The RWStore recycles allocation slots. When a mutation rule is executed, it must ensure that the read-behind point for that rule is protected from the release of allocated records until the rule is finished. This is handled for scale-out using a read-only transaction. A similar integration needs to be performed for the RW Store.
This integration will also require the RWStore to potentially defer the release of allocation slots associated with deleted records. That translates into a requirement to "log" the deleted addresses for eventual release, e.g., using an index over delete blocks containing addresses to be released and associating them with commit times at which they may be released based on the history retention based on the minReleaseAge of the transaction service (or possibly the commit times at which they were allocated).
This problem shows up when running the RDFS closure of a non-trivial data set such as LUBM U50. If you are using the triple store without inference, such as BSBM, then this problem does not show up.