Preserve visual structure when matching unspecified keys in maps and sets

[timothypratley]

Problem

Maps and sets with unspecified keys are at a syntactic disadvantage.

Example

The current tools for working with unspecified keys are:

(m/rewrite {:a 1, :b 2}
  {& (m/seqable [!k !v] ...)}
  {& ([!k !v] ...)})

(m/rewrite {:a 1, :b 2}
  (m/map-of !k !v)
  (m/map-of !k !v))

Both visually obscure the structure.

Observe that vectors do not suffer this problem: [!k ...]

Maps and sets are disadvantaged because unspecified keys must only appear in the & clause. This restriction makes sense because keys in maps and sets are unordered, so specified keys do not overlap with unspecified keys.

Motivation

In isolation, one may argue that a single map is easy to understand, but in the face of nested maps, the structure quickly becomes obscured. Consider:

{!a {!b {!c !v}}}

(m/map-of !a (m/map-of !b (m/map-of !c !v)))

The latter is linguistic, the former is visual.

Furthermore, if we had correspondence (see https://github.com/noprompt/meander/issues/129), nested maps transformations would preserve their structure completely.

Possible Solution

Make memory variables imply unspecified keys when specified as a key in a map or set.

(m/rewrite {:a 1, :b 2}
  {!k !v}
  {!k !v})

!k !v implies {& (map-of !k !v)).

Why does this rule make sense?

• Unspecified key rules MUST apply to all unspecified keys, and MUST NOT apply to specified keys. Breaking these rules would break conjunctive logic. As such we can comfortably realize that the combination of specified keys as values and unspecified keys as meander concepts has clear semantics in the face of multiple clauses of either.

[timothypratley]

@yuhan0 pointed out (in Slack channel) that {?a 1} is useful, so really only !a expressions could imply unspecified keys. The tradeoff then is the subtlety of different behavior based on the variable type vs the visual syntax

[timothypratley]

Based on Slack discussion, an existing option for picky users like me is:

(defn memory-variable? [v]
  (and (symbol? v)
       (re-matches #"!.+" (name v))))

(def extract-unspecified-kvs
  (s/bottom-up
    (s/rewrite
      {& (m/seqable (m/or [(m/pred memory-variable? !unspecified-ks) !unspecified-vs]
                          !specified-kvs) ...)}
      {& (!specified-kvs ... ['& (`m/map-of !unspecified-ks !unspecified-vs)])}

      ?else
      ?else)))

(defmacro | [& rules]
  `(s/rewrite ~@(map extract-unspecified-kvs rules)))

((| {!k !v}
    {!k (m/app inc !v)})
 {:a 1
  :b 2
  :c 3})

;;=> {:a 2, :b 3, :c 4}

^^ use a macro to replace {!k !v} with (map-of !k !v)

This was offered with big warning signs that I don't fully understand yet :)

[timothypratley]

Another idea floated in Slack: This would be a nice way to be explicit about the behavior without losing the structure: {^… !k !v} and #{^... !k} (i.e. using the metadata style of sets for #{^& ?more}

[timothypratley]

I like this suggestion from @yuhan0 :

"There is already precedence in ..?n for splitting a symbol up into components to bind ?n"

{&... [!k !v]}
;; desugars intuitively to 
{& (m/seqable [!k !v] ...)}
;; or
{&!k !v} ;; implicitly includes !v in the repeating pattern
;; or 
{&!k &!v} ;; over-specification?

#{ &!x }

My favourite is {&!k !v} because it preserves the structure (I can copy-paste example data and replace with variables) and &!k is exactly what I want to do, bind the rest of the kvs as memory variables.