Key differences with continuous queries for non-MVCC caches:
1) When event is received, subsequent read of the same key may still return previous value for some time (i.e. no guarantee of happens-before between event and subsequent read)
2) Keys for notification are kept in memory. If transaction is too large, OOME could happen. In order to avoid that, we skip notification of too large transactions.
Roman Kondakov, could you please elaborate on how is this implemented?