Re: Last Call for comments on "SPARQL Query Language for RDF"

Dear all,

below my review on the current SPARQL draft from

http://www.w3.org/TR/rdf-sparql-query/


on behalf of W3C member organization DERI Galway.

Generally, I think the formal definitions have improved a lot, but still 
I am at the same time not 100% sure that all definitions are formally 
water-proof. This affects mainly questions on Section 12 and partly 
unclear Definitions/pseudocode algorithms for query evaluation therein.

HTH,
best,
Axel


-------

Detailed comments:


Prefix notation is still not aligned with Turtle. Why?
Would it make sense to align with turtle and
use/allow '@prefix' instead/additionally to 'PREFIX'
You also have two ways of writing variables... so, why not?


Section 4.1.1

The single quote seems to be missing after the table in sec 4.1.1
in "", or is this '"'?

Section 4.1.4

The form

[ :p "v" ] .

looks very awkward to me!

I don't find the grammar snippet for ANON very helpful here, without
explanation what WS is...  shouldn't that be a PropertyListNotEmpty instead?


Section 5

Section 5 is called Graph patterns and has only subsections
5.1 and 5.2 for basic and group patterns, whereas the other types are
devoted separate top level sections.. this structuring seems a bit
unlogical.


Why the restriction that a blank node label can only be used in a single
basic graph pattern? And if so, isn't the remark that the scope is the
enclosing basic graph pattern redundant?

Why here the section about "extending basic graph pattern matching",
when not even basic graph pattern matching has been properly introduced
yet? If you want to only informally introduce about what matching you
talk here, then I'd call section 5.1.2 simply "Basic Graph Pattern
Matching" but I think I'd rather suggest to drop this section.



"with one solution requiring no bindings for variables"
-->
rather:
"with one solution producing no bindings for variables"
or:
"with one solution that does not bind any variables"

Section 5.2.3

Why you have a separate subsection examples here? It seems 
superfluous/repetitive. Just put the last example, which seems to be the 
only new one, inside Sec 5.2.1 where it seems to fit, and drop the two 
redundant ones. For the first one, you
could add "and thatbasic pattern consists of two triple patterns" to the
first example in sec 5.2, for the second one, add the remark that "the
FILTER does notbreak the basic graph pattern into two basic graph
patterns" to the respective exaple in section 5.2.2.



Section 6:

One overall question which I didn't sort out completely so far:
What if I mix OPTIONAL with FILTERs?

ie.

{A OPTIONAL B FILTER F OPTIONAL C}

is that:

{{A OPTIONAL B} FILTER F OPTIONAL C}

or rather

{{A OPTIONAL B FILTER F} OPTIONAL C}

and: would it make a difference? I assume no, the filter is, in both
cases at the level of A, but I am not 100% sure. Maybe such an example 
owuld be nice to have...


Another one about FILTERs: What about this one, ie. a FILTER which
refers to the outside scope:

?x p o OPTIONAL { FILTER (?x != s) }

concrete example:

SELECT ?n ?m
{ ?x a foaf:Person .  ?x foaf:name ?n .
   OPTIONAL { ?x foaf:mbox ?m FILTER (?n != "John Doe") }  }

Supresses the email address for John Doe in the output!
Note: This one is interesting, since the OPTIONAL part may NOT be 
evaluated separately!, but carries over a binding from the super-pattern!

Do you have such an example in the testsuite? It seem that the last 
example in Seciton 12.2.2 goes in this direction, more on that later

Would it make sense to add some non-well-defined OPTIONAL patterns,
following [Perez et al. 2006] in the document? As mentioned before, I
didn't yet check section 12, maybe these corner case examples are there..


Section 7:

Why "unlike an OPTIONAL pattern"? This is comparing apples with pears...
I don't see the motivation for this comparison, I would suggest to
delete the part "unlike an OPTIONAL pattern".


as described in Querying the Dataset
-->
as described in Section 8.3 "Querying the Dataset"


Section 8

The example in section 8.2.3 uses GRAPH although GRAPH hasn't been
explained yet, either remove this section, start section 8.3 before, I 
think GRAPH should be introduced before giving an example using it.

<you may ignore this comment>
BTW: Would be cool to have a feature creating a merge from named graphs
as well...

ie. I can't have something like
GRAPH g1
GRAPH g2 { P }

where the merge of g1 and g2 is taken for evaluating P.
whereas I can do this at the top level by several FROM clauses.
(Note this is rather a wish-list comment than a problem with the current 
spec, probably, might be difficult to define in combination with 
variables...)
</you may ignore this comment>

Section 8.2.3 makes more sense after the 8.3 examples, and 8.3.2 is
simpler than 8.3.1, so, I'd suggest the order of subsections in 8.3

8.3.2

8.3.1

8.3.3

8.2.3

8.3.4 (note that this example somewhat overlaps with what is shown in
8.2.3 already, but fine to have both, i guess.)



