Networks time out and APIs rate-limit. retry_policy=dg.RetryPolicy(max_retries=3, delay=10, backoff=dg.Backoff.EXPONENTIAL) on an asset or op retries a failed step up to three more times, waiting 10, 20 and 40 seconds. Each attempt appears in the event log as STEP_UP_FOR_RETRY and STEP_RESTARTED, and context.retry_number tells the code which attempt it is.
Not every error deserves a retry. A suspended account or a schema that changed will fail every time. Raise dg.Failure(description="...", metadata={...}, allow_retries=False) to stop at once with a message the on-call person can act on. The description and metadata show on the failed step.
Retries hide symptoms if you let them. Retry what is transient, alert on what is not, and keep max_retries small enough that a real outage still fails within minutes.
Up to four attempts, waiting longer each time.