Hey @sivel, thanks for sharing.
I didn’t take the time to check this out until now, but I wanted to provide some feedback on this.
I’ve tested both of the examples with 2.18.6 and 2.19.0b4, the results are pretty dramatic.
I changed the sleep from 5 to 2 and it is sufficient to see the difference.
For the first example, 2.18.6 completed in 9m23s while 2.19.0b4 completed in 5m33s:
For the second example it’s more or less the same thing:
If we increase the sleep to 5 seconds the disparity increases to 19m23s (!) for 2.18.6 while 2.19.0b4 completed in 5m46s:
I didn’t dig into the specifics for now but one of the things that stood out to me when running the tests is that it seemed the templating (or running the sleep) occurred at least once per batch of forks in 2.18.6.
By this I mean that for 100 hosts and 50 forks, it would seemingly sleep twice per task.
In 2.19.0b4 it felt like it only templated the variable once at the beginning of the task and then it no longer slept.
I could confirm this with more time to test but I will leave it at this for now.
Edit: I later realized I left the sleep at 5 instead of 2 when running the second example. Oops.