[blagie]

The basic problem is that:

1) Relational databases are the best abstraction we've found for storing data. Despite years of attempts (OODB, XML databases, various nosql stores, etc.), we have not been able to improve on it. postgresql, by adding native JSON support, became a better mongo than mongo virtually overnight.

2) Most people never learn databases, and most people who do are idiots working on enterprise three-tier architectures, so mostly it's misused. There is an attempt to hide them.

3) ORMs generally make easy stuff easy, and hard stuff painful.

What I generally want is:

1) Something translating SQL syntax into my native language, but maintaining SQL full semantics and expressiveness.

2) This should allow me to be database-agnostic.

3) This should prevent things like injection attacks.

4) It should not map onto objects. It should maintain 100% of the capability of SQL. The only differences should be syntactic (e.g. instead of writing WHERE, I might write .where() or similar).

5) This may and ideally should have added functionality, for example, around managing and organizing database migrations.

Stored procedures and virtual tables are also very important (but poorly implemented in most databases). These:

1) Allow proper abstraction at the SQL level.

2) Improve performance.

What most programmers fail to understand -- since universities don't teach -- is how powerful and elegant the underlying theory of databases is.

[snorremd]

My best experience integrating with a Postgres database was using Clojure with the Honey SQL library [1]. Essentially Honey SQL operated solely as a query builder library allowing me to express queries as regular Clojure EDN maps and vectors (comparable to objects and arrays in JavaScript), or nested function calls returning the same kind of data. In essence SQL queries were expressed in a Clojure native data format which finally could be transformed into SQL syntax.

The huge win was that I could use the Clojure REPL to run my SQL-generating functions and see that the SQL actually matched what I was aiming for. Caching the query-building was as simple as memoizing the query building functions as queries are parameterised in any case, so can be reused.

There is something really nice about having the full power of SQL at your fingertips. The only drawback compared to ORMs of course is that it outputs the resulting query result in a flat structure. So any aggregation you'd need to handle in code or use something like Postgres JSON aggregation.

[1] https://github.com/seancorfield/honeysql

[sureglymop]

I learned about database theory a few times (and even interned as a DBA before I started uni). But it wasn't until my discrete math class in university that I finally understood, "oh, that's why we call a table a relation, a row a tuple, etc", and from there - started understanding the whole relational model and why it makes sense.

I think ORMs however, became a thing more so due to fear of having to write (correct) SQL. It was meant to reduce complexity for developers but of course that's not really how it panned out in reality.

[everforward]

https://GitHub.com/masterminds/squirrel sounds like what you want, in Go at least. They bill it as a SQL query builder instead of an ORM.

It lets you build SQL queries using code and just returns the built SQL query as a string, the list of variables to pass into the query, and an error if your query was invalid.

Actually executing the query and doing stuff with the result is explicitly outside their scope. I typically use it with sqlx with a struct per query, just to avoid writing tedious row iterators.

Makes it super easy to eg define a function that adds paging to a query and returns the paged query. You can even write a function that takes an HTTP request, reads out standard paging parameters and adds them to a query.

[pjc50]

> It should not map onto objects

? To me, this is the main use of an ORM: don't let it do query building for you (unless it's fairly simple LINQ stuff), but just have something map the rows returned to typesafe objects for you.

[blagie]

Mapping rows returned from a query onto typesafe objects is okay, but doesn't require an ORM. It's quite literally just making sure each row:

1) Has the correct native type

2) Can be accessed by name

Many normal database APIs do that themselves. I guess there's a little bit more for writes.

The point of ORMs is that they generally map objects onto tables. You create a class, and the ORM will create the database table for you based on what you have in the class. For example, for Django, you write:

   class Blog(models.Model):
       name = models.CharField(max_length=100)
       tagline = models.TextField()

       def __str__(self):
           return self.name

(source: https://docs.djangoproject.com/en/5.0/topics/db/queries/)

And it will make the table for you. If you change the class, it will make the migration for you. That's clean, simple, and easy if you're not doing anything complex. It saves a ton of work.

However, this is a Very Bad Idea for complex systems, since exactly as you point out, there shouldn't be a 1:1 mapping between tables and classes. There may very well be a 1:1 or many:1 mapping between queries and classes. More sophisticated ORMs can do a bit of that, but at some point, you run into the equivalent of Greenspun's Tenth Rule, where the ORM turns into a more complex, buggy version of SQL.

[mike_hearn]

Sounds like jOOQ?

https://www.jooq.org/

Note that although advertised as for Java, it really supports any language that can run on the JVM which is many of them these days.

