[peterschmidt85]

Changes the displayed provisioning failure to No quota or No balance in dstack Factory (including Sky) when no available offers remain and all matching offers have that availability. Mixed blockers and failed launch attempts keep their existing failure messages.

Surface	Displayed field	Before	After
dstack ps	STATUS	no capacity	no quota / no balance
dstack fleet instance row	ERROR	no_offers	no quota / no balance
UI run/job lists and details	Status	No capacity	No quota / No balance
UI instance details	Termination reason	No offers found	No quota / No balance

These are failure messages. The lifecycle statuses remain failed for jobs/runs and terminated for fleet instances. The UI instance list still shows Terminated; it does not display the termination reason.

Backward compatibility

Jobs keep termination_reason=failed_to_start_due_to_no_capacity; instances keep termination_reason=no_offers. The specific failure is stored in termination_reason_message, so older replicas can read and process these rows during rolling deployments. Older servers still return no capacity as the job's status_message.

New termination-reason enum values are recognized but are not written by this fix. Emitting them is deferred until all supported rolling-upgrade source versions recognize them.

Notes

When provisioning finds no available offers, it fetches them again including unavailable offers. This repeats quota/balance DB queries in dstack Factory (including dstack Sky), and provider API calls on cache misses.

[r4victor]

Deployment: during a rolling update, old and new servers share the DB. If a new server saves one of the new job failure reasons, an old server fails when reading that job because it does not recognize the value. Safe rolling updates need two releases: first teach every server to read and handle these reasons without saving them; then start saving them. This draft must be split into those two releases before deployment.

Is this addressed?

[peterschmidt85]

@r4victor updated in 5e547cb:

• Rolling deployment: retain existing stored reasons and use termination_reason_message for quota/balance. New enums are recognized but not emitted, so older replicas can read and process the rows.
• Callback: removed along with the result wrapper. Existing list/iterable contracts are restored. A second lookup runs only when no offers are available to identify a common quota/balance restriction.

[peterschmidt85]

@r4victor reworked as you suggested:

• Removed callbacks and wrappers; existing offer lookup contracts are unchanged. Only when no available offers are found, a second lookup includes unavailable offers to check for a shared quota/balance restriction.
• Kept existing stored termination reasons; quota/balance details go in termination_reason_message. New enum values are recognized but not written, so older replicas can process these rows during rolling updates.

Ready for review.