ztp/docs/05-connecting-and-saving.md

13 KiB
Raw Blame History

+++ title = "Databases: Connecting and Using" date = 2023-03-20T17:38:12Z weight = 4 +++

Chapter 3.8: Databases

First, we're gonna expand the configuration we defined previously. The one thing we will require is the password, although eventually I'm going to make that a command-line option using Clap.

#[derive(serde::Deserialize)]
#[serde(default)]
pub struct DatabaseSettings {
    pub username: String,
    pub password: String,
    pub host: String,
    pub port: u16,
    pub database: String,
}

#[derive(serde::Deserialize)]
#[serde(default)]
pub struct Settings {
    pub database: DatabaseSettings,
    pub port: u16,
}

impl Default for DatabaseSettings {
    fn default() -> Self {
        DatabaseSettings {
            username: "newsletter".to_string(),
            password: "".to_string(),
            host: "localhost".to_string(),
            port: 5432,
            database: "newsletter".to_string(),
        }
    }
}

impl Default for Settings {
    fn default() -> Self {
        Settings {
            port: 3001,
            database: DatabaseSettings::default(),
        }
    }
}

Cleanup on Aisle Three

The code's beginning to get a bit messsy. So let's re-arrange. I already have the configuration handler in its own file, but let's clean up the routes. I'm going to take the functions I've already defined and put them in a subfolder, src/routes, and then for each one I'll make a new file. For example:

use axum::{http::StatusCode, response::IntoResponse};

pub(crate) async fn health_check() -> impl IntoResponse {
    (StatusCode::OK, ())
}

Just a note: ignore your IDE's "advice" to remove the unneeded async; Axum will not compile this module correctly, IntoResponse requires that it be an async function. I don't know why rust-analyzer failed to pick that up.

Now make a src/routes.rs (yes, the same name as the folder). You can activate and de-activate routes at will:

use axum::{
    routing::{get, post},
    Router,
};

mod greet;
use greet::greet;

mod health_check;
use health_check::health_check;

mod index;
use index::index;

mod subscribe;
use subscribe::subscribe;

pub(crate) fn app() -> Router {
    Router::new()
        .route("/", get(index))
        .route("/subscriptions", post(subscribe))
        .route("/:name", get(greet))
        .route("/health_check", get(health_check))
}

You don't need the use clauses, you could just say index::index, but I kinda like the way this looks, even with the repetition. It makes it clear what I'm doing and what I'm looking at. This is Rust 2021, where the presence of a file and a folder with the same name indicate a module with submodules. In prior editions of rust, instead of routes.rs, we would have used routes/mod.rs. I'm not entirely sure which I like better.

Now that that's done, in our lib.rs, we use the routes:

use std::net::SocketAddr;
mod configuration;
use configuration::get_configuration;
mod routes;
use routes::app;

pub async fn run() {
    let configuration = get_configuration().unwrap();
    let addr = SocketAddr::from(([127, 0, 0, 1], configuration.port));
    tracing::info!("listening on {}", addr);
    axum::Server::bind(&addr)
        .serve(app().into_make_service())
        .await
        .unwrap()
}

All of the Axum imports have been removed, since they're part of the routes. Nice, clean and clear code. At this point, even cargo clippy is happy.

The Database String

Connecting to a database these days involves a URL. Let's make it easy to generate one. Now, we could implement the trait std::string::ToString, or Display, but that would conflict with any debugging information. Instead, let's define a new method in configuration.rs:

impl DatabaseSettings {
    pub fn url(&self) -> String {
        if self.password.len() == 0 {
            format!(
                "postgres://{}@{}:{}/{}",
                self.username, self.host, self.port, self.database
            )
        } else {
            format!(
                "postgres://{}:{}@{}:{}/{}",
                self.username, self.password, self.host, self.port, self.database
            )
        }
    }
}

It is possible (unwise, but possible) to connect to a Postgres database with no password. In a dynamic language this would have been less wordy, but I can't complain about Rust being a little extra verbose in exchange for being a lot more precise.

Providing the database to the application

Since we're using Postgres, in lib.rs we'll add use sqlx::postgres::PgPoolOptions;, which will allow us to create a pool of workers and limit how many there are. In the run() function, we're going to extract the app() definition and use it:

use axum::Extension;
use sqlx::postgres::PgPoolOptions;

// in: fn run()

    let pool = PgPoolOptions::new()
        .max_connections(50)
        .connect(&configuration.database.url())
        .await
        .expect("could not connect to database_url");

    let routes = app().layer(Extension(pool));

    tracing::info!("listening on {}", addr);
    axum::Server::bind(&addr)
        .serve(routes.into_make_service())
        .await
        .unwrap()

Now we have a configuration that is runnable. If you run it as-is, it will fail because, if you set up your Postgres database correctly, there is no password. Create a ztp.config.yaml file (or JSON, or TOML, or whatever floats your boat) and establish the password:

database:
    password: redacted

At one point, I had this file named only ztp.config, and unfortunately it crashed. Adding #[derive(Debug)] to my configuration structs helped show that the password wasn't getting set, but it took me a few minutes to realize that I'd forgotten the extension so Config didn't know how to read the file. Naming it ztp.config.yaml fixed it, and this illustrates just how useful TOML can be in circumstances like this.

Pulling in a subscription

This turned out to be a bit of a beast, and I'm incredibly indebted to Carlos Armando Marcano Vargas and his blog where he documented much of what I needed to make this work with Axum.

