Sliding Sync: Avoid fetching left rooms and add back newly_left rooms

[MadLittleMods]

Performance optimization: We can avoid fetching rooms that the user has left themselves (which could be a significant amount), then only add back rooms that the user has newly_left (left in the token range of an incremental sync). It's a lot faster to fetch less rooms than fetch them all and throw them away in most cases. Since the user only leaves a room (or is state reset out) once in a blue moon, we can avoid a lot of work.

Based on @erikjohnston's branch, develop...erikj/ss_perf

Todo

• Add tests
• Maybe add in SoftDeleteChainMap, Sliding Sync: Avoid fetching left rooms and add back newly_left rooms #17725 (comment)

Pull Request Checklist

• Pull request is based on the develop branch
• Pull request includes a changelog file. The entry should:
  • Be a short description of your change which makes sense to users. "Fixed a bug that prevented receiving messages from other servers." instead of "Moved X method from EventStore to EventWorkerStore.".
  • Use markdown where necessary, mostly for code blocks.
  • End with either a period (.) or an exclamation mark (!).
  • Start with a capital letter.
  • Feel free to credit yourself, by adding a sentence "Contributed by @github_username." or "Contributed by [Your Name]." to the end of the entry.
• Code style is correct
  (run the linters)

[MadLittleMods]

How do we actually expect a room that you are state reset out of to look when it comes down Sliding Sync?

This is just based off of my best guess in the comments before. The main thing is initial: True and required_state that doesn't include your membership to try to indicate to the client that the new state doesn't include you anymore. But there is an edge with that if you're not requesting your own membership event, then you wouldn't expect it to be missing.

This is better than before though so we could punt until the MSC figures that out more 🤷

[erikjohnston]

I think this is the best we can do for now, and we should deal with this question in the MSC. I don't see any problem with compatibility with landing this as it is

[MadLittleMods]

Is it crazy to introduce something like a SoftDeleteChainMap? Where we can insert sentinel values on map[0] whenever we want to "delete" something from the map and avoid copying the immutable room_membership_for_user_map.

The Python docs encourage making your own ChainMap subclass variations like DeepChainMap.

Can also just tackle this in a follow-up PR.

SoftDeleteChainMap first draft

class Sentinel(Enum):
    # defining a sentinel in this way allows mypy to correctly handle the
    # type of a dictionary lookup and subsequent type narrowing.
    UNSET_SENTINEL = object()


class SoftDeleteChainMap(ChainMap):
    """
    Variant of ChainMap that allows deleting keys without affecting the underlying maps (maps[1:]).
    Only adds sentinel values to the top-most mutable map[0].

    aka `DeletableChainMap`, `NullableChainMap`
    """

    def __getitem__(self, key):
        for mapping in self.maps:
            try:
                value = mapping[key]  # can't use 'key in mapping' with defaultdict
                if value is Sentinel.UNSET_SENTINEL:
                    return self.__missing__(
                        key
                    )  # support subclasses that define __missing__

                return value
            except KeyError:
                pass
        return self.__missing__(key)  # support subclasses that define __missing__

    def get(self, key, default=None):
        if key in self:
            value = self[key]
            if value is Sentinel.UNSET_SENTINEL:
                return default

            return value
        else:
            return default

    def __delitem__(self, key):
        self.maps[0] = Sentinel.UNSET_SENTINEL

    def popitem(self):
        raise NotImplementedError("popitem is not supported")

    def pop(self, key, *args):
        """
        Remove specified key and return the corresponding value.

        If the key is not found, return the default if given; otherwise, raise a
        KeyError.
        """
        if key not in self:
            if len(args) == 0:
                raise KeyError(key)
            # Return the default if given
            return args[0]

        value = self[key]

        if value is Sentinel.UNSET_SENTINEL:
            if len(args) == 0:
                raise KeyError(key)
            # Return the default if given
            return args[0]

        del self[key]
        return value

    def __iter__(self):
        # TODO: Fix/implement this one
        d = {}
        for mapping in map(dict.fromkeys, reversed(self.maps)):
            d |= mapping  # reuses stored hash values if possible
        return iter(d)

    def __contains__(self, key):
        return any(
            (key in m and m[key] is not Sentinel.UNSET_SENTINEL) for m in self.maps
        )

[erikjohnston]

Yeah, I think that could make a lot of sense. Annoying that we have to roll our own

[MadLittleMods]

Removed because we already add it just above

[MadLittleMods]

Just renaming to be proper plural

[MadLittleMods]

This if-tree is a bit messy but the test setup is so similar I think I'd rather have this than duplicating the test a bunch.

The extra self.use_new_tables layer will disappear in the future

[erikjohnston]

Woo, thanks!