We have not yet converged on a best (or even adequate) way to present a structured mountain of information to AI, not already in its training corpus.
AI agents fielded by major AI players still fail at the basic task of providing immediate and correct support for use of the current versions of their products. If a programming language is too new to have adequate representation in the training corpus, there isn't an accepted standard way to provide a reference manual targeting AI agents. Even the best way to include documentation in a large project so new AI agents can take over is controversial. A pile of linked markdown files really isn't an answer, less structured than a codebase itself, that AI is good at navigating.
Other HN posts have discussed using SQL as a backbone for the AI "mind mapping" support we need for AI more critically than for ourselves.
I was hoping that CQL could be an answer to this. Perhaps, but not its current primary goal.
> CQL is not a database management system: it neither stores nor updates data.
The same could be said for SQL. How does CQL differ from SQL? If I squint my eyes just a tiny amount, these ideas become really difficult to separate. I was always under the impression that the relational model is based upon many concepts studied in category theory. To my mind, all of the following things are overlapping parts of the exact same monster:
Set theory
Category theory
Graph theory
Type theory
Discrete mathematics
Relational algebra
Relational calculus
Relational modeling
An actual sql schema
Most attempts to replace/improve SQL derive from the fact SQL was a poorly conceived and designed interface, that originally was meant to be a very small DSL for end users, but unfortunately, was allowed to become a poor, complicated, confusing mess for app developers:
SQL is not an orthogonal language... This is because, in the early days, Ray Boyce and I did not think we were designing a language for programmers. ..
As it turned out, Ray and I were wrong about the predominant usage of SQL...
(same problem from JS, php, etc: Creators don't anticipate that developers will suffer and torture their needs with such anemic ideas!)
---
So the #1 thing any actual replacement or alternative to SQL is how actually become a good language for development, so it is actually composable, can be actually be used to reason about it, has minimal foot guns, etc.
There is a lot of misunderstanding and pushback, similar to how people in the past fight improvements over JS/C/C++ until typescript, rust comes.
The main innovation here seems to be compile time checking of that foreign keys are respected but that is a thing that can be added to SQL and there is at least one proposal for doing so. So I do not really see anything fundamentally different from SQL.
SQL is like Java, CQL is like Haskell. SQL has been around and used in production. CQL is a research language, possibly cleaner foundation but YMMV.
The math fields you list are connected, but whether they are the same monster - again it's kinda like claiming all programming languages and implementations are the same (Turing-complete?) monster.
> Reduce risk of failure through artificial intelligence. CQL contains an embedded automated theorem prover that guarantees the correctness of CQL programs.
Man, it's a rough environment right now marketing-wise. I don't know if they're contractually obligated to say the funny magic words, but the term AI is nearly entirely meaningless at this point. Akin to saying "behold my mighty calculator app: it prevents divisions by zero through artificial intelligence!"
It's what we used to call AI back in the 1970s and 1980s which has advanced a whole lot with little awareness. That is, people thought 10,000 rules was a lot of rules in 1998 and now you can work with 10,000,000 rules. And theorem provers, SAT and SMT all got vastly better.
I know very little about databases, but I work heavily with category theory, so fwiw: I think the main benefit is composition. The edge over SQL shows up when you combine schema mappings - a mapping is a functor, so when you migrate data along it the constraints come with it by construction, and you don't end up writing ETL and hoping integrity held.
As best as I can tell (but i really dont know much about databases) it's probably a narrow advantage - storage and everyday queries still go to Codd's model - but for stitching schemas together it seems like it could work.
I'm using similar math for automated formal verification, where this approach is what makes it tractable.
Since Codd's paper showed that the relational model dominates other approaches (for data storage) I would expect a paper that shows categorical database are not affected by this and what benefit they have.
While the relational model is claimed to be based on relations, the vast majority of the "relations" used in practice are functions, not general relations.
A general relation exists only between the columns of a table that are included in a multi-column primary key.
All columns that are not part of the primary key are functions of the primary key.
Most tables used in practice use a single column as the primary key, which is frequently just a number or a UUID. Most databases contain only tables that are functions, without any table that contains general relations.
The most frequently used kinds of joins are just function compositions.
Thanks for the sharing. It looks interesting but I did not dive deep into it. Just wonder how is it different from SQL trigger which can also ensure integrities?
We have not yet converged on a best (or even adequate) way to present a structured mountain of information to AI, not already in its training corpus.
AI agents fielded by major AI players still fail at the basic task of providing immediate and correct support for use of the current versions of their products. If a programming language is too new to have adequate representation in the training corpus, there isn't an accepted standard way to provide a reference manual targeting AI agents. Even the best way to include documentation in a large project so new AI agents can take over is controversial. A pile of linked markdown files really isn't an answer, less structured than a codebase itself, that AI is good at navigating.
