Currently, pipelined shuffle will fail-fast the moment it receives data from an attempt other than 0. This was done as an add-on check to prevent data being copied from speculated attempts.
However, in some scenarios (like LLAP), it could be possible that that task attempt gets killed even before generating any data. In such cases, attempt #1 or later attempts, would generate the actual data.
This jira is created to allow pipelined shuffle to download data from any one attempt.