Help
Known issues
What does not work in Portal today, what to do instead, and what the releases since launch fixed.
Warnings and limits
- This page lists issues in the release running at portal.clusy.io. Each one is also on the pages it affects.
- An issue stays here until the release that fixes it is running. The limits that are not bugs are on Known limits.
Open
Some CLI commands are console-only on hosted Portal
portal settings KEY VALUE and portal audit --fix answer "Not found." against https://gateway.clusy.io: changing settings and cleaning up leaked resources are an Owner's, in the console. Use Settings, and Machines, Audit, Clean up leaks. Reading settings with portal settings and listing with portal audit work.
A GPU job pinned by --on can be offered a CPU machine after no capacity
When a job launched with an exact --on GPU machine and no --gpu requirement finds no capacity, the choices Portal offers (the same machine on-demand, spot elsewhere) are ranked by the job's requirements, which name no GPU, so one of them can be a machine without a GPU. Give --gpu as well as --on, or choose a GPU offer from portal prices --gpu.
portal connect modal --from-cli needs the Modal SDK
The CLI as installed does not carry the Modal SDK, so --from-cli cannot read your Modal credentials. Connect Modal from the console, or with portal connect modal --file secrets.env. See Modal.
Fixed since launch
| Issue | Fixed (UTC) | Now |
|---|---|---|
portal run --secret NAME and the MCP tool run with secrets were refused ("jobs using stored secrets need the admin scope") | 2026-11-02 | a job in the secret's workspace needs the jobs scope, which the CLI and coding tools hold; each use is an event by name: the request when the job is submitted, the injection when it starts |
A GPU job on AWS showed running for minutes after its program ended, its node lost | 2026-10-30 | the console reads the run as Finishing with what it waits for; nothing after the program's exit is charged |
| Azure Container Apps and AWS Fargate offers waited for a node image | 2026-10-30 | published; a Container Apps job pulls a copy from a registry Portal makes in your subscription, so the public registry's pull limit no longer refuses it |
| A Modal GPU job could sit in provisioning while its image build stalled, with nothing said | 2026-10-30 and 31 | builds are bounded, a failed build is built once more afresh and then the job fails with uv's reason; a GPU job says what it waits for every minute, and a pinned GPU type ends no_capacity after 5 minutes with the other types to choose |
| A small budget could not start a spot job on AWS ("could add $0.0105, budget $0.0100") | 2026-10-30 | the budget is fitted at the price the provider reserves |
| A resume or move accepted just before a release was dropped by it | 2026-10-30 | an accepted request is kept until it runs |
| A release left Portal without an active replica for about a minute | 2026-10-30 | a handover takes a few seconds, and nodes redial at once |
| A move from Modal or Container Apps was refused by the cap room the source had already reserved | 2026-10-30 | the source counts only for the move's window |
| A GPU move from Modal to a VM was refused at restore (torch 2.15.0 against 2.14.1+cu126) | 2026-10-30 | Modal's image installs the same torch build the VMs do |
| 20 small Modal sandboxes committed $198.94 of a $200 cap with $2.78 spent, and the next move was refused | 2026-10-31 | a sandbox holds its lease against the cap, not its 24 hours |
A project that pinned datasets, dill or another package Portal also pins could not build its environment ("No solution found") | 2026-10-31 | Portal's own dependencies are compatible ranges, and your pins win |
portal interrupt on an Azure Container Apps job stopped its execution, Azure ended the replica at once, and the job was lost with no checkpoint | 2026-10-31 | refused: Container Apps gives no interruption notice, and every Container Apps quote, plan and job says so; test resilience there with a suspend or a move |
A script with a StringIO or BytesIO in a global could not pause: its moves waited, and a spot interruption lost the job | 2026-10-31 | both move by value with their positions, and a closed socket moves too |
| Why a move waited was never said | 2026-10-31 | the job's events say which construct keeps the program from pausing |
| A move to a region or instance type the provider cannot price, or from a GPU job to a machine without one, was accepted and failed 40 s later | 2026-10-31 | refused at the request, with the reason |
A machine reclaimed under a notice whose checkpoint failed read running for 11 minutes | 2026-10-31 | the provider is asked 2 minutes past the notice's deadline |
| Several Google Cloud launches at once aborted on "quota unknown" | 2026-10-31 | fixed; a quota that cannot be read says why in the offer's note |
An Azure spot job whose VM Azure had evicted was told "the job goes on" when its suspend was refused, read running with "gone, which is not certain yet" for 10 minutes, was billed 10 min 54 s past the eviction, and its incomplete checkpoint read as kept objects | 2026-10-31 | the refusal says the node is not connected and what Portal checks; the events say what the wait is for; the VM and the checkpoint the eviction cut short are billed to Azure's eviction time; an incomplete checkpoint is said to be not resumable |
Modal jobs failed when their node started: the image's SSL_CERT_DIR held a blocked root certificate | 2026-10-30, the launch release | the blocked certificates are left out and the rest of the store is trusted |
Lambda launches were refused with ssh-key-foreign for accounts an earlier Portal host used | 2026-10-30, the launch release | Portal leaves the old key alone and launches with a key of its own under a name only it holds |