When running compaction, the indexer will create a new file with next revision number (i.e. data.fdb.5 will be compacted into data.fdb.6).
If a restart occurs during compaction, the new indexer is confronted with 2 files. It should pick the lower revision as the higher one is likely incomplete. This is the intend and design of the indexer, however we incorrectly sort the list of files, meaning that .10 will be picked up instead of .9. This also has the side effect of effectively leaking the lower numbered file and consuming disk space.