The current implementation creates a thread per block and in the CSV reader this hurts performance just a bit. However, in the IPC reader this hurts performance even more.
Instead the readahead can move inside the background generator and the background generator task can keep running until the queue fills up and then restart when the queue has drained enough for a substantial amount of work to be done.
In my test CSV case this dropped the # of thread tasks created from ~2.5k to ~100.
Reporter: Weston Pace / @westonpace
Assignee: Weston Pace / @westonpace
PRs and other links:
Note: This issue was originally created as ARROW-12097. Please see the migration documentation for further details.