[t-writescode]

Hand-rolling SQL inside another programming language comes with some unpleasantness, like protecting against SQL injection and making sure the SQL is valid, especially when hand-constructing the query based on input parameters: “sort ascending? Descending? Filter all but things in this group? etc.”

Parameter management in some languages are unpleasant, like how JDBC only has positional arguments; and any time you do string concat in a language, you start getting in danger of misformed SQL.

Ultra basic ORMs, like Exposed (Kotlin), are well-tested frameworks that do exactly what I want. Want a single value in the =? Or want it to be an “in”? Or what if it’s null? Handled. No special string management. Want parameters? Handled.

When I see pure ORM’d code, I can feel safe expecting it to be protected from injection and formatting issues. It’s reduced cognitive load and greater safety.

When I see raw string SQL management, I have to put another layer of care and attention to try and make sure (and maybe still fail) there’s no application-crashing mistakes in that part of code.

It’s kinda like working with typed and compiled code. Greater protection from error.

[webstrand]

It sounds like you're describing a query builders which, unlike true ORMs, don't attempt to mask the object-relational boundary. They only help you build queries in a convenient way and sometimes type-safe way. Query builders are great.

ORMs are not query builders. The problem with ORMs is that they hide the query logic from you. It's not clear what's getting joined in, how its getting joined in, or if its actually 1 or N queries. The further you get from a straight query builder, too, the fewer SQL features you have access to, such as parameterized joins, CTEs, window functions, etc. Sometimes you can hack those into the ORM, but often you have to resort to string concat and build the parameterized query and arguments manually.

I've never used Exposed, but from what I can tell it's kind of a hybrid? the query builder parts look great, but I'd still be wary of the ORM parts.

I've never had a good experience debugging performance issues in ORM-generated queries. Maybe I haven't used the right ones, but the libraries I've used have gone out of their way to hide how the query is actually being executed and only have obtuse levers to tweak that execution. Sure you can see the raw logs of what queries the ORM executed, but you can't easily figure out why its chosen a particular strategy. Or you can't easily make it stop using a pathological join ordering.

[prmph]

If one can update the underlying Db, I've developed an ORM pattern that I use in my projects that works very well.

They keys is to encapsulate most of the (possibly complex) CRUD logic in Db functions (for retrieval these would be table-valued functions) and access these from the application side as virtual tables.

I also have capable but easy to use filtering/sorting/paging operators and combinators in the ORM that are translated into SQL where/sort/limit clauses.

Because the heavy lifting is already done by the db functions (which you have full control of to use whatever SQL you need), the pattern is actually quite powerful but easy to use.

You can define virtual read only tables, several virtual tables that access the same actual table in different way, custom operators that transcend SQL, etc

[webstrand]

Doesn't this break the object-relation mapping the ORM would be trying to do? Since the views would be producing resultsets that do not map to one single object? Or do you just define a new object type with virtual relations?

[prmph]

Exactly. I define a pair of DTOs for each virtual table: an input type (the shape of a record that can be inserted) and an output one (the shape of each record returned when you select from the virtual table, which might include generated, computed, and foreign columns, etc).

These types do not necessarily have to map to the underlying table types.

[andoando]

I HATE ORMs. I end up spending an hour or two trying to figure out why something isnt working for what should be a simple query.

Theyre also seem quite restrictive to what raw sql can do.

[monkeyelite]

You’re arguing against a straw man. All major language sql libraries are not based on string manipulation and provide things like escaping, arguments, etc out of the box.

[kaoD]

Query builders are still an antipattern (what we traditionally think of when we say query builders) because they are still a DSL that (1) you have to learn along with SQL and (2) never map cleanly to actual SQL, so you always have to resort to `.raw` calls when you need anything more complex than a simple SELECT+JOIN.

Even for simple SELECTs, I lost count of how many times I had to enable runtime DEBUG logging in my query builders to get a query that I can copy-paste into my DB client for debugging, data exploring, etc. I should be able to copy-paste my query from code and run it in `psql` with minimal changes.

