It does, mostly, i.e. with some provisos. Past deadline tasks started will be let to run except if the backend has decided the assignment is bad. It's the reason to run with some buffer as any repairs assigned to your nodes will start with a delay. Then the slow/late returners still get a chance to report the task and yours will then be cancelled by instruction from the servers.

What i'm doing at times is killing them when it's obvious there was some hickup somewhere piling up way too many, yes the agent has it's scheduling shortcomings. This particularly happens when the server has the wrong, speak -low- meantime, but the tasks distributed suddenly became much harder.