What's the expected behaviour when using vector as dict(hash) key?

General Usage
1 
a,k=Dict(),[1]
a[k]=1
k[1]=2
print(a," ",a[k]," ");print(a[[1]]," ");println(a[[2]]);
a[k]=2
print(a," ",a[k]," ");print(a[[1]]," ");println(a[[2]]);

What should we expect? Or should we completely forbids using vectors(arrays) as key? Then why does hash for array relies on value rather than objectid?

2

Arrays compare equality by their contents, thus they hash by their (current) contents. If you change the contents of an array that’s being used as a dictionary key, you’ll no longer be able to find it in the dictionary. This is the expected behavior, and I think we have it documented… if not let’s add it as a warning to the documentation.

2 Likes
3

However when using Dict in callback functions, and of course we can’t require the caller to ensure it doesn’t reuse any arrays, disaster comes about. Is this still by design?

4

Assuming that the keys are immutable is pretty common, I think. You could submit a PR to Base that adds an

index = isimmutable(index) ? index : copy(index)

to base/dict.jl setindex! methods and see what people think. Or create a MutableKeyDicts package.

2 Likes
5

I agree. From personal perspect I’d prefer adding MutableKeyDicts, and forbids mutable keys in Dict. Hope no one has opened a similar issue.

6

If you want to hash based on object-id, then you can use the built-in IdDict type:

d = IdDict()
k = [23]
d[k] = "yes"
k2 = [23]
d[k2] # throws exception
get(d, k2, "no") # returns "no"
k[1] = 17
d[k] # returns "yes"

This is probably the behavior you want if you are using arrays as keys (and is also much faster than hashing the array contents).

9 Likes