Limit the cache based on arbitrary measures

[rphmeier]

Motivation: My use-case for lru-cache involves inspecting items' heap size. Rather than hack it in, I think it would be beneficial to make the cache adaptable to whatever limit needs to be imposed.

You might have a set of heterogeneous items, each using a different amount of allocation on the heap. Maybe a more general Meter trait might help:

trait Meter<T> {
    fn measure(&self, item: &T) -> usize;
}

The default Meter could be a Count:

struct Count;

impl<T> Meter<T> for Count { 
    fn measure(&self, _: &T) -> usize { 1 }
}

Here's an implementation that would use the HeapSizeOf trait (I'd recommend this one for inclusion under a feature gate as it's generally useful.)

struct HeapSize;

impl<T: HeapSizeOf> Meter<T> for HeapSize {
    fn measure(&self, item: &T) -> usize { item.heap_size_of_children() + ::std::mem::size_of::<T>() }
}

When you add or remove an item, you add or subtract meter.measure(&item) from the maximum respectively. If the metered total exceeds the maximum while adding an item, other items must be removed in a loop until the total falls below the limit.

Caveats:

• get_mut or interior mutability could allow items to be altered such that their measure at insertion time differs from their measure at removal. We could add a UniformMeter<T> which takes no reference to individual items when measuring -- indicating that the measurement is unrelated to item size. get_mut might only be implemented when the meter is uniform, while fn get(&mut self, key: &Q) -> Option<&T> would serve as an alternative implemented for all caches. This doesn't solve the interior mutability issue, but that can be addressed the same way as in HashMap: documentation specifying that it's a logic error to modify the value such that its measure would be altered.
• insert can remove multiple items -- how do we return each of them to the caller efficiently? (maybe an iterator?)

Alternatively:

• check the measure before and after get_mut and adjust the total correspondingly. The issue here is that it would need to be done through a CacheHandle<'a, T: 'a> which contains that logic in its destructor. This would break backwards compatibility, and would need to handle returning displaced items if the size grows.
• re-measure every item in the cache whenever necessary. this seems prohibitively expensive

Looking forward to your comments and naming suggestions below.

[apasel422]

This is a neat idea. I need to ponder this further, but I think I'm in favor of going the HashMap route and just documenting that changing an item's measure after it's been inserted is a logic error, potentially also providing a method that can be used to update the measure in a controlled way once it's in the map.

[rphmeier]

@apasel422 The situation is a bit different, though, since the measure can be changed through plain old mutability as opposed to just interior mutability like HashMap keys.

I wonder if Meter should be Meter<K, V> and measurement should be done based on both key and value sizes (String as key type comes to mind)

[apasel422]

I do like the idea of a CacheHandle (modulo naming?), even though it's possible to mem::forget it. A recent change to BinaryHeap is similar and declared leaking OK, since it doesn't affect memory safety. As for backwards compatibility, we could just have a semver bump, or declare a new method on LruCache<K, V, M> that returns CacheHandle instead of &mut V and leave get_mut only on LruCache<K, V> (assuming it defaults to Count).

I was also thinking it could be Meter<K, V>.

Is there a way to avoid storing a duplicate usize with Count, since it's derivable from the underlying LinkedHashMap's len()?

[luser]

I have a patch for this mostly done, with the exception of the get_mut bits.