Open li1 opened 5 years ago
Agreed. We won't be able to implement get-else
as a transform. We could either interpret the whole of [(get-else $ e :a v v-else) ?v]
as a special pattern or deviate from the Datomic API and offer something like [?e :a (or ?v false)]
. With the new binding system, we might even be able to do something like [?e :a ?v] (default ?v false)
, where the default
binding would work like a constant binding without any validation.
I like the latter two, with the second one perhaps giving us the most flexibility.
E.g., if we want to calculate a balance, but some users haven't paid / received anything yet so their respective queries currently don't evaluate:
(def balance
'[:find ?person ?balance
:where
(conj/paid ?person ?paid)
(conj/recv ?person ?recv)
[(subtract ?recv ?paid) ?balance]])
With third suggestion:
(def balance
'[:find ?person ?balance
:where
(conj/paid ?person ?paid)
(default ?paid 0)
(conj/recv ?person ?recv)
(default ?recv 0)
[(subtract ?recv ?paid) ?balance]])
Second suggestion looks weird if used on the rules / subqueries (it certainly doesn't look like we're binding 0
to ?paid
):
(def balance
'[:find ?person ?balance
:where
(conj/paid ?person (or ?paid 0))
(conj/recv ?person (or ?recv 0))
[(subtract ?recv ?paid) ?balance]])
...but using it on the subtract
clause makes sense:
(def balance
'[:find ?person ?balance
:where
(conj/paid ?person ?paid)
(conj/recv ?person ?recv)
[(subtract (or ?recv 0) (or ?paid 0)) ?balance]])
Also, we could perhaps even expand the (or ...)
syntax to binding attributes, like this:
(def balance
'[:find ?person ?balance
:where
(conj/paid (or ?person "alfredo") ?paid)
(conj/recv ?person ?recv)
[(subtract (or ?recv 0) (or ?paid 0)) ?balance]])
This would still leave the binding of ?person
for ?recv
, so it means something like "if ?person
hasn't paid anything, use ?paid
from Alfredo, but keep ?person
's receivables."
With suggestion 3 we would sacrifice that granularity and could only speak about ?person
in more general terms (at least without building a more complicated query):
(def balance
'[:find ?person ?balance
:where
(conj/paid ?person ?paid)
(default ?paid 0)
(conj/recv ?person ?recv)
(default ?recv 0)
(default ?person "alfredo")
[(subtract ?recv ?paid) ?balance]])
I prefer suggestion 3, because it is no new syntax, just a new type of binding. The "alfredo" use case seems highly confusing, because it breaks the unification intuition (?person
suddenly doesn't refer to the same entity in all cases).
And another (approximate?) way of expressing that could be:
(conj/paid ?person ?paid)
(conj/recv ?person ?recv)
(conj/paid "alfredo" ?paid-alfredo)
(default ?paid ?paid-alfredo)
Which on the front-end seems like an elegant solution ((default ?paid <constant>)
would simply de-sugar to (default ?paid ?gensym) (constant ?gensym <constant>)
). Not quite sure how the implementation as a PrefixExtender
would look like in that case...
Some more findings, mostly intended as memo.
Default can't be something associated with just a symbol. E.g. (default ?v fallback)
doesn't make sense, because as with negation, defaults need to be w.r.t to a set of entity ids.
I'll try and see how a (default ?e ?v fallback)
feels.
So https://github.com/comnik/declarative-dataflow/commit/1a65b0f98031996e5126d48ae7992d1f3361de19 adds experimental support for something like this.
Indeed it is both simpler and more complex than expected.
More complex, because (for new-query-engine reasons) default bindings are not just dependent on entity ids for context, but also on the attribute that they are acting on.
Simpler, because now [?e :a ?v :else default]
actually becomes a pretty good expression of just that and the implementation fits in nicely with the new query engine (as far as I can tell).
My initial worries were unfounded, because when validating other bindings, this binding will already have a set of entity ids for context.
(Of course this means that we'll have to wait with frontend support, until the new query engine is the default.)
To prevent a situation where queries fail simply due to missing data, it might make sense to introduce default values (similar to
get-else
in datomic).