I've been working on some improvements to our large cluster's space balancing, when I noticed that sometimes the OSD maps have strange upmap entries. Here is an example on a clean cluster (PGs are active+clean): { "pgid": "1.1cb7", ... "up": [ 891, 170, 1338 ], "acting": [ 891, 170, 1338 ], ... }, with an upmap entry: pg_upmap_items 1.1cb7 [170,891] this would make the "up" list [ 170, 170, 1338 ], which isn't allowed. So the cluster just seems to ignore this upmap. When I remove the upmap, nothing changes in the PG state, and I can even re-insert it (without any effect). Any ideas why this upmap doesn't simply get rejected/removed? However, if I were to insert an upmap [170, 892], it gets rejected correctly (since 891 and 892 are on the same host - violating crush rules). Any insights would be helpful, Andras