Behavior with Clojure 1.6.0:
Cause: This is caused by expectations in clojure.lang.RT regarding the type of collections for some methods, e.g. contains() and getFrom(). Checking for contains looks to see if the instance passed in is Associative (a subinterface of PersistentCollection), or IPersistentSet.
Approach: Expand the types that RT.getFrom(), RT.contains(), and RT.find() can handle to cover the additional transient interfaces.
Alternative: Other older patches (prob best exemplified by clj-700-8.diff) restructure the collections type hierarchy. That is a much bigger change than the one taken here but is perhaps a better long-term path. That patch refactors several of the Clojure interfaces so that logic abstract from the issue of immutability is pulled out to a general interface (e.g. ISet, IAssociative), but preserves the contract specified (e.g. Associatives only return Associatives when calling assoc()). With more general interfaces in place the contains() and getFrom() methods were then altered to conditionally use the general interfaces which are agnostic of persistence vs. transience.
Screening Notes re -10 patch
the extra conditions in RT add branches to some key functions. get already has a getFrom optimization, but there is no similar containsFrom or findFrom. Is it worth measuring the possible impact of these?
I believe the interface refactoring approach (not taken here) is worth separate consideration as an enhancement. If this is done, I think leveraging valAt would be simpler, e.g. allowing HashMap and ArrayMap to share code
it is not evident (to me anyway) why some API fns consume ILookup and others do not, among e.g. contains?, get, and find. Possible doc enhancement?
there is test code already in place (data_structures.clj) that could easily be expanded to cover transients. It would be nice to do this, or better yet get some test.check tests in place
Patch: clj-700-14.patch is based on clj-700-10, plus implements Rich's suggestion that entryAt and containsKey should live in ITransientAssociative2 and the existing concrete impls (ATransientMap, TransientVector) now implement it.
If we are not going to get to the type hierarchy change any time soon, and if perf is acceptable (looks like it should be), I think we should consider applying the "RT knows about transients" approach in the current patch.
Added -10 which sorts transient cases to the end of the checks.
Does this solve as well?
Re -11 additions, by adding a method to ITransientAssociative, you have broken all external implementers (which includes several contrib library data structures and from glancing at github, many others). Maybe should make ITransientAssociative2 with the added method instead?
Erik, re - no, that's a separate issue.