> > > > Commands line options: rd.debug rd.udev.debug systems.log_level=debug That willl be incredibly verbose, and slows things down a lot, so in the off chance there's a race, you might get different results. But if not, the log should contain something useful. I like the hypothesis about mdadm metadata version 0.9, however that's still really common on RHEL and CentOS. It was used for /boot prior to about Fedora around Fedora 24 maybe? It could also be a dracut bug, since that's what's largely responsible for assembly. But then it can be confused by a change in udev rules. :-D However, going back to mdadm 0.9 metadata, that version should only be kernel auto detected. Theoretically, it gets activated before dracut gets involved. 1.x versions have no kernel autodetect, instead it happens in dracut (by calling mdadm to assemble and run it). Oh, is this really dm-cache, not lvmcache? That might be a source of confusion, if there isn't lvm metadata present to hint at LVM for proper assembly. Of course, lvmcache still uses device mapper, but with LVM metadata. Anyway is quite an interesting, and concerning problem. Chris Murphy