Man do I have experience with this. I work on a project where we originally stored all of our data using XML, then moved to sqlite. There are many pros and cons to each technology, but it was performance that caused the switchover. Here is what we observed.
For small databases (a few meg or smaller), XML was much faster, and easier to deal with. Our data was naturally in a tree format, which made XML much more attractive, and XPATH allowed us to do many queries in one simple line rather than having to walk down an ancestry tree.
We were programming in a Win32 environment, and used the standard Microsoft DOM library. We would load all the data into memory, parse it into a dom tree and search, add, modify on the in memory copy. We would periodically save the data, and needed to rotate copies in case the machine crashed in the middle of a write.
We also needed to build up some "indexes" by hand using C++ tree maps. This, of course would be trivial to do with sql.
Note that the size of the data on the filesystem was a factor of 2-4 smaller than the "in memory" dom tree.
By the time the data got to 10M-100M size, we started to have real problems. Interestingly enough, at all data sizes, XML processing was much faster than sqlite turned out to be (because it was in memory, not on the hard drive)! The problem was actually twofold- first, loadup time really started to get long. We would need to wait a minute or so before the data was in memory and the maps were built. Of course once loaded the program was very fast. The second problem was that all of this memory was tied up all the time. Systems with only a few hundred meg would be unresponsive in other apps even though we ran very fast.
We actually looking into using a filesystem based xml database. There are a couple open sourced versions xml databases, we tried them. I have never tried to use a commercial xml database, so I can't comment on them. Unfortunately, we could never get the xml databases to work well at all. Even the act of populating the database with hundreds of meg of xml took hours.... Perhaps we were using it incorrectly. Another problem was that these databases were pretty heavyweight. They required java and had full client server architecture. We gave up on this idea.
We found sqlite then. It solved our problems, but at a price. When we initially plugged sqlite in, the memory and load time problems were gone. Unfortunately, since all processing was now done on the harddrive, the background processing load went way up. While earlier we never even noticed the CPU load, now the processor usage was way up. We needed to optimize the code, and still needed to keep some data in memory. We also needed to rewrite many simple XPATH queries as complicated multiquery algorithms.
So here is a summary of what we learned.
For tree data, XML is much easier to query and modify using XPATH.
For small datasets (less than 10M), XML blew away sqlite in performance.
For large datasets (greater than 10M-100M), XML load time and memory usage became a big problem, to the point that some computers become unusable.
We couldn't get any opensource xml database to fix the problems associated with large datasets.
SQLITE doesn't have the memory problems of XML dom, but it is generally slower in processing the data (it is on the hard drive, not in memory). (note- sqlite tables can be stored in memory, perhaps this would make it as fast.... We didn't try this because we wanted to get the data out of memory.)
Storing and querying tree data in a table is not enjoyable. However, managing transactions and indexing partially makes up for it.