In the migrations file, we specified that the primary key of our subscriptions table would be a UUID, and that we were going to record when the subscription happened. To do this, first we have to make sure that we have the Uuid and Chrono crates, and that they're all hooked up for proper serialization:

chrono = { version = "0.4.24", features = ["serde"] }
uuid = { version = "1.3.0", features = ["v4", "serde"] }
sqlx = { version = "0.6.3", features = ["runtime-tokio-native-tls", "macros",
         "postgres", "uuid", "chrono"] }

Both of the new creates need serde as a feature and we'll be using UUID4 for our primary key.

We'll also need a full subscription object to store, and then a way to generate it from a new subscription:

#[derive(serde::Deserialize)]
pub(crate) struct NewSubscription {
    pub email: String,
    pub name: String,
}

#[derive(serde::Deserialize, serde::Serialize, sqlx::FromRow)]
struct Subscription {
    pub id: Uuid,
    pub email: String,
    pub name: String,
    pub subscribed_at: DateTime<Utc>,
}

impl From<&NewSubscription> for Subscription {
    fn from(s: &NewSubscription) -> Self {
        Subscription {
            id: Uuid::new_v4(),
            email: s.email.clone(),
            name: s.name.clone(),
            subscribed_at: Utc::now(),
        }
    }
}

I've renamed "Subscription" to "NewSubscription" to distinguish it from the thing we're going to reliably keep in our database, which has all the fields.

And now we're going to re-write the subscribe() function to actually talk to the database. All of those parens around the payload are necessary to help Rust understand what we want to borrow. You'll note that it's payload.0, not just payload; Form objects can have multiple instances of their content.

pub(crate) async fn subscribe(Extension(pool): Extension<PgPool>,
                              payload:  Form<NewSubscription>) ->
    impl IntoResponse {
    let sql = r#"INSERT INTO subscriptions (id, email, name,
                subscribed_at) VALUES ($1, $2, $3, $4);"#.to_string();
    let subscription: Subscription = (&(payload.0)).into();

    let _ = sqlx::query(&sql)
        .bind(subscription.id)
        .bind(subscription.email)
        .bind(subscription.name)
        .bind(subscription.subscribed_at)
        .execute(&pool)
        .await
        .expect("Failed for some reason?");

    (StatusCode::OK, ())
}

Alternatively, I could have written:

    let subscription = Subscription::from(&(payload.0));

Either way works.

Better error handling

That unused .expect() has Tokio panicking if I give it a duplicate email address, as one of the constraints in our original migration reads:

   email text not null unique,

We want to catch and handle that error. We also want to fix something that's been sitting in the "TODO" list for awhile: make sure that when the subscription form isn't fully filled out, we return a straight 400. That's actually the easier part: we make the payload optional and return our default BAD_REQUEST if it's not there:

pub(crate) async fn subscribe(
    Extension(pool): Extension<PgPool>,
    payload: Option<Form<NewSubscription>>,
) -> impl IntoResponse {
    if let Some(payload) = payload {
    ...
        (StatusCode::OK, ())
    } else {
        (StatusCode::BAD_REQUEST, ())
    }

So, handling the errors is going to be interesting. Going above and beyond, it's time to use thiserror.

Rust has an excellent error handling story, with the Result<T, impl Error>, although mostly this is just a way of distinguishing and annotating errors in a robust, type-safe manner; it's still up to the programmer to handle errors. A library has a set of error states that it can return, and thiserror is a set of macros that automatically annotate your list of errors with messages and values.

Right now we have two possible errors: the form data passed in was incomplete, or the email address submitted already exists. (We're not even checking the email address for validity, just existence.) So let's create those errors and provide them with labels. We'll just put this in a file named errors.rs, in the root of the crate; that seems to be where everyone puts it, and it's a nice convention.

#[derive(Error, Debug)]
pub enum ZTPError {
    #[error("Form data was incomplete")]
    FormIncomplete,
    #[error("Email address already subscribed")]
    DuplicateEmail,
}

And now the error will always be "400: Bad Response", but the text of the message will be much more descriptive of what went wrong.

We have to edit the subscribe function to handle these changes: first, the return type must change, and then the errors must be included:

pub(crate) async fn subscribe(
    Extension(pool): Extension<PgPool>,
    payload: Option<Form<NewSubscription>>,
) -> Result<(StatusCode, ()), ZTPError> {
     // ...
            .map_or(Ok((StatusCode::OK, ())), |_| Err(ZTPError::DuplicateEmail))
    } else {
        Err(ZTPError::FormIncomplete)
    }

We've changed the function to take an Option<Form<>>; this way, if we don't get a form that deserialized properly we'll be able to say "That wasn't a good form," and we've changed the result so that we can return our error. Then we changed our two edge cases to report the errors.

Axum doesn't know what a ZTPError is, but we can easily make our errors to something Axum does understand. We'll just have to implement our own IntoResponse, and we'll put it in errors.rs:

impl IntoResponse for ZTPError {
    fn into_response(self) -> axum::response::Response {
        let (status, error_message) = match self {
            Self::FormIncomplete => (StatusCode::BAD_REQUEST, self.to_string()),
            Self::DuplicateEmail => (StatusCode::BAD_REQUEST, self.to_string()),
        };
        (status, Json(json!({ "error": error_message }))).into_response()
    }
}

... and with that, we can now change all of the tests to watch for 400 codes, as that ancient TODO said, and it will all just work.

End of Chapter Three

Holy Chao, we've made it through the "writing the application" stuff, and a lot of what was in the book is... no longer quite so relevant. There's whole sections on getting Actix to type-match with your code that Axum and 2021 Rust make more or less irrelevant, and by using features like thiserror and dbmate we've managed to route around many of the bulkier difficulties in the book.

The next chapter is about logging and monitoring. ɡɡ