Section 9:

What is "reduced" good for? I personally would tend to make reduced the
default, and instead put a modifier "STRICT" or "WITHDUPLICATES" which 
enforces that ALL non-unique solutions are displayed.

"Offset: control where the solutions start from in the overall solution
sequence."

maybe it would be nice to add: "[...] in the overall solution sequence, 
i.e., offset takes precedence over DISTINCT and REDUCED"

at least, the formulation  "in the overall solution sequence" would
suggest this... however, right afterwards you say:
"modifiers are applied in the order given by the list above"... this
seems somehow contradicting the "in the overall solution sequence", so
then you should modify this to:
"in the overall solution sequence, after application of solution
modidiers with higher precedence" and give an explicit precedence to
each solution modifier....

<you may ignore this comment>
BTW: Why is precendence of solution modifiers not simply the oRder in
which they are given in a query? wouldn't that be the simplest thing to do?

ie.

OFFSET 3
DISTINCT

would be different than

DISTINCT
OFFSET 3

depending on the order.
Anyway, if you want to (which you probably do) stick with what you have
now, it would at least be easier to read if you'd take the suggestion 
with explicit precedence levels for each modifier.
</you may ignore this comment>


Section 9.1

The ORDER BY construct allows arbitrary constraints/expressions as 
parameter...ie. you could give an arbitrary constraint condition here,
right? What is the order of that? TRUE > FALSE? Would be good to add a 
remark on that.

  I would put 'ASCENDING' and 'DESCENDING' in normal font, since it 
looks like keaywords here, but since the respective keywords are ASC and 
DESC.

Stupid Question: What is the "codepoint representation"? ... Since more 
people might be stupid, maybe a reference is in order.


What is a "fixed, arbitrary order"??? Why not simply change

"SPARQL provides a fixed, arbitrary order"
-->
"SPARQL fixes an order"

and

"This arbitrary order"
-->
"This order"

I'd also move the sentence starting with "This order" after the enumeration.


Note that, in the grammar for OrderCondition I think you could write it 
maybe shorter:

Wouldn't simply
  orderCondition ::= ( 'ASC' | 'DESC' )? (Constraint | Var)
do?

  In the paragrpah above the Grammar snippet, you forgot the ASK result 
form where ORDER BY  also doesn't play a role, correct?

Sec 9.2:

Add somewhere in the prose: "using the SELECT result form"...

It is actually a bit weird that you mix select into the solution 
modifiers, IMO, it would be better to mention SELECT first in section 9 
and then introducing the solution modifiers.

Sec 9.3:

REDUCED also allows duplicates, or no? you mention before that reduced 
only *permits* elimination of *some* duplicates... so, delete the "or 
REDUCED" in the first sentence.


Sec9.4:
As for reduced as mentioned earlier, my personal feeling is that 
REDUCED, or even DISTINCT should be the default, since it is less 
committing, and I'd on the contrary put an alternative keyword "STRICT" 
or "WITHDUPLICATES" which has the semantics that really ALL solutions 
with ALL duplicates are given. My personal feeling is that
aggregates, which you mention in the "Warning" box, anyway only make 
sense in connection with DISTINCT. Or you should include a good example 
where not...

Sec 9.5/9.6:

OFFSET 0 has no effect, LIMIT 0 obviously makes no sense since the 
answer is always the empty solution set... So why for both not simply 
only allowing positive integers? I see no benefit in allowing 0 at all.

Section 10:

"query form" or "result form"? I'd suggest to use one of both consistently
and not switch.  Personally, I'd prefer "result form"...

Section 10.1

As for the overall structure, it might make sense to have the whole 
section 10 before 9, since modifiers are anyway only important for 
SELECT, and then you could skip the part on projection in section 9, as 
SELECT is anyway not a solution modifier but a result form...
You should call it also "projection" in section 10.1, ie. what I suggest 
is basically merging section 10.1 and 9.2.


Section 10.2

CONSTRUCT combines triples "by set union"?
So, I need to eliminate duplicate triples if I want to implement
CONSTRUCT in my SPARQL engine?
Is this really what you wanted? In case of doubt, I'd suggest to
remove "by set union", or respectively, analogously to SELECT,
introduce a DISTINCT (or alternatively a WITHDUPLICATES)
modifier for CONSTRUCT...

BTW, I miss the semantics for CONSTRUCT given formally in Section 12.


