Node join problems
Check the recommended Linux releases and prerequisites. For a machine whose previous cluster was removed, use a fresh command from the new ready cluster; keep the old agent state intact so the standard join flow can handle it.
| Symptom | Likely cause | Safe check |
|---|---|---|
| Token exchange fails | Expired, consumed, wrong API, or copied command | Create one new command for that node; do not reuse another machine’s token |
| GPU family preflight fails | Cluster type conflicts with detected GPU or driver stack | Check cluster GPU type, PCI devices, nvidia-smi, amdgpu binding, and /dev/kfd |
| Controller unreachable | WireGuard endpoint/firewall/route or gateway handshake issue | Check outbound UDP, interface state, system time, and the exact logged endpoint |
| k3s registration times out | Stale credentials, service failure, DNS, or controller path | Inspect k3s-agent status and journal; preserve the join transcript |
| Final inventory registration fails | Temporary API or database contention after Kubernetes joined | Keep the transcript and retry the command; create a fresh command if the token is reported expired or consumed |
sudo wg show
sudo systemctl status wg-quick@ccwg0 --no-pager
sudo systemctl status k3s-agent --no-pager
sudo journalctl -u k3s-agent -n 200 --no-pagerNode or listing is not ready
A joined node is not immediately publishable. Confirm its lifecycle state and timestamps for hardware and benchmark collection. If inventory is pending, verify the cluster can schedule the collector and that the node reports expected devices. If benchmarks are pending, check image access, accelerator runtime, disk space, and sustained GPU health.
A listing can show unavailable because capacity is exhausted, the cluster is not ready, no ready nodes remain, the offer is paused, or the cluster composition changed. Fix the cause; do not expand limits beyond real capacity.
Workspace connection problems
| Symptom | Check in order |
|---|---|
| Generate profile disabled | Rental active → workspace ready → no terminal error |
| No WireGuard handshake | Correct profile → only one conflicting interface → outbound UDP → system clock |
| SSH timeout | Handshake connected → private host/port copied exactly → allowed route present |
| SSH permission denied | Selected public key → matching private key → client identity selection |
| Jupyter/API unavailable | SSH works → tunnel session running → correct remote port → unused local port |
Kubernetes access problems
Verify WireGuard before kubectl. Then confirm the downloaded kubeconfig path and current context. A timeout usually indicates the tunnel or route; “Unauthorized” indicates an invalid/expired credential; “Forbidden” can be the expected result of namespace-scoped authorization; “Exceeded quota” means the manifest asks for more than the rental.
sudo wg show
KUBECONFIG=./computecubed-kubeconfig.yaml kubectl cluster-info
KUBECONFIG=./computecubed-kubeconfig.yaml kubectl auth can-i --list
KUBECONFIG=./computecubed-kubeconfig.yaml kubectl get events --sort-by=.lastTimestampBilling problems
If a Stripe payment completed but Rental Balance did not update, keep the checkout session reference, wait for confirmation, refresh Account → Billing, and inspect the top-up transaction status. Do not pay again until the first attempt is known to be expired or failed.
If the Console cannot confirm a rental start after a connection problem, check Rentals. Retrying the same selection in the same tab checks the original request even if its quote refreshed. A connection error does not prove that launch failed; establish the first request's outcome before starting a different selection.
If a rental cost looks wrong, compare contract hourly price, ready/start time, ended time, paused interval, total billable seconds, reservation refund, final charge, host earnings and included marketplace markup in transaction detail. The reservation and final charge are separate ledger events and should not be added together as two costs.
What to send support
Use Report an issue in the authenticated Console. You can preview safe technical details and keep the CC- reference while delivery is confirmed. Do not paste private keys, tokens or confidential workload contents.
After a connection interruption, keep the Console open while it reconnects. It refreshes the visible marketplace, rental and account data automatically; allow a few seconds for current availability to appear. If the connection remains unavailable, check your network and reload the page.
Console sign-in is stored for the current browser tab. A new independent tab may ask you to sign in again. Logging out also signs out your other Console tabs; switching accounts signs out tabs using the previous account. If sign-in cannot be completed, return to the Console and start a new sign-in attempt instead of reusing the callback URL.
If the Console reports that token verification is temporarily unavailable, wait briefly and retry. This temporary service error does not require signing out or resetting your account. If it persists, send support the time and the sanitized error message.
Send the request or operation reference when shown, account email, approximate time and timezone, cluster/node/listing/rental ID, visible state, exact sanitized error, and the steps already tried. For host joins include the transcript after removing bootstrap tokens, public endpoints only if requested, and no private keys. For billing include Stripe reference IDs but never card data.
Missing or delayed metrics do not prevent you from ending a rental or deleting an eligible cluster. Use the resource status and request reference to check progress. Metrics can remain unavailable while the service recovers; they do not change billing or purchased capacity.
The Controller proxy indicator shows the last observation and its time. “Not yet confirmed” and “Observation out of date” do not confirm a connection failure. Wait for a fresh observation and check your WireGuard tunnel separately. Opening the page does not perform a fresh connection test.
If a cluster action reports an identity verification failure, stop retrying changes and contact support with the operation reference. Do not disable security checks or accept an unexpected identity. A replacement must be confirmed before the action can continue.