Hacker News Reader

Comment by scosman | Hacker News Reader

Lots is boilerplate, but eventually you'll need more logic, then some CPU heavy tasks, then a memory heavy task, then a task suited to a specific language, then a task that needs a GPU, then a distributed memory cache.

Web servers are easy to horizontally scale. Even minute to minute if needed. Scaling your database hits limits and is much harder. Keep the database dumb.

View on HN

That’s reasonable advice for growth stage companies aiming for huge user bases.

I work at an Alexa top10k, and our database server is quite small and sits at about 10% load most of the time.

You can grow huge with vertical scaling. Not all the way to “household name” size, but still truly huge. Servers with tens of tb of ram are available by the hour.

Really really expensive servers that sit at 10% load most of the time is one of the points.

This isn’t an expensive server. It’s just not a problem that requires substantial resourcing.

Licensing cost is another factor. DBMS tend to be paid software than Web application server. Scaling up Oralce sometimes very expensive.

Oracle is extremely expensive. I've been using Postgres for a long time and you can eek quite a bit of performance out of it.

> That’s reasonable advice for growth stage companies aiming for huge user bases.

Not really, that's the result of hard-earned lessons that pay off from day 1.

Think about it: you have data in a database, and you put up an API to allow access to it.

Your API needs change, and now you need to make available a resource with/without an extra attribute. You need to provide it as a new API version in parallel with the current API version you're supporting.

What will you do then? Reupdate the whole db with a new migration? Put up two databases? Or are you going to tweak the DTO mapper and be done with it?

I use a database that was designed by someone who understands one sometimes wants to run a migration. It’s not hard to add new fields.

Perhaps my experience is unusual having worked on a distributed database for several years but this stuff just seems like table stakes.

> I use a database that was designed by someone who understands one sometimes wants to run a migration. It’s not hard to add new fields.

We're not talking about adding attributes to a database. We're talking about one of the most basic cases in API design: changing a interface and serving two or more versions.

If the mapping step between getting data from the db and outputting it in a response is expected to be eliminated by getting the DB to dump JSON, how is this issue solved?

And how exactly is a full blown db migration preferable to mapping and serializing a DTO?

The way you are supposed to use Postgrest or Postgraphile is to provide only views and stored procedures as the API. These views and stored procedures are layered over your data schema and should be the only way to query or change the data. This way you can make API changes without requiring migrations.

The point is that instead of having queries that return tables that you then map to DTOs for serialization, you have queries that return JSON directly.

Another way to accomplish the same thing, if you don't want views or stored procedures, is to use a library like jOOQ that does much the same thing in your Java middleware.

Or use an API layer such as GraphQL that doesn’t require versioning for such a minor change since additive changes (or deprecations) to the API do not affect existing queries - each query states exactly what it needs.

Backwards incompatible changes will eventually be required by the business - at least once that business is oriented such that technical concerns do not overrule all other concerns.

> Or use an API layer such as GraphQL (...)

But that already qualifies as the boilerplate code that was supposedly the root problem, isn't it?

But now instead of a simple mapper you've bolted on another complex system.

So exactly how does JSON in the db solve any of the problems that were presented? Other than adding two buzzwords, what were the improvements?

You can use views for that usecase, and you can have multiple available at once.

Reply on news.ycombinator.com