Once web content began to be personalized, that was no longer realistic. Databases became a required part of the stack, web servers ran applications instead of just serving file, we gained massive amounts of functionality, and haven't really looked back.
If all you want to do is run a small project, or a big one where performance matters and you have deadlines, you might just "do it by hand" and simply create your own pseudo-database out of flat files, memory dumps, whatever.
The approach described is what people nowadays would call memcache, but the entire architecture is different because client connections are persistent rather than resetting every page load like the web. They just built a standard game server keeping everything in memory in the regular data structures. Fork-and-persist is a nifty way of using the operating system's copy-on-write functionality to checkpoint state.
This model apparently allowed for quite a few dupe bugs, one of which I published on an old fan site. It was quite involved, in fact I wrote it up so deliberately convoluted that what you were doing was obvious. Did it work, I wasn't quite sure but I did suffer both duplication of items and loss from their system through no effort on my part.
Its kind of funny to learn how they did it this late, reminds me of mechanisms we employed with some early multiplayer door games on BBSes. Supporting multiple players for some early titles was a bear, the least of which was the lack of stability of most connections just plain bad code that bit you in the butt more often than not.
You can get repository level transactions by committing on each consistent state (squashing before push) this lets you do undo/redo and recover from failed operations easily - databases sort of give you this - but only for the stuff they actually contain (so unless you're putting your asset files in to the DB you don't get this).
"Server-boundary crossing edge-cases and race conditions persisted for a long time -- allowing for gold and items to be duplicated, though I think we had largely eradicated the big ones by the time I had moved on to UO2."
you can get a long way with join, awk & grep
Somewhat off topic, but one the pieces of code I'm most proud of was effectively a recursive XML scraper that resolved each leaf (each node was a web request) into a line (like a file path) of a massive text file, which was then grepped when a user searched for something. I hacked it together really quickly to solve a problem, which is solved orders of magnitude faster than commercial solutions that tried to be overly clever. I like it because while it was gross and bad and would be embarrassing to read, in reality grep is crazy fast and the OS cached the text file once you'd read it once. It's a nice personal reminder to me that you don't always need to build some Grand Design, especially when you're just building a tool that has no future evolving path.
Freeshards usally went with a stop-the-world-worldsave. Which means they were infrequent and easily corrupted, leading to "timewarps".
"As Raph also notes, there were no databases originally involved in the storage of game state or player data for UO (disregarding analytics here), everything was kept in flat files. Backups worked by flagging a moment in time where no one was allowed to cross server-boundaries -- during that moment, each areaserver was commanded to fork(), essentially duplicating itself in memory (it's more complicated than this, thanks to Copy-on-Write, but let's simplify). After everyone had fork()ed, the "lock" preventing boundary-crossing was cleared. Then each areaserv began to dump out its huge chunk of memory-state into a file on an NFS server. Those files were then all tarred together and kept as a "backup" of the state of the server. These heavyweight backups happened at half-hour intervals, I believe."