Re: [sqlalchemy] Declarative and deferred

[Michael Bayer]

On Aug 11, 2012, at 10:08 AM, David McKeone wrote:

> so with this helper I can context sensitively build up a result object with just the stuff I need (but without losing the benefits of the associated methods):
>
> deferred = User.get_deferred_except('id', 'password') # Get list of defer() instances for all columns, but those specified
> user = User.query.options(*deferred).first()
>
> # SQL Emitted -- SELECT id, password FROM user
>
> if user.is_valid(the_password):
> # Valid stuff
> else:
> # Invalid stuff
>
> Ok, well that worked great, but now I need to get the patrons name for some runtime specific reason. So I do this:
>
> full_name = " ".join([user.title, user.first_name, user.last_name])
>
> I now emit:
>
> SELECT title FROM user
> SELECT first_name FROM user
> SELECT last_name FROM user
>
> When what I really want at this point, and can predictably know in this case, is:
>
> SELECT title, first_name, last_name FROM user
>
> So, the question is, what is the best way to back-fill an object in a way that you keep the number of SQL queries low, while also getting the advantages of using a declarative instance?

think about just from a python level how that would necessarily have to work. the requirements are:

1. object is loaded with most attributes deferred.

2. code executes against loaded object, which asks for three attributes. the three attributes should load at once. however, the request for those attributes are separate.

Right off, your code is going to be more verbose - "user.title, user.firstname, user.lastname" are three separate instructions.

so no matter what, to follow the requested behavior, it *has* to be:

user = User.query.options(..).first()

load_more_attributes(user, ["title", "first_name", "last_name"])

full_name = " ".join([user.title, user.first_name, user.last_name])

that is, there must be an explicit instruction of some kind that tells it to load these three at once.

So for load_more_attributes, two choices. One is use the function you have:

query(User).options(User.get_deferred_except("title", "first_name", "last_name")).filter(...).first()

because, the identity map will have the effect that the same User instance is the target.

Or use session.refresh(), probably more direct:

session.refresh(user, ["title", "first_name", "last_name"])

also, if the columns you're actually using are along these lines, that is, they aren't 10K text files, I'd strongly encourage you to do some actual profiling to determine if all this complexity is necessary and not actually hurting performance much more. To pull over 3K of text over the wire in a single result is probably a lot less overhead than to pull 1000K of text in two or three queries.


> Additionally, I'd also wonder if it would be possible to make a 'faulting' type of object; one where requesting any of the deferred columns would cause all of the deferred columns in that instance to be loaded.

well this is what the "group" feature does, if any columns in the group are touched, the whole group is loaded. My advice would be to do some profiling, come up with groups that are tailored close enough to the groups of attributes that tend to be called together, and to not overthink it.

[David McKeone]

 

session.refresh(user, ["title", "first_name", "last_name"])

This was the part that I was missing.  It's fairly readable and it does exactly what I'd need.

 

also, if the columns you're actually using are along these lines, that is, they aren't 10K text files, I'd strongly encourage you to do some actual profiling to determine if all this complexity is necessary and not actually hurting performance much more.    To pull over 3K of text over the wire in a single result is probably a lot less overhead than to pull 1000K of text in two or three queries.

At this point I'm really just exploring the boundaries of the tool so that I can select a flexible design.  I still haven't quite found the sweet spot between what can/should be lazy and what cannot/shouldn't be lazy.  In the existing application (the non-ORM one) all of this is done with an abstracted form of direct SQL (kind of like SQLAlchemy core).  I'd like to convert some of those sections to use declarative objects instead, so the point of of this is to know that if I do go down that path then I could still optimize the columns if I needed to (read: after I profiled it and determined that it was necessary) without having to drop all the way down to SQLAlchemy core and then change things from passing objects around to passing keys in some circumstances.  Although it's very likely that you are correct and that the complexity induced from using this kind of system may outweigh the over-the-wire savings -- I guess we'll see when I get there.

well this is what the "group" feature does, if any columns in the group are touched, the whole group is loaded.  My advice would be to do some profiling, come up with groups that are tailored close enough to the groups of attributes that tend to be called together, and to not overthink it.

I will certainly look into this some more, since there are certainly groups of columns that can be naturally grouped.  Plus I imagine that session.refresh() would load the entire group if an attribute from a group was passed to it.  So that could be an interesting way to chunk it.

Thanks for responding with such a great post.  Its certainly helped me think through the issues from different angle.

[Michael Bayer]

On Aug 11, 2012, at 3:43 PM, David McKeone wrote:

> Plus I imagine that session.refresh() would load the entire group if an attribute from a group was passed to it. So that could be an interesting way to chunk it.

I think the attributes to session.refresh() trump any deferred rules. It will load just what you send it.