problem

The threaded termtagger is much faster which leads to some needed modifications to deal with new problems out of that in translate5.

solution

We have to use persistent connections for the threaded termtagger.

Persistent connections are needing pinned DNS entries then (so each worker talks to the same node / same IP) instead of just using the termtagger hostname and landing on different nodes per request.

For pinned DNS we need a fallback if a node is regularly going down.

Open is the adoption for spellchecker, so this code is currently purely in TermTagger Worker.

The described behaviour is only enabled when termtagger threads are > 1 in the translate5 configuration