Is it normal for group_by to take 4 minutes to evaluate 2800 servers ?
We often start off with a group_by to simplify the playbook (and not overdesign our inventory script) and have noticed that with a large number of systems it doesn't scale very well.
It was version 1.6.3, but meanwhile it seems that 1.7.2 has this fixed although I haven't confirmed it for every use-case in production. Will check on Monday.
Good deal, my other thing I’d suggest checking is whether the inventory script has the “_meta” optimization, as it may be trying to get hostvars on those hosts it has not asked about it yet.