[PH95VuimJjqBqy]

> Most people never learn databases, and most people who do are idiots working on enterprise three-tier architectures

Tell me how you really feel :)

My biggest problem with ORM's is that it causes people to sprinkle the ORM code throughout their codebase but they'd never do that with SQL, they'd want to try and sequester it to a system whose responsibility is the retrieving and updating of data.

It puts people into a poor mindset around their data.

[feoren]

That's like saying: my problem with Functions is that people sprinkle them throughout their codebase, instead of having one single place where all their Functions are defined.

For most applications, "Retrieving and updating data" is the entire point, or at least a deeply fundamental part of how they work. It shouldn't scare you that such a core function -- the entire reason your application exists -- is "sprinkled around" your codebase. It should scare you if it isn't! I think the reason it scares you is because you imagine the performance nightmare of every random function possibly hitting the database N times. That's using an ORM terribly. Good ORMs let you plan database actions, and compose those plans.

Guys, you really, genuinely, don't have to choose between (a) writing SQL strings in code, and (b) using an ORM badly. You can truly do better!

[PH95VuimJjqBqy]

putting repeated data requests behind a function is what I'm suggesting...

sprinkling your ORM code all throughout your codebase is the equivalent of NOT using functions but instead copy/pasting the implementation everywhere it's needed.

Most ORM's recognize this is a problem, which is why they often try to bake in some sort of solution for re-use, only it's always done badly. They'll generally attach it to the model and it will turn into some sort of unobvious behavior or they'll attempt to be able to attach SQL fragments by name. What it ends up doing is effectively spreading an SQL query over several files. Often by convention so there's no real way to know if part of the query is going to be in file X without just checking or knowing ahead of time.

[blagie]

I kind of disagree. If you're using an ORM, the ORM is the abstraction layer which should be sprinkled throughout the code. You shouldn't wrap a layer in a layer.

If that's the right abstraction, use an ORM. If it's not, don't.

If all you have are lists of items (a todo list, a set of employee records, a set of products), an ORM is a fine abstraction. If you have or expect to do something more complex, it's probably a bad abstraction.

[PH95VuimJjqBqy]

> If you're using an ORM, the ORM is the abstraction layer which should be sprinkled throughout the code. You shouldn't wrap a layer in a layer.

preventing people with this belief from sprinkling ORM code all throughout the codebase is reason enough to ban the use of an ORM.

The specifics of how you retrieve data is an implementation detail. If someone wants to use an ORM have at it, but don't sprinkle it throughout the codebase, place it behind well defined functionality (behind a system dedicated to pulling data).

The fact that ORM's have their own query language should be all you need to know.

One could easily argue that you should be able to manually interact with db readers all throughout the codebase, after all, it's an abstraction.

But no one would ever actually argue that. Like such forward-only db readers, ORM's are a way to pull data, they are not the abstraction layer for pulling data for your application.

[blagie]

> The specifics of how you retrieve data is an implementation detail.

Correct.

> (behind a system dedicated to pulling data).

That's the textbook definition of a good ORM.

> One could easily argue that you should be able to manually interact with db readers all throughout the codebase, after all, it's an abstraction.

It depends on the abstraction they present. SQL throughout the code would be a train wreck. On the other hand, something like:

* A stored procedure to have a filtered list of gizmos...

* Mapped onto an AJAX or a function call in a data layer...

* Being called whenever one needs a filtered list....

is the way to do it. If all you have are tables of simple data, that's exactly what a good ORM would do.

[brainlessdev]

Can it be database-agnostic and 100% expressive at the same time? Perhaps expressiveness is different depending on the engine.

SQLx comes close to this btw.

[blagie]

Yes.

The basic data model was invented a half-century ago. Expressiveness of basic SQL is not different depending on the engine. Relational algebra is the same however it's implemented.

Some databases have extensions, and those are okay to include. You should be able to either (1) choose to not use those or (2) lock yourself into a subset of databases which support the extension you need.

It's good if those extension were namespaced. For example:

* If I want to use a postgres JSON type, it should live under postgres.json (or if it cuts across multiple databases, perhaps extensions.json or similar)

* Likewise, database built-in functions, except for common ones like min/max/mean, should be available under the namespace of those databases.

* It's okay if some of those are abstracted out into a namespace like extensions.math, so I can use extensions.math.sin no matter whether the underlying datastore decided to call it SINE, SIN, SINE_RADIANS, SIN(x/180*PI), and if it doesn't have sine, an exception gets raised.

The basic relational data model provides the best expressiveness for representing data and queries created to date, and doesn't differ. It's a good theoretical model, and it works well in practice. There's good reason it's survived this long, despite no lack of competition. The places expressiveness differs are relatively surface things like data types and functions.

