Sadly we can't do it yet, the improvements colliding with our usage of creating the index with mappings.
But: My initial problem is solved.
Just for information (and i cant show it), since we use dynamic_templates there was a real big mapping at the end (didnt expect that), the mapping got a size of 194MB at the crash, now its back to 10KB, since we changed some implementation detail.
Since this will show only on the indexing phase, not with the index creation, we couldnt see it before. For real we only saw it because i stopped the master and the errors with "mapping not found" were in there then. If i didnt stop it, i would still searching for a solution.
So @jprante, the reason why we could create such heavy lot is, that we really used massive dynamic_templates, and the cluster couldnt keep up.
Thanks again @all for all you suggestions, improvements and comments.