openstack team mailing list archive

[Jay Pipes <jaypipes@xxxxxxxxx>]

On 06/12/2012 12:21 PM, Adam Young wrote:
> On 06/12/2012 04:24 AM, Gabriel Hurley wrote:
> > That said, we have also considered the case you propose where you
> > effectively "request everything and handle it on the client-side"...
> > however, I see that as a tremendously lazy solution. On the
> > service-provider end you have access to powerful database methods that
> > can do these operations in fractions of the time the client-side can
> > (especially with good indexes, etc.). And if you've ever worked in
> > mobile applications you'll know that minimizing data across the wire
> > is crucial. The only argument I've heard in favor of that is basically
> > "it's easier for us not to add API features".
>
> At the expense of loading your Database. Serverside paging and filtering
> both require one of two things: caching or additional Database queries,
> and both increase your server footprint. For small datasets, or for
> limited queries, this is not a problem, but for scalability you want to
> limit the work you do on the server.

This is actually incorrect for relational databases. Passing filter and 
pagination/offset parameters in the API allows *more efficient* queries 
to be executed on the database server (given proper indexing, of 
course...). Not passing in these parameters in the API means more full 
table scans, more rows transmitted over the wire, and more work done by 
the database server.

> For Keystone using the LDAP backend, caching and pagination are
> extremely expensive, and not something I would like to support.

This is a problem with LDAP, not with SQL backends, which are 
specifically built for querying in this manner. That said, however, 
there *are* certain filters that LDAP can deal with pretty well, right? 
Things like limiting to a specific OU can allow LDAP to winnow results, 
correct?

> an LDAP
> query is not guaranteed to come back in any particular order, so you
> can't just do the SQL trick of executing the query at offset + window
> size. You have to do the equivalent of a Cursor, and this places serious
> load on the LDAP server, the Keystone server, and possibly impacts other
> apps dependand on LDAP.

If this is the case, then one solution might be to raise 
NotImplementedError in the case of when an API filter is not supported 
by a backend and leave it up to the client to retry the "full set of 
results" request and do the filtering/pagination itself?

> > To speak on the specific feature of pagination, the problem of
> > 'corruption' by simultaneous writers is no excuse for not implementing
> > it. You think Google, Facebook, Flickr, etc. etc. etc. don't have this
> > problem? If you consume their feeds you'll notice you can fetch
> > offset-based pagination with ease. You'd never expect to see a
> > navigation element at the bottom of Google search results that said
> > "take me to results starting with the letter m".
>
> There is a major difference. We are working with data that has to be
> ACID. Google, Facebook and flickr do not. Before you migrate a VM, you
> need to know if the host meets the criteria for the VM. If it changes
> between when you check and when you reserve the space for the VM, you
> have just over committed. "Get it right eventually" does not work for
> management apps.

This isn't necessarily true. Nova's compute layer goes through a number 
of steps to ensure a semi-transactional nature to certain operations 
like resizing. Certain times a query needs to indicate that it intends 
to make a reservation of resources (see quota/reservation system now .. 
this is the SELECT FOR UPDATE paradigm) and other times, the query 
doesn't care about such things. In the latter case, there aren't 
expectations that the list returned is 100% accurate according to the 
state of the database at a particular timestamp of when the transaction 
occurred. In this case, filters and optimistic pagination works 
perfectly fine, IMHO.

> > None of this is a case of "someone might use it". The Horizon team has
> > been loudly asking for these features for 8+ months now. And not just
> > from Keystone but from all the projects. I have a list a mile long of
> > API features we need to really deliver a compelling experience. I was
> > just adding some items to it today, in fact.
> >
> > The rest of your points I have no strong feelings on and generally
> > agree, but when it comes to API features... I feel *very* strongly.
>
> Note that I am not saying "don't do pagination" as I agree, it is
> essential for good user experience. What I am stating is that we need to
> be smart about the techniques and technologies we choose, as there is
> always an upside and a downside.

Sure, agreed. :)

Best,
-jay