It hasn't always been stated there, and it was a warning on a README somewhere. The server would run happily on 32 bits, and never crash or produce any user-visible errors, it would just silently corrupt data while pretending the insertion went great.

And that's the real issue I would think. If it shut itself down or went into read-only and generated errors it wouldn't be so bad. The limitations 32-bit architectures are why we have 64bit architectures. I think it's fair as a developer to require x64 instead of jumping through hoops to support both architectures. But if you're going to release 32-bit binaries for something like development or testing you should be explicit about the use cases and limitations and avoid failing silently when they're exceeded.

This stems from their choice of "mmap and done" as their entire IO strategy, for earlier versions. The data loss is silly though and shows their attitude.

Oracle 9 on 32 bit can handle 32GiB databases without any issue, and bigger setup to use more that a single file per table space. And Oracle 9 it's old.

Note: I hate Oracle DB, but I must work with Oracle 9/10/11/2 DBs because is what our clients have.