We went through this exercise, though our starting point was ubuntu 16.04 / nautilus. We reduced our double builds as follows:
- Rebuild each monitor host on 18.04/bionic and rejoin still on nautilus
- Upgrade all mons, mgrs., (and rgws optionally) to pacific
- Convert each mon, mgr, rgw to cephadm and enable orchestrator
- Rebuild each mon, mgr, rgw on 20.04/focal and rejoin pacfic cluster
- Drain and rebuild each osd host on focal and pacific
This has the advantage of only having to drain and rebuild the OSD hosts once. Double building the control cluster hosts isn’t so bad, and orchestrator makes all of the ceph parts easy once it’s enabled.
The biggest challenge we ran into was: https://tracker.ceph.com/issues/51652 because we still had a lot of filestore osds. It’s frustrating, but we managed to get through it without much client interruption on a dozen prod clusters, most of which were 38 osd hosts and 912 total osds each. One thing which helped, was, before beginning the osd host builds, set all of the old osds primary-affinity to something <1. This way when the new pacific (or octopus) osds join the cluster they will automatically be favored for primary on their pgs. If a heartbeat timeout storm starts to get out of control, start by setting nodown and noout. The flapping osds are the worst. Then figure out which osds are the culprit and restart them.
Hopefully your nautilus osds are all bluestore and you won’t have this problem. We put up with it, because the filestore to bluestore conversion was one of the most important parts of this upgrade for us.
Best of luck, whatever route you take.