It varies between clusters. External facing workloads - I try not to overcommit at all memory wise. With Cross campus (multi site) clusters try to keep < 50 percent so that a failover event will not impact external facing workloads, internal cross campus clusters > 50 % < 75% so a failover would result in approximately 0-50% overcommitment.
Internal Production level development. clusters (Things that developers and testers use daily to do their work but are not used by external clients or for internal users servicing clients) up to 40-50% overcommittment.
Internal production - 0-20% though sometimes we go higher (One cluster approached 40% and we saw intermittent performance issues (DB servers don't like vswap, neither do jvm instances).
You can implement shares to prioritize workloads until you get more capacity.
Two things that vary between environments -
1) Hhow dynamic is the environment? How often do you add or subtract workloads and what size of new deployments are "typical"?
2) How fast can you procure and deploy new capacity? Are you using EVC mode in case new capacity has newer cpu features that could result in incompatibilities?
Those answers will also help you figure out how much if any overcommittment is appropriate for your environment.