Reconsider our recommendations for heap size?

[jpountz]

Our documentation states "Set Xmx and Xms to no more than 50% of your physical RAM" but many on-prem deployments and Cloud set it to exactly 50%. Given how we've increasingly moved memory usage off-heap (#52405 in particular), we are less and less relying on the heap but also increasingly relying on the filesystem cache. Should we update our recommendations? If yes, what should be our guidance?

[elasticmachine]

Pinging @elastic/es-docs (>docs)

[elasticmachine]

Pinging @elastic/es-core-infra (:Core/Infra/Core)

[rjernst]

Just adding a note here that this issue is partially moot now in that the documentation quote above no longer exists as we now have auto heap sizing. However, for data nodes, that is still 50%, so we should probably look at what that should be going forward. I'm tempted to close this issue though, or perhaps relabel it for future performance testing work, since our recommendation has essentially changed to "let us figure out heap for you", which indirectly answers the question raised by this issue. @jpountz thoughts?

[mosche]

I'm tempted to close this issue though

Still tempted to close @rjernst ?