Section 10.2.1

<you may ignore this comment>
What if I want a single blank node connecting all solutions? That would 
be possible, if I could nest constructs in the FROM part...
</you may ignore this comment>


Section 10.2.3

  Hmm, now you use order by, whereas you state before in Section 9.1 
that ORDER BY has no effect on CONSTRUCT... ah, I see, in combination 
with LIMIT!
  So, would it make sense in order to emphasize what you mean,  to 
change in section
9.1

"Used in combination"
-->
"However, note that used in combination"

10.3/10.4

I think that ASK should be mentioned before the informative DESCRIBE, 
thus I suggest to swap these two sections.

Section 11

- Any changes in the FILTER handling from the last version? Is there a 
changelog?
- As mentioned earlier, I am a bit puzzled about the "evaluation" of 
Constraints given as an argument to ORDER BY especially since there you 
don't want to take the EBV but the actual value to order the solutions.
(Note that what it means that a solution sequence "satisfies an order 
condition" is also not really formally defined in Section 12!)

Apart from that, did not check the section in all detail again since it 
seems to be similar to the prev. version , but some comments still:

"equivilence"?
Do you mean equivalence? My dictionary doesn't know that word.

The codepoint reference should already be given earlier, as mentioned above.


Section 11.3.1

The operator extensibility  makes me a bit worried as for the 
nonmonotonic behavior of  '! bound':
  In combination with '! bound', does it still hold that
"SPARQL extensions" will produce at least the same solutions as an 
unextended implementations and may for some queries, produce more 
solutions... I have an unease feeling here, though not substantiated by 
proof/counterexample.


Section 12 :

12.1.1


Is the necessity that the u_i's are distinct in the dataset really 
important?
Why not also define the data corresponding to the respective URI as 
graph merge then, like the default graph?


12.2

The two tables suggests there is a corellation between the patterns and 
modifiers appearing in the same line of the table, which is not the case.

Also, why in the first table is RDF Terms and triple patterns in one 
line and not separate?

Why do you write
    FILTER(Expression)
but not
   ORDER BY (Expression)
as the syntax suggests?

Moreover, the tables should be numbered.

You use the abbreviation BGP for Basic graph pattern first in the second 
table which wasn't introduced. Actually, it would be more intuitive, if 
you'd use actually *symbols* for your algebra, like e.g. the ones from 
traditional Relational Algebra, as was done in [Perez et al. 2006].

"The result of converting such an abstract syntax tree is a SPARQL query 
that uses these symbols in the SPARQL algebra:"
-->
"The result of converting such an abstract syntax tree is a SPARQL query 
that uses the following  symbols in the SPARQL algebra:"
or maybe even better:
"The result of converting such an abstract syntax tree is a SPARQL query 
that uses the symbols introduced in Table 2 in the SPARQL algebra:"

What is "ToList"?

12.2.1

The steps here  refer to the grammar?
The steps obviously take the parse tree nodes of the grammar as the basis...
anyway this is neither explained nor entirely clear.

then connected with 'UNION'
-->
connected with 'UNION'

What do you mean by

"We introduce the following symbols:"

1) what you define here is not 'symbols'
2) This doesn't seem to be a proper definition but just a bullet
   list without further explanation.

as said before, the symbols, should indeed be symbols and be defined 
properly in section 12.2 with the tables, in my opinion.

The algorithm for the transformation is a bit confusing, IMO. It seems 
to be pseudo-code for a recursive algorithm, but it is not clear where 
there are recursive calls.

Is the observation correct that in this algebra (following the algorithm)

     A OPTIONAL {B FILTER F}

would be the same as

    A  FILTER F OPTIONAL {B}

???

ie, both result in:

  LeftJoin(A,B,F)

That is not necessarily intuitive in my opinion.
Take the concrete exampe from above:

SELECT ?n ?m
{ ?x a foaf:Person .  ?x foaf:name ?n .
   OPTIONAL { ?x foaf:mbox ?m FILTER (?n != "John Doe") }  }

As I said, in my understanding, this query could be used to supress
email addresses for a particular name, whereas the algorithm suggests
that this is just the same as writing:

SELECT ?n ?m
{ ?x a foaf:Person .  ?x foaf:name ?n . FILTER (?n != "John Doe")
   OPTIONAL { ?x foaf:mbox ?m  }  }

Is this intended? If yes, the last example of section 12.2.2 is wrong.

BTW: If so, it seems that the whole second part of the algorithm can be 
simplified to:
-- 
If F is not empty:
   If G = LeftJoin(A1, A2, true) then
         G := LeftJoin(A1, A2, F)
   Else
         G := Filter(F, G)
-- 
where, as I said, the first branch puzzles me a bit... and actually,
it seems to be contradicted by the last example in seciton 12.2.2!


