https://bugs.openldap.org/show_bug.cgi?id=9316
--- Comment #1 from Howard Chu hyc@openldap.org --- The results are pretty much expected; the design of LMDB's freelist management in LMDB 0.9 isn't suited to workloads like this. With large records of mixed sizes, fragmentation in the freelist will kill write performance over time. We've experimented with several improved designs for LMDB 1.0 but nothing has panned out yet.