acoustic-carpenter-78188
04/28/2026, 11:57 PMmax_runtime.
• "How long am I willing to wait for this to start?" — unanswerable. An action stuck in QUEUED or WAITING_FOR_RESOURCES waits forever (or until a parent timeout fires elsewhere).
• "What is the total wall-clock budget for this work, including retries?" — unanswerable. With retries=5 and max_runtime=1h, an action can legally consume 6 hours plus queue time before it gives up.
Retries have the same flattening problem. retries=N means "try up to N+1 times back-to-back," with no way to say "wait 30s, then 1m, then 2m before retrying" — which is exactly the right policy for a flaky external API. Authors today either embed manual backoff inside the task body (which loses observability) or accept that retries will hammer a recovering downstream into the ground.
We want to support
@env.task(
retries=flyte.RetryStrategy(
count=5,
backoff=flyte.Backoff(base=timedelta(seconds=30), factor=2.0, cap=timedelta(minutes=5)),
),
timeout=flyte.Timeout(
max_runtime=timedelta(minutes=10),
deadline=timedelta(hours=1),
),
)
async def bounded_retries():
...
flyteorg/flyteacoustic-carpenter-78188
05/23/2026, 6:26 PM