I'm also curious if there is a way to give feedback to users without breaking existing installations. Maybe @s1monw has a clue?

@mikemccand to fix this the right way I agree we should be loud. I think we have to upgrade the settings in the clusterstate first and that can on work well if we do it only in a major version? On the other hand what if you have machines with 8 CPUs and others with 80 CPU? I mean this is really a per-node setting and I think we should maybe not allow to set this globally? I think you patch is good as a first step but we should really think about how to expose these settings?

I talked to @s1monw ... we decided to add a logger.warn here if you tried to set to a too-large value, but otherwise be silent for now, and to open a separate issue that this setting should not be global and dynamically updatable.

OK I pushed another commit w/ the feedback; I think it's ready.

LGTM

Thanks @s1monw, I'll port to 2.2.0 as well.