Repo GC strategy.

[jbenet]

@whyrusleeping @wking @tv42

Currently we explicitly mark + store indirect pins. @tv42 brought up that we perhaps should mark-and-sweep. the discussion on https://github.com/ipfs/go-ipfs/pull/1274 puts me strongly in favor of that approach. let's use this note to coordinate discussion + potential upgrades to do.

[jbenet]

• gc must be safe (don't ever gc objects the user explicitly asked to keep)
• @wking mentioned how git uses a time-based grace period to make gc safer.

[tv42]

I've pondered this a bunch for https://bazil.org/ , though I'm not at the point where I'd have written anything yet. Quick notes on what differentiates CAS GC from normal:

• immutable objects
• you can memoize all referred objects of a root with bloom filters etc
• can view storage as append-only; links can only point backward in time
• if you rely on time-based grace for new objects, linking to an object must refresh that timer

Good background material to read:

• Erlang GC
• papers on the Foundation project
• the Train algorithm

There's tons more there, but I've barely scratched the surface myself.