The first issue involved copied users which would lead to user updates
causing faster evictions of the cache than was expected.
The second issue involved users that weren't bound to an internal
lifetime eviction policy. These users would not get evicted.
For example, a user without mutual guilds or being part of the internal
cache in general (messages, DMs) would never end up being evicted for
some strange reason. To handle this case, store_user would get a
counterpart named create_user which would create a user without
potentially storing them in the cache. That way only users with a
bound lifetime within the library would be stored.
This changes the lookup from unnecessary O(n) to two amortised O(1)
lookups. This event pretty much always has a guild_id so the original
code was always a performance bottleneck.
The previous protocol based tag type caused significant overhead
(in the magnitude of seconds). Removing this should simplify object
creation by removing typing.Generic from the __mro__
Profiling showed that WeakValueDictionary caused rather significant
and noticeable slowdowns during startup. Since the only thing it was
used for was to automatically remove the key from the mapping when
the reference count reaches zero, the same could theoretically be
accomplished by using the __del__ special method. There is a chance
that this could lead to a memory leak since the __del__ method is not
always called, but the only instances of this happening are during
interpreter shutdown to my knowledge and at that point the mapping
is the least of my concern.
The difference in speed seems negligible at start up, which is when
most time is taken for actually parsing JSON. I could potentially be
missing something but profiling didn't point to any discernable
difference.
