[odie5533]

I'm talking knowing what the shape of the response and shape of the data going in ahead of time and being consistent. SQL is like a black box in and out. I mainly use Python. For that, it's nice to have things like Dataclasses for DTOs or Pydantic models or some sort of DTO class that has known field names and known types. When you use raw SQL, you lose all that or have to roll it yourself. And at that point, you're most of the way to an ORM or at least the data mapping portion of SQLAlchemy.

[sgarland]

Mapping a DB response to a Pydantic model is hardly an ORM.

[dijksterhuis]

> Pydantic serves as a great tool for defining models for ORM (object relational mapping) libraries. ORMs are used to map objects to database tables, and vice versa.

https://docs.pydantic.dev/latest/examples/orms/

[primering]

No, that is exactly what an ORM is, plus mapping it back. Anything around that is additional toolings that no ORM needs to be ORM, but is nonetheless usefull.

[halfcat]

Pydantic isn’t an ORM, any more than JSON.stringify() and JSON.parse() are an ORM.

Pydantic knows nothing of your database. It’s schema-on-read (a great pattern that pydantic is well suited for), or serialization, or validation, but not an ORM.

[odie5533]

Do you advise then use Pydantic for data mapping to/from raw SQL, to avoid using a full ORM? My thinking is you're almost at an ORM with that method with tools like SQLModel that I'm unsure what the benefit is to the plain Pydantic method.

[halfcat]

In one sense you’re right, but (at least in data projects) the goal is a bit different. We’re often reading not only from SQL databases, but also parquet files, CSV, JSON, APIs, piped input from another process’s STDOUT, and so on.

Basically we don’t always know what the future unknown data source we may be reading from, and also the schema of the source might change, but we can define what we expect on the receiving end (in pydantic), and have it fail loudly when our assumptions change.

[odie5533]

The nice thing about SQLModel is they're still Pydantic models so you can use them with custom data mappers like parquet, csv, json, etc. I think you make a good point about keeping the data model pure so you're not dependent on a data source. But I think SQLModel largely accomplishes that, and so does SQLAlchemy's declarative dataclass mapping (though I've not used the latter).