Sun introduced a filesystem layout back in the 80's with SunOS 2 (I think), where /usr was a largely imutable filesystem.
What this allowed was the /usr filesystem of a system serving diskless clients to share it's own /usr filesystem with the clients.
If anybody cares to remember, the diskless client model meant that Sun 2, 3 and 4 workstations could just be CPU, memory, display and network, with no local persistent storage. Back when SCSI disks were very expensive, this allowed you to centralise the cost in a large server, and keep the cost of the workstations down.
The model was that all filesystems were mounted over NFS, with /, and /var (a new filesystem in this model) mounted (IIRC - myy memory could be faulty and confused by the differences between the Sun and IBM models) from /export/root/clientname and /export/var/clientname on the server as read-write filesystems, and /usr, (and later /usr/share) mounted read-only, served either from the /usr and /usr/share if the clients ran the same architecture and OS level, or from some other location which mirrored /usr if the clients ran a different version (this allowed SPARC architecture systems to be served from Motorola ones, or vice-versa).
Directories such as /etc, /var/adm, /usr/spool, /usr/tmp, which would have been on read-only or read-mostly became symlinks into /var (which was unique to each client as it was mounted from a different directory on the server).
Other vendors including IBM and Digital adopted very similar layouts for clusters of diskless clients. With IBM in 1991, it appeared with AIX 3.2 (and refined in 3.2.5). The filesystem layout meant that no machine should really write into /usr except during an upgrade, containing any variable files into /var. Unfortunately, many people (including IBM software developers) forgot this, and over the years, software expected to be able to write into directories below /usr.
Interestingly, the IBM 9125-F2C, aka Power7 775, supercomputer running AIX reintroduced the concept of diskless clients in 2011. The filesystem layout was modified slightly, with the concept of a statefull read-only NFS filesystem (STNFS), which allowed changes to the read-only filesystem to be either cached in memory for the duration of the OS run (a bit like a filesystem Union), or files/directories to be point-to-point mounted over entities on the read-only filesystem into a read-write filesystem.
/ became a STNFS read only mount, /usr was a read-only filesystem, and /var was a read-write mount off of an NFS server. /tmp was left on the / filesystem, meaning files were lost on a reboot, and also that writing lots of files into /tmp reduced the amount of RAM the node had!
Work related filesystems were mounted over GPFS for performance (NFS was just too slow), although any paging did actually work over NFS (obviously, paging was a major no-no for these performance optimised machines, but we could not get AIX to run without a paging space).
Unfortunately, as I found out, the hot-swap process for adapters, run over RMC from the HMC (Hardware Management Console) had a habit of trying to construct scripts in /usr/adm/ras (on the read-only part of the file tree) to execute to enable the swap, and as a result, we were unable to hot-swap adapters, which caused problems on more than one occasion. I did raise a PMR with support/development, but had trouble arguing the problem through, as the systems were so niche, that the support droids could not understand the problem.