#961 Correct docs for interruptible retry behaviour Pull request opened by fg91 The documentation suggests that an interruptible task is retried on a spot instance after one preemption and that this doesn't count against the user-defined retry budget. In contrast to this, in the OSS community sync (21st February 23) @freezing-airport-6809 explained that when

n

retries are allowed,

n-1

of them are on spot instances and for the last attempt, Flyte resorts to a regular instance. I tested the behaviour using this workflow ...

from flytekit import task, workflow, Resources

@task(interruptible=True, retries=5, requests=Resources(gpu="1"))
def train():
    import time
    time.sleep(600)

@workflow
def wf():
    train()

... and by simulating a preemption with

gcloud compute instances simulate-maintenance-event <node name> --zone <zone>

. Result: • Pods with

FLYTE_ATTEMPT_NUMBER

from 0-4 had the spot/preemptible node selector. • The pod with

FLYTE_ATTEMPT_NUMBER=5

did not have the spot/preemptible node selector anymore. flyteorg/flytesnacks GitHub Actions: Mark github pre-release as Release GitHub Actions: Publish artifacts to github release GitHub Actions: Create Prerelease GitHub Actions: Bump Version ✅ 26 other checks have passed 26/30 successful checks