Server crashes mean so much more than lost data; it’s lost income, loss of reputation and lost time spent cleaning up the mess. Availability is crucial. But not all availabilities are equal. Should a loss of a data store mean goods can’t be sold?
Dan Pritchard argues that in the absence of data, replies to requests are predictable and robust. Queue data updates, delay all data access, and for goodness sakes test the damn thing. Pritchard goes into enough detail, but don’t expect a HOWTO. Those steps alone are worth a book.