It’s not just the language, it’s schema evolution, data distribution, and exposed APIs.

I don’t want to give other teams direct access to a DB and have them Not only take a dependency on the schema, but have the ability to run arbitrary queries that may exhaust resources in ways that impact normal operations. If I expose an API, I control the access patterns and can evolve the schema separately to suit the workload.

If other teams need a replica to perform their arbitrary queries, I’d much rather have them using a richer data model that they can normalize into whatever form suits their needs than have to conflate that into a source of truth data store.

If you have a single business unit and can get away with commingling concerns within a small team, great, throw it all in a single DB. If it makes sense to split, however, do it quick and early to avoid a decoupling hell that is more expensive then having split in the first place.

That’s not a new problem though? The classic way to solve this in database is to expose the data API as stored procedures/views and restrict query access on the actual tables to just the DBAs - I think even MySQL which was late to the party here has had this ability for some time now.

> If I expose an API, I control the access patterns and can evolve the schema separately to suit the workload.

And then they will depend on your API data schema...

Yes, schema evolution is hard, but databases have many tools to help here that you will either have to recreate on your APIs or live without and have a harder time. Either way, all the trouble comes from data evolution, and any schema-only change is trivial to deal with.

Data distribution is something that varies from one DB to another, they usually have very good performance that is hard to replicate on your application layer, but are very hard to setup and keep running. But the point about control of resource usage is a good one.

> And then they will depend on your API data schema...

+100 to this

But that API schema is unrelated to the underlying DB schema. I’ve been able to run services with different backends (eventually consistent & low latency vs transactional) exposing the same API. That would not have been possible by just giving consumers access to a DB.

The DB can do everything can’t seem to understand this, for some reason.

Moving the goalposts, I can see.

Of course, merging different backends can not be done on the backend. Either you add a layer or you do it on the client.

That seems to be an odd way of looking at it, IMO. Most of the time, the entire point of building an API is to present consistent functionality to any consumer, not only ones under your control. Also, a well-behaved API is versioned so as to allow evolution of the API without breaking existing clients who can upgrade to new versions as they are able.

Any nontrivial change to the base model will mean a lot of complexity in the API layer and degraded performance. Maybe that's worth it for you, maybe if you're exposing this data to hundreds of external users who don't need high performance. But I feel that for most usecases, barebones DB access is the better option.

> That would not have been possible by just giving consumers access to a DB.

That depends on the DB's foreign table support, or equivalent, if it has it. It's certainly not categorically impossible with a DB.

> And then they will depend on your API data schema...

Which we have methods of evolving. I can put my API in gRPC and know exactly which changes will or won't break compatibility. Try doing that with a database.

> I don’t want to give other teams direct access to a DB [...]

You don't have to. Package up necessary queries into views and/or stored procedures and grant permissions only on those. Views can also shield from schema changes.

That also requires all data to be in the same DB. Often that’s not practical or possible.

Postgres FDW's do wonders to centralize this type of access.

In SQLServer you can define external tables that access data on another server, or even parses a file.

is there something fundamental from us making a new frontend to something like postgres? I think all the solutions that compile to SQL kinda work but it would be nice to have a new native interface that sucks less.

Nope. Nothing but industry-wide inertia is so massive by now that it just doesn't make sense to switch. Nosqls tried hard and went nowhere.

What i find funny is that most dbs translate sql into an internal representation that is remarkably similar to a proper relational algebra and optimise on that. I'd really just prefer the alebraic language as described in the original ages old paper.

There were also a few dbs trying to push sql-but-better languages... haven't heard about them for a while.

I have experimented with a different query model from sql for time series data. A query took the form of a Rhai script. (Rhai is a scripting language that has great interop with rust, so it was similar to how lua would be used to script parts of a game.)

Each query script would act on a few objects in global scope: `db` (handle to database), `start`, and `end` (time range of grafana dashboard that query was for).

I found being able to write imperative (rather than declarative) code to build a query to be extremely powerful, especially for storing variables and looping over things.

e.g. query script - just to get a feel for it:

    let dalmp = db.ts("pjm-da-lmp/western-hub") // ie retrieve the timeseries named 'pjm..'
        .with_time_range(start, end);
    
    let rtlmp = db.ts("pjm-5min-lmp-rt-lmp/western-hub")
        .with_time_range(start, end)
        .resample("1h", "mean");
    
    let da_err = rtlmp.diff(dalmp);
    
    #{
        dalmp: dalmp,
        rtlmp: rtlmp,
        da_err: da_err,
    }

A query script would be expected to return a dictionary-like object. the keys would be used as labels and the values would each be a time series object.

This is not the perfect solution for every problem but though it might be interesting to see an example of a very different approach to querying compared to sql.

As someone who absolutely loves the power of SQL, I abhor the footguns involved. Especially with null-based ternary logic that is incomprehensible to most people.

You have to learn how your database handles NULLs, different databases do it slightly differently. Once you do that, you're fine.

One big thing is that NULL should never "mean" anything in a business sense. It is the absence of a value, hence it cannot mean anything.

By the same token you need to understand how your database handles concurrency. Different databases do it differently.

> One big thing is that NULL should never "mean" anything in a business sense. It is the absence of a value, hence it cannot mean anything

I always had a problem with that notion. I mean, it has a memory representation, it has a set of operators you can apply to it, defined behavior in UNIQUE and FOREIGN KEY constraints etc. All this is well documented (though can behave slightly differently between databases, as you mentioned).

So, it has a set of valid values (just one: NULL) and a set of valid operations, so it's a type!

And now you have a type that looks somewhat similar to a null in "normal" programing languages, and SQL generally lacks the mechanisms for inventing your own types, so why wouldn't you use it in your business logic where it makes sense?

The design of NULL seems like a historical accident anyway. The best I can discern, there was a need for something to behave as "excluded" in the context of FOREIGN KEYs and outer joins, and so that semantic was just passed along to other areas where it made less sense.

I think a better type system and a better separation between comparison logic and the core type would have obviated most of the NULL's weirdness and made it far less foot-gunny in the process...

Many businesses are absent of value :)

The problem is that it isn't just me. It's everybody. Everybody that uses an SQL system has to learn a version of logic that looks like something they've learned before but actually has no relation to it. People can learn it, but it is a chore and takes a lot of real world experience (i.e. costly mistakes) to get it drilled into their head just like it did with me. As someone who does a lot of mentoring of data engineers, it's infuriating that they all have to go through this at some point.

Unfortunately the problem is made worse by the proliferation of languages making the equally heinous mistake of treating null as "false-y". Bad form, Peter, bad form!

Agreed. The power of SQL is awesome. But really wish people started making better languages that compile to SQL; like it happened with JavaScript

Same. Edgedb is doing something good but I wish I didn’t have to deploy a new service on top of my own db. Postgres only as well.