feat: move docs map to docroom

[kptdobe]

This is a first attempt to fix #76

@karlpauls and I brainstormed and the current hypothesis is:

• The DocRoom is the Durable Object.
• It contains multiple properties, including the storage.
• It creates ydoc instances which are cached in the docs map.
• This docs object is a global object, not necessarily bound to the Durable Object.
• Depending on how the worker memory is managed, it may happen that this global object still contains references to the DO (especially since the docs are attached to the storage)
• Sometimes, it may happen that the DO instance or the global object are updated / refreshed / recreated / ? and then there is some old reference still around.
• This "might" lead to the error

This is a pure hypothesis and we need to make sure:

1. the code is doing the same thing - all tests still work as before, just needed to tweak some of them
2. deploy in production and monitor to see if this occurs again

In any case, I think we should continue re-working the code: the DocRoom code should be isolated in its own file and, in the ideal scenario, contain all code it need - not sharing anything, to avoid those kind of memory mix. This would allow to clearly visualize what belongs to the DO (super singleton) vs what belong to the worker.

[github-actions]

LCOV of commit 4a1b60c during Install, lint, and test #206

Summary coverage rate:
  lines......: 95.3% (1652 of 1734 lines)
  functions..: 96.7% (58 of 60 functions)
  branches...: no data found

Files changed coverage rate:
                  |Lines       |Functions  |Branches    
  Filename        |Rate     Num|Rate    Num|Rate     Num
  ======================================================
  src/edge.js     | 2.5%    360| 0.0%     8|    -      0
  src/shareddoc.js| 3.5%    569| 0.0%    20|    -      0

[bosschaert]

Looks good to me, thanks for looking into this!

[karlpauls]

I'm not sure this is correct. Wouldn't this basically remove the doc if a connection goes away?

[kptdobe]

On connection close, the doc is removed from the map ("local cache"). This is the same logic as before, just moved the code here to have access to the cache.

[kptdobe]

Ok, what was missing in the new version is the check if there is no more connections. Added.

[karlpauls]

I think this is not doing what happened before. IIRC, the idea was to use the promise to make sure the doc doesn't get bound to persistence called concurrently. @bosschaert, are you sure this is still correct (I would have expected to have this move up to the caller like it used to for the await)...

[bosschaert]

Ah yes, the doc.promise should only be set if it wasn't there yet. Then we await on it in the next line of code. If it resolved already the await will be a no-op.

[kptdobe]

ok, I restored the logic but split the method in 2: createYDoc and setupYDoc. Create is only called once, setup each time a new session starts.
Too bad this is not tested.

[bosschaert]

Looks good to me, but I would also get approval from @karlpauls

[auniverseaway]

Do we want to pull this down to stage given the comments?

[kptdobe]

We could go to stage first but this means someone would test there. I tested it locally as much as I can, basic features work as expected (editing, saving, awareness). What I cannot control is the long run on CF but this will be true also on staging. We can merge to prod and I'll monitor as I do for all changes I am doing on this project anyway.