requirement 3.10 (was Re: Minutes of RDF DAWG telecon Tuesday 2004-06-08 for review)

on specifying an upper bound ...

note that federation is just the general case. the simple
case has the same problem, where the simple case is a 3-tier
architecture consisting of an RDF application talking
to an RDF query processor talking to a database store
containing RDF triples.

you suggested a work-around in which the query processor can
pass the LIMIT to the database and then process the results
(filter/infer/whatever). if the result set size is less than
the LIMIT, the query processor should discard the result and
resend the query to the database with a higher LIMIT.

this requires the query processor to completely buffer the
result set before returning any partial result to the
application (to ensure that the result set size reaches the
LIMIT). is this acceptable, i.e., having to buffer the result?
i thought not. it also forces a more complicated architecture
on the query processor.

kevin 



Eric Prud'hommeaux wrote:
> 
> On Thu, Jun 10, 2004 at 07:06:29AM +0900, Eric Prud'hommeaux wrote:
> >
> > In the above, I assume tuples as the unit of measure. Rob raised
> > the issue that (correct me here) the client could be doing some
> 
> oops, sorry, make that Kevin.
> 
> On Wed, Jun 09, 2004 at 04:12:28PM -0700, Kevin Wilkinson wrote:
> >
> > re: requirement 3.10, i'll restate my concern about making
> > the LIMIT an exact bound on the number of "results" returned.
> > i think an exact limit may have a negative impact on performance
> > for a common query processing scenario. i think it should be
> > an upper bound. here's why.
> 
> I fear that this encourages returning 0 results when the results to your
> federated queries didn't happen to line up. Specifics down 100 lines.
> 
> > assumption: in a common scenario, the query processing is distributed
> >    among a client module (providing a query API to an application
> >    program) and one or more servers (data sources), likely accessed
> >    across a network. the client module may do additional processing
> >    over the results from the servers (e.g., for inferencing, union,
> >    whatever) and this may add or remove results.
> >
> > assumption: the primary motivation of the LIMIT clause is to
> >    reduce query latency for queries where the result size is unknown
> >    or known to be undesirably large.
> >
> > to support an exact limit, the client module first decomposes the query
> > and dispatches subqueries to the servers. to retrieve the results,
> > the client module must open a stream connection and keep that stream
> > open until the LIMIT is reached by the client module.
> >
> > the problems with this approach are (1) creating a stream connection
> > is more expensive than returning a single batch of results (2) in the
> > common case of JDBC access to a database, the server is required to
> > materialize the entire result set of the subquery. so, the LIMIT clause
> > cannot be extended to the servers. consequently, the application may see
> > little or no reduction in query latency (which is a primary motivation
> > for the LIMIT clause).
> 
> I believe the spottiness of JDBC setMaxRows(int) implementation comes
> from the lack of LIMIT in SQL92 entry or intermediate specifications.
> JDBC drivers have to use server-dependent hacks to provide this
> functionality.
> 
> This argues *for* putting such a function into DAWG-QL.
> 
> The MySQL JDBC driver could use their "LIMIT x" extension to SQL, but
> it doesn't actually parse the SQL query and therefor doesn't know how
> to stick the "LIMIT x" clause to implement setMaxRows or scrollable
> result sets.
> 
> This argues that even if we put it in the QL, folks might be too lame
> to use it.
> 
> > an alternative approach would allow the client module to pass a LIMIT
> > clause on to the servers. the servers would then process their
> > subqueries
> > with that LIMIT (which could be larger than the original query LIMIT)
> > and return that result as one batch to the client module.
> >
> > the problem with the alternative is that, if the client module does
> > a lot of filtering on the partial results, then the LIMIT is not
> > reached, despite the fact that are more matching results available on
> > the servers.
> >
> > the alternative has much better behavior with respect to system
> > resources and latency because it can pass the LIMIT clause to the
> > servers. since this is the primary motivation for the LIMIT clause
> > (faster response time), i prefer the alternative. in the case that
> > the client module filters out too many results (which, i expect
> > would be rare), i propose that an error condition or indicator
> > be raised to the application which can then increase the LIMIT.
> >
> > as an aside, we chose not to support aggregates. yet, making
> > LIMIT an exact bound makes it a type of aggregate query (it's
> > really doing a count query).
> 
> I'm still fuzzy on aggregate queries. Can nth queries reference
> variables used in the 1st query?
> 
> Suppose I want some people at the New York site with chronic conditions:
> 
> SELECT (who, dept)
> AT <http://example.com/personel>
> FROM (who p:department dept)
>      (who p:site "New York")
> USING p FOR <http://example.com/p-schema>
> 
> [ aggregate with ]
> 
> SELECT (illness)
> AT <http://example.com/medRecords>
> FROM (who m:record x)
>      (x m:chronic illness)
> USING m FOR <http://example.com/m-schema>
> 
> is the second "who" defined by the results of the first query?  It
> seems that, regardless of whether users will know how to use them,
> LIMITs on clause 1, clause 2, or the whole thing could have easily-
