Struggling With OOP Concept

Say you have a "wheel" class, which loads its data from the wheels table in its constructor

Constructors should not be doing any work. Instead they should contain only assignments. Otherwise you make it very hard to test the behavior of the instance.

Now, we have a "car" class, which loads its data from the cars table joined with the cars_wheels table and creates wheel objects from the returned wheel_ids:

No. There are two problems with this.

Your Car class should not contain both code for implementing "car logic" and "persistence logic". Otherwise you are breaking SRP. And wheels are a dependency for the class, which means that the wheels should be injected as parameter for the constructor (most likely - as a collection of wheels, or maybe an array).

Instead you should have a mapper class, which can retrieve data from database and store it in the WheelCollection instance. And a mapper for car, which will store data in Car instance.

$car = new Car;$car->setId( 42 );$mapper = new CarMapper( $pdo );if ( $mapper->fetch($car) ) //if there was a car in DB{    $wheels = new WheelCollection;    $otherMapper = new WheelMapper( $pdo );    $car->addWheels( $wheels );        $wheels->setType($car->getWheelType());    // I am not a mechanic. There is probably some name for describing     // wheels that a car can use    $otherMapper->fetch( $wheels );}

Something like this. The mapper in this case are responsible for performing the queries. And you can have several source for them, for example: have one mapper that checks the cache and only, if that fails, pull data from SQL.

Do I really have to choose between beautiful OOP code with a million queries VS. 1 query and disgusting, un-OOP code?

No, the ugliness comes from fact that active record pattern is only meant for the simplest of usecases (where there is almost no logic associated, glorified value-objects with persistence). For any non-trivial situation it is preferable to apply data mapper pattern.

..and if I make a "city" class, in its constructor I'll need to join the cities table with the cities_dealerships table with the dealerships table with the dealerships_cars table with the cars table with the cars_wheels table with the wheels table.

Jut because you need data about "available cares per dealership in Moscow" does not mean that you need to create Car instances, and you definitely will not care about wheels there. Different parts of site will have different scale at which they operate.

The other thing is that you should stop thinking of classes as table abstractions. There is no rule that says "you must have 1:1 relation between classes and tables".

Take the Car example again. If you look at it, having separate Wheel (or even WheelSet) class is just stupid. Instead you should just have a Car class which already contains all it's parts.

$car = new Car;$car->setId( 616 );$mapper = new CarMapper( $cache );$mapper->fetch( $car );

The mapper can easily fetch data not only from "Cars" table but also from "Wheel" and "Engines" and other tables and populate the $car object.

Bottom line: stop using active record.

_{P.S.: also, if you care about code quality, you should start reading PoEAA book. Or at least start watching lectures listed here.}

my 2 cents

ActiveRecord in Rails implements the concept of lazy loading, that is deferring database queries until you actually need the data. So if you instantiate a my_car = Car.find(12) object, it only queries the cars table for that one row. If later you want my_car.wheels then it queries the wheels table.

My suggestion for your pseudo code above is to not load every associated object in the constructor. The car constructor should query for the car only, and should have a method to query for all of it's wheels, and another to query it's dealership, which only queries for the dealership and defers collecting all of the other dealership's cars until you specifically say something like my_car.dealership.cars

Postscript

ORMs are database abstraction layers, and thus they must be tuned for ease of querying and not fine tuning. They allow you to rapidly build queries. If later you decide that you need to fine tune your queries, then you can switch to issuing raw sql commands or trying to otherwise optimize how many objects you're fetching. This is standard practice in Rails when you start doing performance tuning - look for queries that would be more efficient when issued with raw sql, and also look for ways to avoid eager loading (the opposite of lazy loading) of objects before you need them.

In general, I'd recommend having a constructor that takes effectively a query row, or a part of a larger query. How do do this will depend on your ORM. That way, you can get efficient queries but you can construct the other model objects after the fact.

Some ORMs (django's models, and I believe some of the ruby ORMs) try to be clever about how they construct queries and may be able to automate this for you. The trick is to figure out when the automation is going to be required. I do not have personal familiarity with PHP ORMs.