- From: Steve Harris <steve.harris@garlik.com>
- Date: Mon, 21 Dec 2009 23:56:57 +0000
- To: Andy Seaborne <andy.seaborne@talis.com>
- Cc: SPARQL Working Group <public-rdf-dawg@w3.org>
On 21 Dec 2009, at 18:07, Andy Seaborne wrote: > Steve and I were discussing aggregates on IRC: this is for the > record because we didn't have time to complete the discussion there > and then. > > What we both want is that > > sum(1, unbound) -> 1 > > sum(1, 2/0) -> error > > i.e skip unbounds in the group but return error on bad expressions. > > COALESCE does not help because it treats both as the same. And does > not work for min and max which are multi-datatype. > > The key text is: > > [[ > Definition: Aggregation > > Any unbound values exp · μ' do not appear in the resulting value > multiset > ]] > > which (my reading) that says evaluate exp on μ' then see if it's > unbound. > > But the single variable case sum(?x) has an expression of ?x so exp > · μ' is called on unbound ?x (which is an error) then the test is > made to see if it's unbound (it's not - it's now error). > > We could special-case evaluation of aggregates. If it's a single > variable, then return value or unbound. If it's any other > expression (including ?x+0), then evaluate as normal, i.e. error. > There are otherways for sum (e.g. the value of unbound is always > zero) but is using the special feature that zero is the identity > value for + so does not work for other aggregates. Another option would be to say that if there are any variables appearing in exp (other than as an argument to COALESCE) that are unbound in μ', the result in unbound. It's unfortunate for aggregates that eval(?x) for unbound ?x is an error. - Steve
Received on Monday, 21 December 2009 23:57:29 UTC