The Berkeley DB transactional data store offers all the data integrity promises that you'd expect from a world-class database system. But every rose has its thorn, and so we must note some known limitations of Berkeley DB.
Berkeley DB environments are not portable. You cannot simply copy a Subversion repository that was created on a Unix system onto a Windows system and expect it to work. While much of the Berkeley DB database format is architecture-independent, other aspects of the environment are not.
Second, Subversion requires the use of Berkeley DB in a way that will not operate on Windows 95/98 systems—if you need to house a BDB-backed repository on a Windows machine, stick with Windows 2000 or later.
While Berkeley DB promises to behave correctly on network shares that meet a particular set of specifications, most networked filesystem types and appliances do not actually meet those requirements. And in no case can you allow a BDB-backed repository that resides on a network share to be accessed by multiple clients of that share at once (which quite often is the whole point of having the repository live on a network share in the first place).
If you attempt to use Berkeley DB on a noncompliant remote filesystem, the results are unpredictable—you may see mysterious errors right away, or it may be months before you discover that your repository database is subtly corrupted. You should strongly consider using the FSFS data store for repositories that need to live on a network share.
Because Berkeley DB is a library linked directly into Subversion, it's more sensitive to interruptions than a typical relational database system. Most SQL systems, for example, have a dedicated server process that mediates all access to tables. If a program accessing the database crashes for some reason, the database daemon notices the lost connection and cleans up any mess left behind. And because the database daemon is the only process accessing the tables, applications don't need to worry about permission conflicts.
These things are not the case with Berkeley DB, however. Subversion (and programs using Subversion libraries) access the database tables directly, which means that a program crash can leave the database in a temporarily inconsistent, inaccessible state. When this happens, an administrator needs to ask Berkeley DB to restore to a checkpoint, which is a bit of an annoyance. Other things can cause a repository to “wedge” besides crashed processes, such as programs conflicting over ownership and permissions on the database files.
Berkeley DB 4.4 brings (to Subversion 1.4 and later) the ability for Subversion to automatically and transparently recover Berkeley DB environments in need of such recovery. When a Subversion process attaches to a repository's Berkeley DB environment, it uses some process accounting mechanisms to detect any unclean disconnections by previous processes, performs any necessary recovery, and then continues on as though nothing happened. This doesn't completely eliminate instances of repository wedging, but it does drastically reduce the amount of human interaction required to recover from them.
 Berkeley DB requires that the underlying filesystem implement strict POSIX locking semantics, and more importantly, the ability to map files directly into process memory.