Put the clusters in the chat's scope
This works with clusters you already connected to Skyportal, for example with a read-only service account token on EKS, GKE or AKS. Each connected cluster shows up as a host.
In the web chat, click the "Scope:" pill in the chat panel and tick each cluster. A ticked cluster opens its namespace list: "All namespaces (cluster-wide)" first, then one checkbox per namespace, with its environment tag if you set one. Pick the namespaces SARA may touch. Until you pick at least one, the cluster shows "Select at least one namespace, or "All namespaces", to run commands on this cluster."
From the terminal, the CLI sets the same scope when it creates the chat. Repeat --server for each cluster and --namespace as SERVER_ID=NAMESPACE (__all__ means the whole cluster):
skyportalai chat send \
--server 21 --server 22 --server 23 \
--namespace 21=api --namespace 22=api --namespace 23=api \
--wait \
"Which pods in the api namespace are not ready on these clusters?"
Later in the same chat, skyportalai chat select-servers changes the scope. The IDs come from skyportalai kubernetes list.
One question, every cluster at once
When the scope holds more than one cluster and you ask about all of them, SARA can send a single command to all of them. She can also run it once in each of several namespaces, on every cluster in the call. The runs start together instead of one cluster after another.
Questions that fit this well:
- "Which pods in the api namespace are not ready on these clusters?"
- "What image is the api Deployment running on each cluster?"
- "Show the last 20 events in the payments and api namespaces everywhere."
SARA gets the output back as one block per cluster and namespace, labelled with the cluster's name, the namespace and the exit code, with a line like "2/3 succeeded" on top and failures first. If one cluster doesn't answer, only its block fails and the others still come back. Each run has the usual kubectl time limit of 120 seconds, or the command's own --timeout plus 30 seconds.
What the command can look like
The fan-out runs plain kubectl or helm. There is no shell, so pipes, &&, ;, redirects and xargs are refused. If you want grep, ask SARA to read the output instead.
Namespaces have to be clear. When SARA runs a command across several namespaces, the command itself can't carry -n, --namespace or -A, and a cluster-wide command like kubectl get nodes runs once per cluster, not once per namespace. Squeezed flags such as -nprod or -lapp=nginx are refused with a request to write -n <ns> instead, and -A only works when the scope covers every namespace on that cluster.
Clusters and SSH hosts don't mix in one call. If the scope has a GPU VM next to your clusters, a call that targets both is refused, and SARA has to target only the clusters or run the VM part on its own.
One approval card, one checkbox per target
A command that needs approval doesn't run anywhere until you answer. On a fan-out you get a single "Batch Approval" card, "Approve 3 commands on 3 hosts" for example. Its reasons start with where it runs, like "Runs on prod-eks in namespace api", followed by why it asks, such as "Not on any allow list".
Every cluster and namespace is its own row, with an "ns" tag and a checkbox. Say you asked SARA to restart the api Deployment everywhere:
kubectl rollout restart deployment/api -n api
Untick the production row and approve. The restart runs on staging and dev, and SARA is told the production target was refused by you at approval, so she doesn't try it again unless you ask. If you untick everything, the card asks you to pick a target or click Deny.
The card never offers "Always allow" when it covers several hosts. An allow-list entry belongs to one host, and on Kubernetes it is tied to one namespace, so you set it from a single-cluster card.
Production stays read-only
If you marked a namespace as Production, or gave its environment a read-only policy, a fan-out respects it. A restart or a delete that would reach a read-only namespace is blocked for the whole call, with "Reads still run there." SARA can still run get, describe and logs across all your clusters, and you split the write into a call that skips production.
Reads still ask by default. Two things let them through without a card: a matching entry on that host's allow list, or the Read-Only Mode toggle, which runs safe reads without asking. exec, cp, attach and port-forward always ask, whatever the allow list says.
When SARA's command points at a namespace that isn't in the chat's scope, nothing runs. For a single target, she asks with a card like "Add namespace payments on dev-aks to this chat and run ...", and approving adds it to the scope and runs the command.
Get started
Connect your clusters, tick them in the Scope pill and ask SARA your question once. Nothing in this feature is tied to a plan, and the Free plan lets you connect three hosts, each cluster counting as one, which is enough to try it on production, staging and dev. Sign up here, and get the CLI from PyPI.