12.2.3

Why do you need ToList?


Projection: You only mention SELECT here.. shouldn't you write here

"If the query is a SELCECT query"
??

"length defaults to (size(M)-start)."

"size(M)" isn't defined anywhere.

It would be probably more elegant to interpret 0 as parameter for LIMIT
as ALL, since you can't know the size of the solution set upfront ... As 
you mention above 'LIMIT 0' doesn't really make sense anyway.

In the definition of compatible mappings, you might want to change

"every variable v in dom(&mu;1) and in dom(&mu;2)"
to
"every variable v &isin;  dom(&mu;1) &cap; dom(&mu;2)"

"Write merge(&mu;1, &mu;2) for &mu;1 set-union &mu;2"

Why not use the symbol &cup; here?


12.3.1

"A Pattern Instance Mapping, P, is the combination of an RDF instance 
mapping and solution mapping. P(x) = &mu;(&sigma;(x))"

Should this be:

"A Pattern Instance Mapping, P, is the combination of an RDF instance 
mapping &mu; and solution mapping &sigma;: P(x) = &mu;(&sigma;(x))"

What is x here? I assume you want P to be defined as a mapping from
RDF-T cup V to  RDF-T?
&sigma; (instance mappings) are defined for graphs, not for variables!
Something seems strange to me here.


12.3.2

You use the terms answer and answer set several times in that section 
which haven't been defined... You should either do so, or refer to 
solution, solution set, as defined.

12.4

Filter:
"a effective boolean"
->
"an effective boolean"



Move the explaining sentence:

"It is possible that a solution mapping μ in a Join can arise in 
different solution mappings, μ1and μ2 in the multisets being joined. The 
cardinality of  μ is the sum of the cardinalities from all possibilities."

before the definition of Join


Note: the semantics of OrderBy seems to suggest, that any 
(non-deterministically chosen?) sequence which satisfies the order 
condition, is valid... correct?


Definition of Project:
- What is i in [i]???
- The use of V is ambiguous here, since in the initial defs this was the 
set of all possible SPARQL query variables.
- The use of P is ambiguous here ,since P was used before to define a 
pattern instance mapping in Sec 12.3.1 ... BTW: it would help a lot if 
Definitions were numbered!



"The order of Distinct(Ψ) must preserve any ordering given by OrderBy."

hmmm,
you mean:  "The order of Distinct(Ψ) must preserve any ordering given by 
any nested OrderBy."?
That is a bit weird, since the order by's have been resolved previously, 
right?

I think the problem is with this notation:

"Write [x | C] for a sequence of elements where C(x) is true."

because this imposes a condition on the element and not on the whole 
sequence.


12.5

The operator List(P) is nowhere defined.
I still don't have totally clear why you need to introduce the ToList 
operator.


A general comment:

I miss a section defining the *Semantics of a query* and of different 
result forms.
The Evaluation semantics given here rather is a mix of functions having 
partly multisets of solution mappings and sequences thereof as result, 
but all are called "eval()".
  E.g. eval for BGP returns a multiset, whereas eval returns a list for 
ToList, etc.

The semantics of a *query* is not really clearly defined yet, it seems.
This needs another revision, I guess.

12.6

In this section again, the terms answer set and answers are used for 
solutions.
As mentioned above, I guess this needs to be introduced to be clear.

In the "Notes", item (d):

"the current state of the art in OWL-DL querying focusses on the case 
where answer bindings to blank nodes are prohibited."

It would be helpful to give references here.

"The same blank node label may not be used in two separate basic graph 
patterns with a single query."

Isn't this restricting? I see no good motivation for this restriction, 
if the group patterns refer to the same graph, ie are no graph patterns, 
to be honest. Anyway, you can remark that variables shall be used 
instead, where one would feel that
such overlapping blank nodes would be necessary, right?


Note that, in the context of bnodes however, I have a problem with this one:

"(c) These conditions do not impose the SPARQL requirement that SG share 
no blank nodes with AG or BGP. In particular, it allows SG to actually 
be AG. This allows query protocols in which blank node identifiers 
retain their meaning between the query and the source document, or 
across multiple queries. Such protocols are not supported by the current 
SPARQL protocol specification, however."

Note that this seems to be a bit worrying to me. It seems to suggest 
that extensions of SPARQL allow to treat BNodes different to existential 
variables, which is what would become possible if you allow them to 
retain their meaning over stacked queries. I am a bit worried, if this 
"backdoor" is really compatible with the intention of bnodes in RDF.


</review>

-- 
Dr. Axel Polleres
email: axel@polleres.net  url: http://www.polleres.net/





Received on Tuesday, 17 April 2007 16:54:15 UTC