I had a similar problem. Jobs seemed to stay in pending and get processed very arbitrarily, dashboard seemed off as well.

It seems horizon needed to be using the same redis connection as the queue.

We had split domains a bit, so default was a connection that used redis DB 0, Session used redis DB 5 and Cache used Redis DB 10.

At first we put Horizon on default via it's 'use' => 'default' config and we had the queue on it's own connection using DB 2.

Once we put them both on default connection using DB 0 it worked as expected.