Changes to HighLevelGraph:

• Order-of-magnitude speedup of __getitem__ in the most common use cases
• various minor speedups thanks to set arithmetics
• type annotations to declare layers, dependencies, and all other internal collections as read-only
• copy() now also copies the memoized data of get_all_dependencies()
• new method to_dict(), which is a memoized wrapper around ensure_dict()
• keys(), items(), and values() are now memoized. The extra RAM usage compared to the previous design should be trivial (I replaced a set with a dict where all the values are already referenced objects)
• keys(), items(), and values() now respect the Mapping protocol and are respectively KeysView, ItemsView and ValuesView instead of lists
• deprecated keyset()
• 50x speedup in keys(), items(), values(), upstream of memoization:

import dask.array as da 
a = da.random.randint(0, 10, size=10_000, chunks=10) 
# Create more layers for HighLevelGraph 
for i in range(10): 
    a = a + 1


def bench(k):
    try:
        del a._to_dict
    except AttributeError:
        pass
    list(getattr(a.dask, k)())

%timeit bench("keys")
%timeit bench("values")  
%timeit bench("items")  

Before:

3.19 ms ± 8.28 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
3.2 ms ± 5.39 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
3.1 ms ± 46.6 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

After:

57.1 µs ± 493 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
58.1 µs ± 152 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
243 µs ± 1.16 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

This iterates upon #4918.