ORMs force the thinking "objects first" (ie. work with objects, the SQL will be handled for you). This translates into a design of "per object" architecture. Business logic is abstracted for handling single rows only - one at a time - encapsulation being the design idiom. However, this is a trap that immediately becomes apparent when you realise that you've forgone the powerful set-based capabilities of SQL. SQL doesn't "need to see" a single row of a table as anything special to warrant ring-fencing in the name of encapsulation. The whole (related) data model is a single algebraic representation of state.
One could extend that for more complex operations - instead of immediately executing the operations, it'd collect the operations and convert them into efficient queries. In a way, this is already done in "fluent interfaces" like LINQ, but with a more flexible language (Lisp?), I'm sure one could use standard control flow instead of a custom layer.
I'm my opinion this is the least desirable behavior. It abstracts the data layer away to the point where you don't think about what's happening under the hood when you type foo = Klazz.new("id"); foo.bar; foo.baz;
Was that one, two or three separate queries ? Did a query run at all? Does it matter, after all you have an object to work on, right in your language of choice.
Sure the ORM may be smart about it and do the right thing - I have no doubt that most decent ones day. But that misses the point - as the person using this interface you get divorced from the direct awareness that this data has an origin, and there is a cost to access it. After all, you're just instantiating an object.
Patterns that use interfaces that make it so simple to do a thing that you don't need awareness that you're doing it tend to degrade quickly after the original engineers who did have that understanding leave a project.
Isn't knowing what queries are issued giving you a false sense of knowledge about how the data is actually fetched?
And much the same way you can request a query plan, the ORM could provide the list of queries issued for a given piece of code.
records = ORM.search(type='X')
for record in records:
record.a = record.b + 30
and only trigger a single UPDATE query?This is the example in their docs:
Entry.objects.all().update(n_pingbacks=F('n_pingbacks') + 1) update records
set a = b + 30
where type = 'X'Reminds me of, "in theory, there's no difference between theory and practice, in practice there is".
For the app I work on, the DB is only viewed as an object store for very small sets of data - configuration, specifically. It's impossible to map it to objects in any realistically performant way for our user's data, not least because that would require the data round-tripping through the application. It's much, much more efficient to execute my multi-million row updates and inserts on the database server directly rather than drag the data down to the app, tweak it slightly, then persist it back.