lossless paging (was Re: stable paging)

On 02/19/2014 05:49 PM, Steve Speicher wrote:
>
>
>
> On Sun, Feb 16, 2014 at 2:49 PM, Sandro Hawke <sandro@w3.org 
> <mailto:sandro@w3.org>> wrote:
>
>     Sorry I didn't propose this months ago.   It's taken me some time
>     to see this solution clearly.   Thanks to Eric and Andrei for
>     being sounding boards.   They haven't seen this particular
>     framing, though.
>
>     This is my proposal for "stable paging", a way to prevent the
>     "lossy paging" problem in the currect spec and highlighted in
>     RFC-5005.  I believe it can be implemented efficiently, and isn't
>     prohibitively complicated.  I'd like to see it in the LC draft,
>     marked "at risk" in case it turns out to be harder to implement
>     than I think.
>
>     *static paging*
>
>     First, let me define something simpler, which I'll call "static
>     paging".  With static paging, the server embeds the etag of the
>     paged resource into the paging URLs. When it responds to the
>     paging URLs, it checks to make sure the etag is still the same. 
>     If it's not, then the server responds with a 410 GONE.  Now the
>     client knows it needs to re-initiate paging.
>
>     With static paging, the client gets a perfect view of the
>     resource, with no loss, no corruption.  And it's trivial to
>     implement.  The problem is that resources that are large enough to
>     require paging are likely to change frequently, so a client would
>     be unable to follow PREV/NEXT links very many hops before the
>     resource changed and the client had to start again.  It might be
>     it could never get more than 10 pages in before it had to start
>     again.  That doesn't seem acceptable.
>
>     *stable paging definition*
>
>     With "stable paging" I propose a looser requirement.  We allow the
>     graph to change during paging, but each triple which could
>     theoretically ever be in the graph is assigned to a particular
>     page. This results in certain guarantees for the client. 
>     Specifically, if the client traverses all the pages with stable
>     paging, (1) every triple it sees must have been in the graph at
>     some point during the traversal, and (2) every triple it does not
>     see must have been absent from the graph at some point during the
>     traversal.  In other words, the error is limited to exactly those
>     triples that changed during the traversal.  (With lossy paging,
>     the client may miss triples that were always in the graph.)
>
>
> To be very clear, what you mean by "exactly those triples that 
> changed" meaning only those triples whose object position has changed 
> (subject and predicate have not changed).  So it could be a delete + 
> insert of a triple with same subject and predicate is also a "triple 
> that changed" and not a "delete, then an insert", correct?  I guess 
> you could argue there is no difference from the client's perspective.

In all of this, I'm thinking of general RDF graphs, so I'm not treating 
any position in the triple as special.

At the point paging starts, the paged resource is an RDF graph 
containing the set of triples G0.

As time goes on, and the client slowly flips through the pages, the 
paged resource might happen to be updated.  After the first change, it 
contains the set of triples G1.  Then, after another change, it contains 
the set of triples G2.   Etc, up to Gn, when the client does its last 
page GET.    (Of course, the server doesn't know this is the last one, 
unless it decides to cut the client off and start replying 410 GONE.)

Let's define union(Gi, Gj, ...) and intersection(Gi, Gj, ...) to be the 
set union and set intersection, respectively, of the set of triples in 
each given graph.

Now, we can say: to do lossless paging, the server MUST construct the 
pages such that the union of all the pages is a superset of 
intersection(G0, ... Gn) and a subset of union (G0, ... Gn).

In other words, the client will see every triple that is there the whole 
time the traversal is happening, and it MAY see triples that are there 
part of the time the traversal is happening.

Again, I think this is fairly intuitive behavior -- if triples are added 
while you're reading, it shouldn't surprise you that you might ore might 
not see them (depending exactly when it happened, and where you were in 
your reading).

This isn't nice as static paging, where the client simply sees G0, but 
it's much cheaper.   Compared to static paging, lossless paging can 
produce inconsistent views of the data, as some triples of a structure 
are seen but not others.   With paging containers, though, since each 
triple is independent, that's not a problem.

Contrast this with lossy paging, as ATOM, where if triples are added 
while you're reading, COMPLETELY UNRELATED TRIPLES might be omitted from 
what you see.    The guarantee there is just that what you'll see is a 
subset of union (G0, ... Gn).   But it could be any random subset.

BTW, instead of "single-page resource", I suggest the term "subset page" 
or "subset page resource" or "subset resource".   To me a "single-page 
resource" sounds like a resource that doesn't happen to be paged.

