The new service, let's call it content cache, should work similar to the existing parser cache:
* semi-permanent
* support splitting on things like target language
* get automatically purged when page content changes
In addition, support
* more than one object per page (per MCR slot, but also other things like graphoid data or Wikibase constraint checks)
* multiple "stages" per page, e.g. "current" and "stable". This is presently hacked in by the FlaggedRev extension.
* multiple "targets" per object, e.g. per language, but also annotated vs. folly resolved parsoid output, just pre-processed output, etc.
* be accessible from outside MediaWiki code, so standalone service can use it.
This is currently just a draft, which would probably become an RFC in time.