[spec] generator for `any?` occasionally generates `Double/NaN` for which equality semantics don't apply, and that is a problem for the :ret spec of many functions.


The generator for `any?` will occasionally give back Double/NaN value(s). Since NaNs & equality (via `=`) don't get along, :ret spec'ing a fn which transforms/processes a collection according to a predicate, becomes rather problematic. That's because the most obvious thing to check under :ret (the case where the predicate didn't return true for any value, and so the output coll should be equal to the input coll because nothing was transformed/processed), cannot be expressed trivially.

The workaround I've come up with in my own specs is to spec the elements of the collection with `(s/and any? (complement double-NaN?))` instead of just `any?`, and it works. However, even though I can live without NaNs in the tests, I must admit it still feels sort of hacky.


1) The generator for `any?` could be hardcoded to never return Double/NaN. Sounds rather invasive.
2) The generator for `any?` could be reworked to somehow be configurable wrt allowing/prohibiting Double/NaNs. Then perhaps a dynamic-var and/or a macro (e.g. `without-NaNs`) could expose this (just brainstorming here).
3) The generator for `any? could stay as is, but a new equality operator could be added (e.g. `clojure.spec/===`), which somehow ignores NaNs (a naive implementation for instance might walk the data-structures and replace all NaNs with keywords, and only then perform a regular comparison).


Ubuntu 16.10 - Oracle Java 8


Alex Miller
November 8, 2016, 4:29 PM

Should consider whether this change is more appropriate in test.check or in the spec generator for any?.

Dimitrios Jim Piliouras
November 11, 2016, 6:31 PM

It turns out that my workaround does not fully work. I literally just stumbled in the following case:

{nil {[] {NaN 0}}}

which is a conforming value for:

(s/def :ersistent-map
(s/map-of ::anything-but-NaN ::anything-but-NaN)) ;; (s/and any? (complement double-NaN?))

So basically, the inner collections can still have NaNs. So far I've got 4 specs that I've written and faced this problem on all of them.

Wes Morgan
March 6, 2018, 9:37 PM

exclude_NaN_from_any_and_some_generators.patch is my attempt at addressing this by preventing NaNs from appearing in any? and some? generated values. It seemed to solve the problem in my testing.




Dimitrios Jim Piliouras






Affects versions