Raw SQL is the way to go[0]. Period. You just need a thin wrapper on top of it that adds escaping, arguments, type-safety and other QoL.

[0] https://gajus.medium.com/stop-using-knex-js-and-earn-30-bf41...

[esafak]

That thin wrapper is a query builder.

[kaoD]

> antipattern (what we traditionally think of when we say query builders)

[monkeyelite]

I'm with you.

[ameliaquining]

Only for parameterization over scalar values. If you want to do any kind of composition more sophisticated than that, you're either stitching together strings or using some kind of more heavyweight abstraction like an ORM.

[monkeyelite]

That’s because the composition is supposed to be inside sql. Views, functions, etc.

This is another reason why the ORM is a leaky abstraction - it hides all the best features from you.

[ameliaquining]

I suspect the biggest reason those aren't more popular is that they usually have to be stored as state in the database, which isn't what you want when developing an application. You want all of your query logic to be versioned with your application code.

[branko_d]

> You want all of your query logic to be versioned with your application code.

SQL can be stored in version control just as well as any other code. This can include application-level queries as well as the DDL SQL which defines the actual structure of your database.

It's sad that tooling for this kind of workflow doesn't seem to be particularly good across the board, Visual Studio being somewhat of an exception.

[ameliaquining]

The problem isn't version-controlling the SQL; it's making sure that, when a particular build of your app executes a query that calls out to a user-defined function, it uses the SQL function code from the same commit that the app itself was built from. Otherwise you have a potential source of version skew, and those are always really annoying and unpleasant to deal with.

I think Postgres sort of supports this but it's really clunky, and also I think you'd have to run the function-creation code on every connection; I don't know whether that would create any performance problems.

What does Visual Studio do?

[monkeyelite]

In most organizations a database is broader than any individual application - both in lifecycle and scope. So it makes sense that this state exists in a different way.

I suspect it’s because people never learned to use them, but they did learn to use the ORM.

[ameliaquining]

The textbooks all say that, but is it really true in practice? I suspect it's a bit of holdover conventional wisdom from an earlier era of software. Usually, when I've created a database for a particular app, it was never used by anything else besides that app (and manual queries, for which versioning is mostly not relevant).

You might choose to have a set of views and functions that are versioned separately from your app, for the same reasons you might choose to set up a microservice. But as with microservices, it should be a deliberate decision that you make because you're getting something out of it that outweighs the complexity costs and version-skew headaches; it should not be the default way to do composition.

[branko_d]

> Only for parameterization over scalar values.

ADO.NET has full support for table-valued parameters.

[ameliaquining]

That's part of .NET Framework and therefore legacy, right? Do the database libraries from recent versions of .NET do this?

In any case, it's just one framework; previous comment said "all major languages". And it's useful to be able to abstract and compose over expressions and predicates and such, not just data values, which this still doesn't help with.

[branko_d]

> That's part of .NET Framework and therefore legacy, right? Do the database libraries from recent versions of .NET do this?

ADO.NET is available both in the legacy Windows-only .NET Framework and in the new cross-platform .NET (previously known as .NET Core).

> In any case, it's just one framework; previous comment said "all major languages".

Well, you are not implementing a piece of code in "all major languages" - you can pick the one that fits the problem best.

> And it's useful to be able to abstract and compose over expressions and predicates and such, not just data values, which this still doesn't help with.

You can do that via LINQ - there is even special query-like syntax built right into C# for that that looks like this:

    var companyNameQuery =
        from cust in nw.Customers
        where cust.City == "London"
        select cust.CompanyName;

This does NOT load the entire table in memory just to filter on City. It actually transpiles to SQL which does the filtering on the server.

But anything non-trivial is much better done in SQL proper, IMO. Most of the time, at least for OLTP, you'll be using static SQL - that is you will not need to change the text of the SQL query, just parameters. But dynamic SQL is a thing and can be very useful on occasion - which is string concatenation with all the problems that might bring.