Routine settings

Most routine settings live in the definition, under defaults, and a step can set its own timeout_ms and retry. A run can ask for another durability when it starts. Definitions has the whole format.

Runs

Durability

What a crash may lose. ephemeral runs in memory and journals nothing: the fastest, and a crash loses the run. standard commits the journal in groups, so an effect is durable before the steps that need it run. strict commits the intent and the result of every effect step.

• Field: defaults.durability and durability on definition and run
• Default: standard (defaults.durability, definition), the routine's defaults.durability (durability, run)
• Values: ephemeral, standard or strict
• Invalid value: 422 routine_invalid, 400 invalid_request
• Guide: Durability

json
{"defaults": {"durability": "strict"}}

Concurrency

At most limit runs active per group. A new run in a full group waits (queue), ends canceled (skip), or cancels the oldest active run first (cancel_running).

• Field: defaults.concurrency.limit and defaults.concurrency.on_limit on definition
• Default: none: every run starts (defaults.concurrency.limit), queue (defaults.concurrency.on_limit)
• Values: 1 to 1,000 (defaults.concurrency.limit); queue, skip or cancel_running (defaults.concurrency.on_limit)
• Invalid value: 422 routine_invalid
• Guide: Concurrency

json
{"defaults": {"concurrency": {"limit": 1, "group": "${{ inputs.customer_id }}", "on_limit": "queue"}}}

Pause sandboxes while waiting

Pauses the sandboxes a run execs in while it waits long and nothing else of it runs, so they bill no CPU; the next exec or the end of the run resumes them. Anything else using those sandboxes freezes meanwhile, so set it only for sandboxes the routine's runs have to themselves. Needs sandboxes:write on the publishing key.

• Field: defaults.pause_sandboxes on definition
• Default: false
• Values: true or false
• Invalid value: 422 routine_invalid

json
{"defaults": {"pause_sandboxes": true}}

Steps

Step timeout

How long one attempt of a uses step may take, or a whole map step. A step's own timeout_ms wins over defaults.timeout_ms. A job step such as render/video@1 that sets neither gets its operation's timeout, one hour. Wait and sleep steps are not affected.

• Field: defaults.timeout_ms and steps[].timeout_ms on definition
• Default: 30,000 ms (defaults.timeout_ms), defaults.timeout_ms (steps[].timeout_ms)
• Values: 1 to 86,400,000 ms
• Invalid value: 422 routine_invalid
• Guide: Steps

json
{"defaults": {"timeout_ms": 120000}}

Retries

How often a failed step runs again, and how long it waits between attempts. defaults.retry applies only to pure and read steps; a write step retries only with its own retry, and an irreversible one never. Only failures the callee marks retryable are retried, and every attempt carries the same idempotency key.

• Field: defaults.retry on definition
• Default: none: a step runs once
• Values: max_attempts: 1 to 20; initial_ms: 1 to 3,600,000 (default 200); factor: 1 to 10 (default 2); max_ms: 1 to 86,400,000 (default 30,000); jitter: true or false (default true)
• Invalid value: 422 routine_invalid
• Guide: Retries

json
{"defaults": {"retry": {"max_attempts": 3, "initial_ms": 500}}}

Fixed in the beta

These work one way in the beta, and no request field changes them.

When sandboxes pause

After 30 seconds of waiting. With pause_sandboxes, a run's sandboxes pause once it waits at least 30 seconds (a wait, a sleep, or a signal wait with no deadline).

Lookahead

On. While a step that waits on I/O runs, the node creates the sandbox the next sandboxes/exec@1 step needs. A pre-warmed sandbox whose step does not run is deleted after 30 seconds, and that time bills.

Run history

Kept 30 days. Runs, their steps and their events are kept 30 days.