Welcome! Please see the About page for a little more info on how this works.

0 votes
in Clojure by
h2. Problem

To do runtime coercion, specs need to be walked twice to strip away the branching information: {{s/conform}} + {{s/unform}}. This introduced extra latency (see the sample below).

h2. Proposal

New versatile {{s/walk*}} to support generic spec walking.

h2. Current status

* {{s/valid?}} allows us to do quickly check if a value conforms to a spec
** [https://dev.clojure.org/jira/browse/CLJ-2115] could help to return the errors in a single sweep.
* for coercion, we do {{s/conform}} + {{s/unform}} / {{s/explain}}
** [https://dev.clojure.org/jira/browse/CLJ-2116] would separate conforming from specs

Still, when running {{s/conform}} + {{s/unform}}, we walk the specs twice - which is performance-wise suboptimal. Below is a sample, with Late 2013 MacBook Pro with 2,5 GHz i7, with JVM running as {{-server}}.


(require '[clojure.spec.alpha :as s])

(s/def ::id int?)
(s/def ::name string?)
(s/def ::languages (s/coll-of #{:clj :cljs} :into #{}))
(s/def ::street string?)
(s/def ::zip string?)
(s/def ::number int?)

(s/def ::address (s/keys
                   :req-un [::street ::zip ::number]))

(s/def ::user (s/keys
                :req [::id]
                :req-un [::name ::address]
                :opt-un [::languages]))

(def value {::id 1
            :name "Liisa"
            :languages #{:clj :cljs}
            :address {:street "Hämeenkatu"
                      :number 24
                      :zip "33200"}})

; 2.0 µs
(cc/quick-bench
  (s/conform ::user value))

; 6.2 µs
(cc/quick-bench
  (s/unform ::user (s/conform ::user value)))


Despite {{s/conform}} is relatively fast, we triple the latency in the sample when running also {{s/unform}}. As we know already that we are not interested in the branching info, we could just not emit those.

h2. Suggestion

{{s/walk*}} to replace both {{s/confrom*}} and {{s/unform*}}, maybe even {{s/explain*}}. It would take extra {{mode}} argument, which would be a Keyword of one of the following:

* {{:validate}} - return false on first failing spec
* {{:conform}} - like the current {{s/conform*}}, maybe also return {{s/explain}} results?
* {{:unform}} - like the current {{s/unform*}}
* {{:coerce}} - {{s/conform*}} + {{s/unform*}}, could be optimized (e.g. if no branching info, just return the value)

The public apis could be remain the same (+ optional extra argument with CLJ-2116), and a new {{s/coerce}} to call the {{s/walk*}} with {{:coerce}}.

h2. Results

Single sweep validation & coercion. Happy runtime.

5 Answers

0 votes
by

Comment made by: ikitommi

Renamed the issue. Instead of Keyword argument, it should take a function to walk the spec to support arbitrary walking applications.

0 votes
by

Comment made by: marco.m

hello, any news ?

0 votes
by

Comment made by: alexmiller

No plans to look at this before the next batch of implementation changes, so it will be a while.

0 votes
by

Comment made by: ikitommi

There are three different versions for ~this in spec-tools now:

1) spec-walker: using s/form, walks the specs and values and calls a callback function with both, returning a new value. It can be used for things like coercion. It doesn't validate, need to call s/valid? and s/explain-data separately (2-3 calls in total)

2) spec-visitor: just walks the forms and return new forms. Used in things like spec => json-schema transformations. Not transforming values.

3) wrapped specs with overridden s/conform**, returning new value or error. It's the saddest of the three, because everything needs to be wrapped, but still the only one that knows how to walk over properly regex specs. Bundled transform + validate, like with Schema.

...

https://cljdoc.org/d/metosin/spec-tools/0.9.0/doc/spec-coercion has examples of the first/coercion, hopefully highlighting why support for coercion is important.

Is this issue on a (near) roadmap? Can we help with it? If not on the roadmap, is there a suggested way to do runtime value transformations with spec?

0 votes
by
Reference: https://clojure.atlassian.net/browse/CLJ-2251 (reported by ikitommi)
...