Use tinyset to optimize in memory cache RAM usage

[james7132]

A good chunk of the overhead of in-memory cache comes from the large number of HashSets used in a lot of it's data structures.

tinyset provides a set type optimized for holding types that boil down to 64-bit integers. For a good number of cases, it will not allocate any heap memory until several elements are inserted (like small/arrayvec), and is only the size of a pointer otherwise.

This PR includes one unsafe function, because tinyset::Fits64 requires one unsafe fn as a part of the trait.

To likewise maximize memory savings, when converting to a u64 for use in Set64, Id<T> shuffles the snowflake parts around to shift most of the 1s towards the lower bits.

As an alternative, a roaring bitmap might be another alternative.

[7596ff]

Could you provide benchmarks at a hundred, ten thousand, or a million guilds? Also, any unsafe code needs to be removed and CI needs to be addressed.

[james7132]

The unsafe fn is required by the trait. The actual implementation is 100% safe. I'll set up a benchmark soon.

[james7132]

Units are in bytes allocated by the system allocator.

guild count  | hashset | tinyset | change
1 | 196817 | 195697 | -0.57%
10 | 224218 | 214534 | -4.32%
100 | 416956 | 343188 | -17.69%
1000  | 1427224 | 983548 | -31.09%
10000 | 13006648 | 8395768 | -35.45%
100000 | 102821832 | 66121672 | -35.69%
1000000 | 1636599912 | 1049397352 | -35.88%

Benchmarking memory usage is a lot harder than it needs to be, orz.

[baptiste0928]

I think this should be behind a feature flag in the model crate to avoid adding unnecessary dependencies for those who don't use the in-memory cache.

[HTGAzureX1212]

This test doesn't seem to test for panic on zero.

Documentation should be updated?

[7596ff]

This hasn't had much movement since the initial discussion in Discord. Ultimately, we will not be using tinyset, and if any changes are to be made here, they should be made in a separate PR.