the only path i can see is to

@catch(var=task_error)

and save the task

self.input

as an artifact on the task, say to

task_input

then in the job

start

step, we could use the Client API to check all tasks from the previous run to see if any have

task_error

. For the ones that do, aggregate the

task_input

data to form the input for the run. is that a bonafide way to go about it, im not sure if this would somehow break checkpointing. or is there another option?