The reason is performance. Reading and writing from RAM is orders of magnitude faster than from disk.
If a clock producer node is down, it will be replaced by another block producer according to the votes.
If you're referring to a witness node for a specific dApp, then you wouldn't want a single point of failure. A serious dApp which relies on a witness node will probably have several copies in separate data centers, using bare metal or a cloud data center such as AWS or Azure or others.
In that witness node, you can certainly use a plugin to save the state in a conventional DB and resync faster in case it is needed.
Block.one is adding support for faster resync and external DB in version 1.1.0 of the EOS node software. See details here