I have been using rpm for almost 2 years now and that's the last time I did
a full system install, using rpm-2.4.something, I think. From then on, I
upgraded my system regularly, upgrading rpm itself several times on the
way. Some rpm versions had to rebuild the database in order to convert it
to a newer format, IIRC version 2.5.x and 2.9x. Having done these, I have
never had problems.

The problem is, however, that erasing a package from the old days now
sometimes gives me a core dump, and I've found out that some packages in
the database are missing the RPMTAG_FILESTATES tag, and that rpm seems to
assume that they exist. Maybe they were forgot when rebuilding and
upgrading the database. I have used the following quick hack to get rid of
the problem of not being able to upgrade these old packages:

--- rpm-3.0.4/lib/transaction.c.fix     Fri Nov 26 21:51:50 1999
+++ rpm-3.0.4/lib/transaction.c Fri Dec  3 12:25:32 1999
@@ -982,7 +982,7 @@
            }
            if (XFA_SKIPPING(fi->actions[i]))
                break;
-           if (fi->fstates[i] != RPMFILE_STATE_NORMAL)
+           if (fi->fstates && fi->fstates[i] != RPMFILE_STATE_NORMAL)
                break;
            if (!(S_ISREG(fi->fmodes[i]) && (fi->fflags[i] &
RPMFILE_CONFIG))) {
                fi->actions[i] = FA_REMOVE;
@@ -1379,8 +1379,9 @@
                                fi->fflags, fi->fc * sizeof(*fi->fflags));
            fi->fmodes = memcpy(xmalloc(fi->fc * sizeof(*fi->fmodes)),
                                fi->fmodes, fi->fc * sizeof(*fi->fmodes));
-           fi->fstates = memcpy(xmalloc(fi->fc * sizeof(*fi->fstates)),
-                               fi->fstates, fi->fc *
sizeof(*fi->fstates));
+           if (fi->fstates)
+             fi->fstates = memcpy(xmalloc(fi->fc * sizeof(*fi->fstates)),
+                                  fi->fstates, fi->fc *
sizeof(*fi->fstates));
            fi->dil = memcpy(xmalloc(fi->fc * sizeof(*fi->dil)),
                                fi->dil, fi->fc * sizeof(*fi->dil));
            headerFree(fi->h);

Is that a dangerous way? And maybe you could double check the rebuild db
feature...