Migration from Akka 2.5 to 2.6 caused deadlock

Akka
,
1

I used to maintain pretty huge app on top of Play 2.7 (Akka 2.5). Bumping Play version to 2.8.8 caused random deadlocks during startup.
Unfortunately, app does not strictly follows “do not block” imperative; for example, some blocking code in beans’ constructors (i. e. Await.result(dbFuture, Duration.Inf)) takes place. In previous years of support I managed to make it work (including deployments on single-core CPU servers) by using this default-dispatcher configuration:

akka {
  actor {
    default-dispatcher {
      type = Dispatcher
      fork-join-executor {
        parallelism-min = 8
        parallelism-max = 64
        parallelism-factor = 3.0
      }
    }
  }
}

While updating to Play 2.8 and Akka 2.6, I noticed webserver becomes unavailable on 1 or 2-core servers and on dev machines (if parallelism-max is reduced to 8). Thread dump showed application-akka.actor.default-dispatcher locked by each other in tree-like manner, some of them stuck at those Await.result(dbFuture, Duration.Inf).
Increasing parallelism-min solves current problem, but in unstable way.
I also found this note about internal dispatcher, that was introduced in 2.6. If I rollback new behaviour with this line in application.conf:

akka.actor.internal-dispatcher = akka.actor.default-dispatcher

deadlock seems to be gone away.
My question is why this happens. It would be fine if my code may not work while internal and default dispatcher were the same, but this situation is directly opposite.