It's also okay to have compound types, where I am explicit about what the type is in different databases. e.g.: string_type = {mysql: 'TEXT', postgresql: ...

[random_kris]

https://orm.drizzle.team/ Basically this ?

[iccananea]

Drizzle does exactly what I described in the article: it re-implements SQL in the target language (in this case TypeScript).

[random_kris]

But you have raw SQL operator and all drizzle operations does is to build you the SQL query.

Well other orms do that as well ofc. But I feel this is a low level as it can get

[feoren]

> What I generally want is:

> 1) Something translating SQL syntax into my native language, but maintaining SQL full semantics and expressiveness.

You have just described a good ORM used well.

> 2) This should allow me to be database-agnostic.

Meh, you sacrifice some powerful features if you demand total database agnosticism, and how often do you actually switch databases? Being database-agnostic is a side benefit of writing your logic simply against a good abstraction. The biggest benefit is composability. You can write (and optimize) one query against abstractions, and re-use that query in lots of different ways.

> 3) This should prevent things like injection attacks.

As all ORMs automatically already do.

> 4) It should not map onto objects. It should maintain 100% of the capability of SQL.

If it's not mapping onto objects, what is it doing? The problem is that your mental model of what "objects" means includes awful design decisions like deep inheritance trees, mutability, and lots of reference cycles (this.parent.child[0] == this, etc.) If your object model is already clean and following relational principles, then mapping into that model is exactly what you want.

It should not strive to maintain the capability of some bastardized pseudo-language which is despised by the progenitors of relational logic. It should strive to support the relational model. That's not the same thing.

> 5) This may and ideally should have added functionality, for example, around managing and organizing database migrations.

No, because your code versions and your database schemas advance together. To reliably run a database migration in code, you'd need to run it with the code version that exactly matches each step in the schema. That means for each step in the migration, you'd need to checkout the correct commit in git, compile, and run the ORM code in that version. Either that, or you're maintaining a code model that is compatible with every historical database schema, which is way worse.

But what ORMs should be able to do (and I haven't found one that does this well) is generate SQL migration scripts for you, which you store. Those would be frozen relative to the database schema version, so all the above problems go away.

> What most programmers fail to understand -- since universities don't teach -- is how powerful and elegant the underlying theory of databases is.

The underlying relational model is powerful and elegant. SQL itself is not. SQL is disliked by the founders of the relational model. Good ORMs let you incorporate the relational model into your code.

[blagie]

> You have just described a good ORM used well.

No. I did not. ORM is an "object–relational mapping." It maps data relations onto objects.

https://en.wikipedia.org/wiki/Object%E2%80%93relational_mapp...

> how often do you actually switch databases?

Very often. Virtually all of the systems I write have at least two back-ends to maintain that flexibility. At the very least, I'd like my systems to run well locally for development but also to scale. The easiest way to do that is to e.g. support both SQLite and Postgres, but there are other ways which make sense.

In proprietary settings, I like having a BATNA. The architectural flexibility means I get better prices on hosted services and aren't liable to turning into a cash cow through lock-in. That's savings even if I'm not switching.

> If your object model is already clean and following relational principles, then mapping into that model is exactly what you want.

This is where your thinking broke. A good object model is NOT a relational model, and a good relational model is NOT an object model.

Learn the theory of both. They're both good theories, but they're different.

An ORM makes sense if you want to use an object model, but want the backing store to be an RDBMS.

> But what ORMs should be able to do (and I haven't found one that does this well) is generate SQL migration scripts for you, which you store. Those would be frozen relative to the database schema version, so all the above problems go away.

I believe that's one instantiation of what I wrote: "This may and ideally should have added functionality, for example, around managing and organizing database migrations." It's actually exactly what I was thinking.

Some ORMs do this not badly, actually. Don't let the perfect be the enemy of the good. A simple system which does 90% of the work of generating a migration (with manual verification and tweaks) is often better than a complex one which tries to do 100% of the work.

> The underlying relational model is powerful and elegant. SQL itself is not. SQL is disliked by the founders of the relational model.

Citation required.

In either case, ORMs aren't translating just syntax, but also semantics. That's where the problem lies. If you're not doing that, you're not an ORM.

> Good ORMs let you incorporate the relational model into your code.

You're confused about what an ORM is. ORMs essentially map an RDBMS onto an object model (which an OODBMS does natively). The two models are fundamentally different. It's a translation layer.

Any good database library will let me "incorporate the relational model into my code." That's not an ORM.

