The pkgset phase takes around 35 minutes in current composes.
Around 20 minutes of that is spent creating these per-arch
subsets of the global package set. In a rather roundabout way
(see #1794 ), I figured out that almost all of this time is
spent in this cache check, which is broken for a subtle reason.
Python's `in` keyword works by first attempting to call the
container's magic `__contains__` method. If the container does
not implement `__contains__`, it falls back to iteration - it
tries to iterate over the container until it either hits what
it's looking for, or runs out. (If the container implements
neither, you get an error).
The FileCache instance's `file_cache` is a plain Python dict.
dicts have a very efficient `__contains__` implementation, so
doing `foo in (somedict)` is basically always very fast no matter
how huge the dict is. FileCache itself, though, implements
`__iter__` by returning an iterator over the `file_cache` dict's
keys, but it does *not* implement `__contains__`. So when we do
`foo in self.file_cache`, Python has to iterate over every key
in the dict until it hits foo or runs out. This is massively
slower than `foo in self.file_cache.file_cache`, which uses the
efficient `__contains__` method.
Because these package sets are so huge, and we're looping over
*one* huge set and checking each package from it against the cache
of another, increasingly huge, set, this effect becomes massive.
To make it even worse, I ran a few tests where I added a debug log
if we ever hit the cache, and it looks like we never actually do -
so every check has to iterate through the entire dict.
We could probably remove this entirely, but changing it to check
the dict instead of the FileCache instance makes it just about as
fast as taking it out, so I figured let's go with that in case
there's some unusual scenario in which the cache does work here.
Signed-off-by: Adam Williamson <awilliam@redhat.com>
c8fe99b1aa