Looking at apr's 1.3.x branch:
In r817810 a change to the hash table implementation was introduced, which moves the data of a hash table from the main pool to a subpool (as seen from the pool passed to apr_hash_make). This makes the contents of the hashtable unavailable from pool cleanup handlers registered on that same pool. (subpools are cleared before the cleanup handlers of the pool itself)
For Subversion this is a breaking change as we commonly use pool cleanup handlers which read hashtables. (E.g. we close the open data files by iterating a hash in the cleanup handler)
Switching to this apr version (once released) will make our code use pointers to already freed memory.
A valid alternative would be to allocate a subpool in the parent pool of the pool passed to apr_hash_make, and clearing/destroying that pool via a cleanup handler once the passed pool is cleared. (Or just keeping the old behavior of realocating in the same pool).