Critical Description
In the documentation for the product (and general sizing advice) we tell people to allocate no more than 80% of their memory for the Server Quota, to leave headroom for the views, disk write queues and general OS usage.
However on larger[1] nodes we don't appear to enforce this, and instead allow people to allocate up to 1GB less than the total RAM.
This is inconsistent, as we document and tell people one thing and let them do another.
This appears to be something inherited from MB-2762, which the intent of which appeared to only allow the relaxing of this when joining a cluster, however this doesn't appear to be how it works - I can successfully change the existing cluster quota from the CLI to a "large" value:
$ /opt/couchbase/bin/couchbase-cli cluster-edit -c localhost:8091 -u Administrator -p dynam1te --cluster-ramsize=127872
ERROR: unable to init localhost (400) Bad Request
[u'The RAM Quota value is too large. Quota must be between 256 MB and 127871 MB (memory size minus 1024 MB).']
While I can see some logic to relax the 80% constraint on big machines, with the advent of 2.X features 1024MB seems far too small an amount of headroom.
Suggestions to resolve:
A) Revert to a straightforward 80% max, with a --force option or similar to allow specific customers to go higher if they know what they are doing
B) Leave current behaviour, but document it.
B) Increase minimum headroom to something more reasonable for 2.X, and document the beaviour.
([1] On a machine with 128,895MB of RAM I get the "total-1024" behaviour, on a 1GB VM I get 80%. I didn't check in the code what the cutoff for 80% / total-1024 is).