[feoren]

You're being condescending, telling me to "learn the theory". We've clearly had different experiences and work with different kinds of systems. You're automatically discounting what I say by assuming I must be a newbie because I disagree with you. No. We both have things to learn from each other and you telling me "go RTFM, newbie" is shutting down that opportunity.

> Virtually all of the systems I write have at least two back-ends to maintain that flexibility.

It sounds like that's a worthwhile tradeoff for you, but it is a tradeoff, giving up some of the unique power of each individual database in order to support both. Realize that most people don't do this.

> A good object model is NOT a relational model, and a good relational model is NOT an object model.

Maybe this is our biggest disagreement. I believe there is a model that can get the biggest benefits of both. I'm not alone in this; papers in Software Engineering, "Out of the Tar Pit", prescribes this as a possible solution to a lot of woes in software engineering:

"The classical ways to approach the difficulty of state include object oriented programming which tightly couples state together with related behaviour, and functional programming which — in its pure form — eschews state and side-effects all together ... We argue that it is possible to take useful ideas from both and that — when combined with some ideas from the relational database world - this approach offers significant potential for simplifying the construction of large-scale software systems"

https://curtclifton.net/papers/MoseleyMarks06a.pdf

I agree with them: Object Oriented Programming alone is nice for UIs but otherwise has failed to live up to its hype; Functional Programming is elegant and pure but hard to actually get anything done with; the Relational Model is simple and powerful but writing an application entirely in database procedures is a Lovecraftian Horror. So look for a model that takes the best from all three, and apply that model -- at least conceptually -- in both your database design and your application code. It is actually possible, and it's wonderful. But it will piss off the zealots on both sides.

I'm not sure what you consider "a good object model", but there is very little agreement in our industry on what exactly that looks like. If yours permits the CS 101 inheritance examples like "Dog : Animal" then I'd strongly disagree with you. Or maybe yours is the Smalltalk message-passing version: better, but still not immune to improvement. Don't assume that everyone who doesn't 100% line-toe bog standard OOP and bog standard SQL RDMSes needs to "RTFM".

>> SQL is disliked by the founders of the relational model.

> Citation required.

"SQL isn’t just user hostile, it involves some very serious departures from relational theory ... Suffice it to say that those departures are so serious that I honestly believe SQL has no real right to be called relational at all." -- Chris Date, who worked closely with EF Codd and helped spread his ideas.

https://red-gate.com/simple-talk/opinion/opinion-pieces/chri...

[blagie]

Citations appreciated! I will not respond in-depth to most parts, because I agree with what you're writing. But I will make one or two points. To clarify a piece of confusion:

> Maybe this is our biggest disagreement. I believe there is a model that can get the biggest benefits of both. I'm not alone in this; papers in Software Engineering, "Out of the Tar Pit", prescribes this as a possible solution to a lot of woes in software engineering:

I don't necessarily disagree, Moseley does not describe an OO model. If you showed their proposal to someone who does e.g. just Java, they would puke.

> I'm not sure what you consider "a good object model", but there is very little agreement in our industry on what exactly that looks like

A more accurate statement is that there are many models. The very first one I learned formally was the Booch method, over a quarter-century ago. My productivity fell about tenfold when I began applying it. Second one was Java. Less than tenfold. Like you, I eventually rejected OO for most things other than UX, or relatively small objects which are more-or-less little more than data types.

-----

But onto the main point:

- Models have theoretical properties which work well in isolation.

- Mixing models often destroys those properties.

- For example, I can slice code vertically or horizontally. Either works well. If I mix the two, I have no abstraction or modularity left.

- Java got rid of MI for a reason. It's not that MI is bad, but it doesn't fit into the Java OO model.

- If I add a little bit of OO to functional, I get mutation, and almost all of the benefit disappears in a poof. (note: There are systematic ways to do this which maintain benefits of both)

- Aside from theoretical properties, models are a way to communicate. You know what to expect in code.

You can design hybrid models, but you can't just mix models. Code kinda falls apart. None of the OO models map cleanly onto relational, or vice-versa. It's possible to do other models that perhaps combine aspects of both, but they're no longer OO.

An ORM is a bastardization of the two models, by trying to mesh them together. In contrast, I like Moseley Marks a lot, which is it's own model.

-----

And a footnote: I no longer find it hard to get things done with functional programming. Most of my code is pretty close to pure functional, with a few well-controlled places with state (on the front-end, with Redux). Part of that is experience on my part, but part of that is that both Python and modern JavaScript surface many of the most useful aspects of functional in ways which are linguistically nice.

-----

As a second footnote: A lot of this is based on domain. For example, numerical code works great in functional, while as you pointed out, UX maps well onto OO.