[ANN] Dictionaries.jl - Improved productivity and performance of dictionaries in Julia

Package Announcements
,
15

I find the design choice that dict[key] = value means strict update interesting… in the sense I am so used to it to mean “upsert” it’s a bit hard to imagine how reading/writing code would be like. But I think I get the analogy to arrays and consistency with other parts of the interface.

Yes this one is quite interesting. I fretted over this one for a while, but I did use Dictionaries.jl for a little while for some “real” work before releasing, and honestly I never missed this. On the other hand, the feeling of certainty that the keys aren’t shifting under you while you are mutating values is quite comforting and very much seemed worth it to me (and set! is a nice partner to get!).

I suppose this is intentionally like this also for broadcasting and mapping?:

Yeah, generic functions delegate to similar and empty as factories to materialize new containers, so anything that preserves keys ends up going through similar. This is quite nice because I found I frequently ended up using fast co-iteration with the input and output, and also I rarely (if ever?) needed to mutate the keys of the output in practice.

Is there a plan to “fix” this (or is this expected)? Maybe using copy-on-write indices? Immutable/persistent data structure?

Yes, they are both valid options. My personal favorite is freeze from #31630 - I’d possibly have a freezekeys or freezeindices thing that you call first in order to get what is the current behavior, and revert similar on mutable keys to making a copy. We should fix this, for sure - but I feel this is the exact same problem with SubArray (mutating the indexer part can cause segfaults, since the bounds check on the indexer values are done at construction time, and the same goes for resizing the indexee) which we should fix first (and using freeze on the indexer and freezekeys on the indexee is only the way to achieve that which I have thought of, while I’m hoping the compiler can realize copying is a not needed in the majority of cases via escape analysis/rust-style borrow checker) . Currently as it stands, consider Dictionaries.jl to be a sharp tool.