W3C home > Mailing lists > Public > public-rdf-dawg-comments@w3.org > May 2011

Re: Pagination in SPARQL OFFSET and LIMIT needs ORDER BY

From: Jerven Bolleman <jerven.bolleman@isb-sib.ch>
Date: Fri, 13 May 2011 10:01:22 +0200
Message-ID: <4DCCE552.7000305@isb-sib.ch>
To: public-rdf-dawg-comments@w3.org
Dear workgroup,

I realized that I might not have been so clear in describing the problem.

Assume that you maintain a publicly available SPARQL endpoint.
You want to support both a HTML view and the official SPARQL formats.

Lets say a user executes the query
SELECT * WHERE {?s ?p ?o}
This will download every triple in your store. In my store this will 
mean trying to download 160gb of triples via a single HTTP connection.
This is not likely to work and if it did most browser will crash on the 
HTML view.

Therefore I would like to always put a LIMIT on the query to make sure 
that the result will match the capabilities of a common HTTP connection.
e.g. default LIMIT 1000

But I do want people to download more than just the first 1000 results 
to their query. I just want them to do it in multiple requests that are 
likely to complete and not crash their browsers.

So I need pagination i.e. OFFSET. In practical terms this does exactly 
what I need (having briefly tested OWLIM and Virtuoso).
i.e. page 1 SELECT * WHERE {?s ?p ?o} OFFSET 0 LIMIT 1000
      page 2 SELECT * WHERE {?s ?p ?o} OFFSET 1000 LIMIT 1000
Until there are no more results. However, this is not specified to work 
in the current public draft.

Having the following 2 triples in a store.
<_:1> <lala> "hi"
<_:1> <lala> "by"

The following query
SELECT * WHERE {?s ?p ?o}
Can evaluate to either a)
<_:1> <lala> "hi"
<_:1> <lala> "by"
or b)
<_:1> <lala> "by"
<_:1> <lala> "hi"

i.e. ordering is random but all results are returned.

The following query, assume the implementation always returns ordering a)


Can return
<_:1> <lala> "hi"
And in the same store it is valid to return this for
As well.

So while the chunks are small I am not guaranteed to get all valid 
results. I need to add an ORDER BY clause. However, I can't without 
changing the query as you can not add ORDER BY *. Nor is this always 
desired because ORDER BY actually means that you need to ORDER the 
results. This can be very expensive relative to executing the query.

Therefore, I would define OFFSET more specifically.

When a implementation returns a result set for a query. Then it should 
do so in a deterministic manner. i.e. executing the same query twice on 
a store with constant data will return results in the same order.
The OFFSET parameter is then interpreted as discard the first X results 
that a the same query without OFFSET would have generated.

This means that for a query A with N results. The concatenation results 
of queries A OFFSET 0..N LIMIT 1 is equal to the result of the query A.

Jerven Bolleman

P.S. the original source of this discussion is.

On 05/12/2011 04:32 PM, Jerven Bolleman wrote:
> Dear workgroup,
> I was recently made aware that there is no easy way to get a guaranteed working pagination.
> i.e. QUERY OFFSET 0 LIMIT 5 page 1
>        QUERY OFFSET 5 LIMIT 5 page 2
>        QUERY OFFSET 10 LIMIT 5 page 3
> Without adding an ORDER BY clause. Adding any kind of ORDER BY clause would be enough to ensure pagination worked. I would therefore like to see an  ORDER BY * or ORDER BY ANY option. To ensure that the results come in some implementation specific order and that this can be used to show all possible results.
> Trying a few public current SPARQL implementations. With ORDER BY * showed that this is currently not implemented. Although pagination with OFFSET and LIMIT without an ORDER BY clause  seems to work as a naive user (e.g. me) would expect. Meaning that for current SPARQL implementers it is no work at all other than dealing with a slightly different SPARQL grammar.
> Pagination guaranteed to succeed would then be
>        QUERY OFFSET 10 LIMIT 5 ORDER BY ANY page 3
> The other option is to expand the description of the OFFSET clause. For example the use of the OFFSET clause should guarantee that query results come back in a consistent order.
> I hope this concern makes sense.
> Regards,
> Jerven

Received on Friday, 13 May 2011 08:02:01 GMT

This archive was generated by hypermail 2.2.0+W3C-0.50 : Friday, 13 May 2011 08:02:02 GMT