Thanks for taking the time to expand on your ideas, Roman Zaiev . Now I see your architectural approach much better. I think the key difference between our views comes down to what we consider the main responsibility of the persistence layer. You focus strongly on transparency and control, and for Hypha it clearly works well. I get that. For me, ORM’s biggest value isn’t query abstraction, but maintaining long-term consistency between the domain model, schema and business rules, without re-implementing that logic manually in every repository or factory. That’s why I still think ORM is often the more efficient choice, especially in systems that evolve quickly or involve many contributors. It prevents rebuilding a lot of what ORM already provides out of the box (migrations, typing, constraints, semantic exceptions, domain-level validation). But if the raw SQL + factories approach keeps Hypha simpler and more predictable, that’s awesome. Different contexts, different trade-offs. 👌 Great discussion, appreciate the exchange of perspectives! 👍

I fully agree that using Postgres’ native capabilities consciously gives a huge advantage in terms of performance when executing code. However, I don’t agree with the statement that ORM is “bad”. You just need to know how to use ORM properly. It’s like saying: why use a jackhammer if it costs electricity, when I can do the same manually with a hammer and a chisel? Some things that I think are missing in this perspective: maintainability – in a project with dozens of tables, the need to maintain raw SQL (whether as templates or – even worse – strings in Python) is a recipe for a nightmare. Not to mention SQL validation, a syntax checker or at least a formatter. Even the smallest schema change (e.g. column name) requires updating dozens of places in the codebase. migrations and validation – this is another area we have to watch very carefully when a table schema changes, and manually handle additional serialization, eg. we cannot simply pass Python True into a boolean column, instead we must convert it into SQL literal TRUE . exception handling – ORM gives more semantic domain-specific errors instead of generic driver messages like: ERROR: duplicate key value violates unique constraint . consistent domain model – without ORM, we still map records into objects, only that mapping happens… inside the developer’s head. For me the right approach should rather look like: using ORM as the default tool for CRUD and the domain model, and raw SQL where it gives a real advantage (aggregations, optimization, reporting). But even in the second case, you can create advanced queries in a clean and efficient way, however, you must know and understand the specific ORM at an advanced level.