Other HN posts have discussed using SQL as a backbone for the AI "mind mapping" support we need for AI more critically than for ourselves.
I was hoping that CQL could be an answer to this. Perhaps, but not its current primary goal.
> CQL is not a database management system: it neither stores nor updates data.
The same could be said for SQL. How does CQL differ from SQL? If I squint my eyes just a tiny amount, these ideas become really difficult to separate. I was always under the impression that the relational model is based upon many concepts studied in category theory. To my mind, all of the following things are overlapping parts of the exact same monster:
> How does CQL differ from SQL?
Most attempts to replace/improve SQL derive from the fact SQL was a poorly conceived and designed interface, that originally was meant to be a very small DSL for end users, but unfortunately, was allowed to become a poor, complicated, confusing mess for app developers:
https://cacm.acm.org/research/50-years-of-queries/
(same problem from JS, php, etc: Creators don't anticipate that developers will suffer and torture their needs with such anemic ideas!)
---
So the #1 thing any actual replacement or alternative to SQL is how actually become a good language for development, so it is actually composable, can be actually be used to reason about it, has minimal foot guns, etc.
There is a lot of misunderstanding and pushback, similar to how people in the past fight improvements over JS/C/C++ until typescript, rust comes.
But, oh boy, SQL need their typescript!
The main innovation here seems to be compile time checking of that foreign keys are respected but that is a thing that can be added to SQL and there is at least one proposal for doing so. So I do not really see anything fundamentally different from SQL.
https://keyjoin.org/
Full disclosure: I am one of the co-authors of this paper and an associated patch implenting it in PostgreSQL that we have proposed.
I am happy to see more people than us think this is useful.
Here is a tl;dr as well: https://keyjoin.org/tldr.html
> How does CQL differ from SQL?
SQL is like Java, CQL is like Haskell. SQL has been around and used in production. CQL is a research language, possibly cleaner foundation but YMMV.
The math fields you list are connected, but whether they are the same monster - again it's kinda like claiming all programming languages and implementations are the same (Turing-complete?) monster.
SQL is not an imperative programming language.
7 replies →
[dead]
> Reduce risk of failure through artificial intelligence. CQL contains an embedded automated theorem prover that guarantees the correctness of CQL programs.
Man, it's a rough environment right now marketing-wise. I don't know if they're contractually obligated to say the funny magic words, but the term AI is nearly entirely meaningless at this point. Akin to saying "behold my mighty calculator app: it prevents divisions by zero through artificial intelligence!"
It's what we used to call AI back in the 1970s and 1980s which has advanced a whole lot with little awareness. That is, people thought 10,000 rules was a lot of rules in 1998 and now you can work with 10,000,000 rules. And theorem provers, SAT and SMT all got vastly better.
There was a good blog post on how the category theoretic ideas behind this applies to data frames
What Category Theory Teaches Us About DataFrames https://news.ycombinator.com/item?id=47561426
I know very little about databases, but I work heavily with category theory, so fwiw: I think the main benefit is composition. The edge over SQL shows up when you combine schema mappings - a mapping is a functor, so when you migrate data along it the constraints come with it by construction, and you don't end up writing ETL and hoping integrity held.
As best as I can tell (but i really dont know much about databases) it's probably a narrow advantage - storage and everyday queries still go to Codd's model - but for stitching schemas together it seems like it could work.
I'm using similar math for automated formal verification, where this approach is what makes it tractable.
Since Codd's paper showed that the relational model dominates other approaches (for data storage) I would expect a paper that shows categorical database are not affected by this and what benefit they have.
My (amateur) take. CDB model (based on functions) has three advantages over RDB model (based on relations):
1. Easier modelling sum types (inheritance) due to duality.
2. Better handling of null due to labelled null.
3. Better foundation of elementary types (they're just another table ids). (Column stores often do that already, if your question is about storage.)
While the relational model is claimed to be based on relations, the vast majority of the "relations" used in practice are functions, not general relations.
A general relation exists only between the columns of a table that are included in a multi-column primary key.
All columns that are not part of the primary key are functions of the primary key.
Most tables used in practice use a single column as the primary key, which is frequently just a number or a UUID. Most databases contain only tables that are functions, without any table that contains general relations.
The most frequently used kinds of joins are just function compositions.
Thanks for the sharing. It looks interesting but I did not dive deep into it. Just wonder how is it different from SQL trigger which can also ensure integrities?
It's not much really, CDBs are based on foreign key relationships as a fundamental building block, rather than on relation.
The difference is more in theory than in practice.
Not to be confused with https://cql.hl7.org/ which is the CQL I know of.