Hello Frank,
It is a known open issue where the broker gives you wrong stats for messages queued. As you can see (with vmq-admin session command, for example), you have no offline messages queued.
The counters are valid for the lifecycle of a VerneMQ boot to boot. What we count here is every message that enters a queue, and every message that leaves a queue. Now, quite obviously, in some specific situations, we either count too many in's, or too few out's.
So the most likely reason is that we missed an edge case scenario (re-send, inter-node delivery...) that makes the counters go off.
We will eventually fix this, as just has been done for the connection counters. I'm aware it is annoying.
Best regards,
André