Skip to main content
One attempt may run ten minutes. A failure gets two more attempts, thirty seconds apart, so the call runs at most three times before .remote() raises or .map() yields None for that input.

​
Cap one attempt

timeout_seconds is the limit for one attempt. Functions default to 3600 and endpoints, ASGI apps, and realtime handlers to 180, per request. An attempt past its limit ends as a timeout failure and the retry policy decides what happens next. A queued function call also waits up to timeout_seconds for a container, and an endpoint request waits up to its limit for a free slot before 429.

​
Retry a failed call

retries is the number of extra attempts after a failure, 3 by default for functions and 0 for endpoints. A call made with the default can therefore run four times, so writes should be safe to repeat, for example keyed on the input, or the function takes retries=0. retry_delay_seconds is the wait between attempts. An attempt retries when your code raised, when it timed out, or when its container was lost. An import error or a failing on_start is a load error, which fails the task without retrying, since the next attempt would fail the same way.

​
Add backoff

RetryPolicy sets the same thing with backoff and a choice of which failures retry:
max_attempts counts the first attempt, so this is retries=3. The delay doubles from two seconds and stops growing at thirty. retry_on_statuses defaults to ["failed", "timeout"]; this policy leaves a timed-out attempt alone, which suits work where a timeout means the input is too large rather than the network flaky. A policy is used as written. retries or retry_delay_seconds given next to one replace its max_attempts or delay_seconds, and the default of three retries applies only when none of the three options is set.

​
Do not repeat long work

retries=0 avoids repeating six hours of work, and preemptible=False keeps the machine from being reclaimed mid-run. Checkpoints in a volume let a new run resume.

​
What a preemption does

Workloads run on preemptible capacity by default, which the provider can reclaim without notice. A preempted attempt goes back to the queue at once, with no delay, and does not spend a retry. The task’s attempt limit grows by one for each preemption. After five preemptions the task fails with a message that names preemptible=False as the way out.

​
Follow the attempts

A task is queued again between attempts until its delay passes, and uv run lazycloud task show <task-id> prints the attempt as 2 of 4 next to the status. The on_retry hook runs in the container after a failed attempt that will be retried, and the task’s callback carries attempt_number and max_attempts with every event.