[SOLVED] Curiosity: OSD Out - Backfill. OSD Destroy. Backfill again?

[AllanM]

Interesting observation:

Steps:

1. Set an OSD to "Out"
2. Wait for Backfill (rebalance) to complete. Healthy status, nice green circle.
3. Set OSD to "Stop"
4. Observe healthy status, no rebalancing needed.
5. Destroy the Out/Stopped OSD.
6. Backfill operations begin again, just like in step 2.

Why does it take 2 rebalances to remove an OSD?

There's probably a perfectly reasonably explanation. I just don't know what it is.


Thanks!

 

[mira]

The placement of objects might change if there's one less OSD to consider.
So the first backfill was to make sure every object has the configured number of replicas.

Then the 2nd backfill was to update the placement of objects based on the layout after the number of OSDs in the cluster changed.

 

[AllanM]

That makes perfect sense! Thanks Mira!