I would like to point out that Python allows only immutable objects as dictionary keys, presumably to prevent this exact problem. Nevertheless, my own preference is also to leave things as they are (no new rules), since Julia's purpose is different from Python's.
Having said this, I would still like to suggest two changes: (1) Explain this potential pitfall in the manual. (I will be happy to write the appropriate text and make a pull request.) (2) Make available a library function "checkcorrectness(Dict)" in case a user is afraid that he/she has clobbered his/her dict by unwisely mutating keys. I could probably write this, but the original author of Dict would do a much better job than me. -- Steve On Monday, October 13, 2014 12:27:35 PM UTC-4, Stefan Karpinski wrote: > > Yeah, it's a bit sketchy, but the alternatives are: a) only allow > immutable values as keys, which seems annoyingly finicky; or b) hash > mutable values by identity, which seems really unintuitive. You can think > of using a value as a dict key, as an implicit promise not to mutate it. > Copying keys before storing them in dicts doesn't really solve the problem > since you can still get a hold of the copy that's stored and mutate that; > this is only slightly harder to do. > > On Mon, Oct 13, 2014 at 12:07 PM, Tamas Papp <[email protected] > <javascript:>> wrote: > >> As a Julia newbie, I see no problem with that. Modifying the keys of a >> hash table is an operation with undefined consequences ("just don't do >> it"), so the programmer should focus on good practices that prevent this >> from happening (it is not that hard). >> >> The alternative is to only allow bits types or containers which are >> immutable recursively (eg an immutable type which contains a mutable >> container is ruled out). In practice, this can be done, bit it would >> lead to a lot of boilerplate in case wants to use other types as keys. >> >> Best, >> >> Tamas >> >> On Mon, Oct 13 2014, nbbb <[email protected] <javascript:>> wrote: >> >> > as a julia outsider (for now), it seems odd to me that mutable >> containers >> > are allowed as dictionary keys at all. >> >> >