> defined semantics. eg. A LIMIT on clause 1 controls how many whos
> we are looking at. A LIMIT on clause 2 controls how many illnesses
> we look FOR EACH who. A LIMIT on the whole thing controls how many
> results we get back at the end of the day.
> 
> There is no way to say how a LIMIT on the whole thing would translate
> to LIMIT 1 without knowing something about the data. Perhaps each
> person is allowed no more than two chronic illnesses, but the very
> nature of being in New Your guarantees at least one illness.
> 
> I think that the complexities of query federation should not prevent
> our protocol/ql from having otherwise critical features. Whether LIMIT
> is critical? Up to the jury.
> 
> SPECIFYING AN UPPER BOUND
> 
> I think the only way a federater can take advantage of the upper bound
> approach is to carry the aggregate LIMIT to each of the sub clauses.
> Suppose the user imposes an (aggregate) LIMIT of 10. We then search
> for 10 employees in New York. We then impose the LIMIT on clause 2.
> We are asking "of the 10 people I happened to bind in the first query,
> give me up to 10 of their illnesses*." If illnesses are scarce (maybe
> 1 in 21? not so good at stats), we are likely to return 0 answers.
> This is not what the person asked for.
> 
> Instead I think we have to do what Kevin said, keep a stream open to
> the first server and keep asking the second server for matches on each
> batch of results from the first server until we reach the limit or the
> end of the solutions from the first server. But this implies a stream,
> and that makes me think of cursors and they are mentioned in the
> charter as being out of scope.
> 
> A stateless way to do this is to passs the aggregate LIMIT to the
> first (and second) server. If, at the end, we get too few results,
> multiply by 10 and retry. Then multiply by 100...
> 
> Federation isn't easy. I don't think we have to write query
> specifications that make it trivial to implement federaters.
> 
> * Names only, please; I don't want to gave the illnesses myself.
> 
> > Eric Prud'hommeaux wrote:
> > >
> > > On Wed, Jun 09, 2004 at 02:07:21PM +0100, Seaborne, Andy wrote:
> > > >
> > > > > > -- 3.10 Result Limits
> > > > > >
> > > > > > straw poll: who's convinced this is a requirements? are we close
> > > > > > to a decision here?
> > > > > AndyS: RDQL does not have a limit. Joseki provides limits.
> > > > > PatH: it could mean 2 things: I don't want more than 10 answers
> > > > >   and the next ones after that or *just* 10
> > > >
> > > > It occurred to me that one way of differentiating between exactly 10 and
> > > > more than 10 (and an unknown number) when we have "LIMIT 10" is to ask for
> > > > 11.  If 11 results return up, it is more than 10.  Otherwise, it is exact
> > > > and 10 or less.  This is making the (trailing) flag we discussed a matter of
> > > > whether the 11th slot gets filled at all.
> > > >
> > > > Hence, no requirement for a special flag and the client can decide whether
> > > > to ask for 11 if it needes to know the difference.  It still might be a
> > > > better design overall to have a flag.
> > >
> > > This came up during the telecon. I think the outcome was that it
> > > depended on what extra bit you had in the respones. It could be
> > > a boolean say there were more results; it could be a count of
> > > those results:
> > >
> > > query   response        server  network
> > > limit   extra bit       effort  burden
> > > -----   ---------       ------  ------
> > > none    none            full    full
> > > 10      none            light1  light
> > > 10      boolean         light2  light
> > > 11      none            light2  light
> > > 10      count           full    light
> > >
> > > where "light" meant the effort was shortcutted by the limit.
> > > "light2" means the server had to compute one more tuple (or
> > > the existance of) than "light1".
> > >
> > > In the above, I assume tuples as the unit of measure. Rob raised
> > > the issue that (correct me here) the client could be doing some
> > > inferencing which would give more [or fewer, i note] results than
> > > specified in the limit. I think that's fine, that it does't reduce
> > > the utility of specifying a limit.
> > >
> > > > > AndyS: seen this to prevent error conditions
> > > > > EricP: SQL has it
> > > > > RobS: there are other things in SQL that we don't have
> > > > > Steve: can affect latency
> > > > > AndyS: can be either a query or a protocol issue
> > > > > EricP: arguing from the point of view of a straightforward
> > > > >   implementation (for both client and server)
> > > > > Yoshio: but then we need a way to specify the snapshot of the database
> > > > > RobS: what is understood as a "result"
> > > > > Kevin: do we want an exact upper bound? could cause more impl burden on
> > > > > the server
> > > > > DanC: does "upper bound" appeal to somebody?
> > > > > Kevin: yes
> > > > > FarrukhNajmi: +1
> > > > > PatH: DQL could, by request, give a count of results
> > > > > DanC: motivation to spend time on this is in discussing/bounderaries
> > > > > with
> > > > > other groups
> > >
> > > --
> > > -eric
> > >
> > > office: +1.617.258.5741 NE43-344, MIT, Cambridge, MA 02144 USA
> > > cell:   +1.857.222.5741
> > >
> > > (eric@w3.org)
> > > Feel free to forward this message to any list for any purpose other than
> > > email address distribution.
> 
> --
> -eric
> 
> office: +1.617.258.5741 NE43-344, MIT, Cambridge, MA 02144 USA
> cell:   +1.857.222.5741
> 
> (eric@w3.org)
> Feel free to forward this message to any list for any purpose other than
> email address distribution.

Received on Thursday, 10 June 2004 19:34:55 UTC