>
>     *directory example *
>
>     Let's consider the common case: we're paging an LDPC which has a
>     sort order.  (In fact, I expect this will be the only case where
>     paging will be implemented by most LDP servers.  When else would
>     it ever actually be useful?)  Maybe it's a directory of employees,
>     sorted by the foaf:name of each person.  The FIRST page has the
>     folks at the beginning of the alphabet; the LAST page has the
>     folks at the end.
>
>     It's a big directory, and people are being added and removed all
>     the time.
>
>     With lossy paging, and a server that naively just puts 100 people
>     per page, if a name were deleted during forward paging or added
>     during reverse paging, the client would simply, silently miss a
>     person. Displayed to a user this might be okay, but if the client
>     is the payroll system, or an acl system, or a spam filter lookup,
>     or an emergency alert system, missing even a single person is
>     inexcusable. Static paging might be nice, but isn't practical. 
>     But stable paging is pretty good.  With stable paging we might
>     still miss a person, but only if that person were added or deleted
>     during paging.  That seems like perfectly reasonable behavior,
>     about what people would expect.
>
> I guess I'm confused by a couple of the statements here, namely:
> "With stable paging we might still miss a person, but only if that 
> person were added or deleted during paging"
>
> Isn't that the same argument you used why lossy paging is bad?
> "if a name were deleted during forward paging or added during reverse 
> paging, the client would simply, silently miss a person."
>
> Perhaps your point was more on the silent failure part (more on that 
> below).  I believe you meant with stable paging it will not miss a 
> person but I wanted to make sure.
>

Hopefully I made that clear above.

>     So, how does stable paging work?
>
>     Instead of just putting 100 names per page, with stable paging the
>     server creates a function mapping names to pages independent of
>     what else is in the graph.  It might be the first page has names
>     starting with A, the second has names starting with B, etc.  Or it
>     might be the first is "" through "Anderson", the second is
>     "Anderson" through "Aubrey", etc.  The paging function can be
>     created by looking at the database, but once a page has been
>     served to a client and certain triples included/excluded from it
>     because of some boundary test, that boundary MUST NOT change.
>
>     Each paging URL then includes an identifier for that paging
>     function. As with static paging, the server can, at any time, give
>     up on a particular paging function and answer 410 GONE for those
>     URLs.  With the directory, I imagine one would only do that when
>     the directory changes size significantly.  I can also imagine one
>     would have two paging functions, one for big pages and one for
>     small pages, and the server would pick one based on a client
>     PREFER header.
>
>
> So after the ("Anderson" < page <= "Aubrey") is served to the client, 
> say "Arthur" is inserted.  My client (perhaps after waiting for some 
> period) requests ("Aubrey" < page <= "Brook"), the server now will 
> have to replay the paging function to the entire paged resource and 
> see if the page sizes are consistent with the boundaries defined (i.e. 
> nothing has changed).  The "Arthur" change is detected, then 410 (or 
> some indicator) is sent.  This right?   So there is a decent cost 
> involved with determining for each page, if paging needs to be 
> reinitiated.

In the implementation I was proposing here, I was assuming page sizes 
didn't need to be exact.   If a few adds or deletes happen, the page 
sizes will be off a bit, and that will be okay.

I think the implementation I mentioned earlier today [1], where you put 
the boundary values in the URL, is better than this one in ever case.   
So, no need to worry about the logic here.   The core of my proposal is 
the definition of lossless paging above, and that we require it of LDP 
servers if they're going to do paging.

        -- Sandro

[1] http://lists.w3.org/Archives/Public/public-ldp-wg/2014Feb/0058.html

>
> (apologies for replying in reverse order on these, I should have 
> comprehended this first).
>
> - Steve Speicher
>
>
>     *microblogging example*
>
>     In the application Andrei and I are working on (microblogging),
>     the server will have a container of postings on each "channel". 
>     It will keep this sorted by the date of posting, so the LAST page
>     will be the most recent postings.  It's fine if some system posts
>     every 10 seconds (3 million posts per year) because the clients
>     just jump to the LAST page, and maybe follow a few PREV links,
>     depending how far back they want to go. Lossy paging would result
>     in postings not being shown to some people in some circumstance,
>     which is likely to be unacceptable.  For us, the paging function
>     will just be constructed by partitioning the range of dates
>     present, and re-partitioning if a lot of postings are inserted
>     into the middle or deleted.  In the normal case of postings being
>     added "now", we're just adding to the end, so no re-partitioning
>     is needed.
>
>
>     *conclusion*
>
>     I know this is a little complicated, but without it, it seems to
>     me paging is worthless, or even harmful, at least for the
>     applications I'm thinking of.  I'm willing to help draft text for
>     the spec if people think this is reasonable to include.
>
>           -- Sandro
>
>