axum-impl-database

Axum Database Access with sqlx

Overview

An Axum service talks to a SQL database through a sqlx connection pool. The pool is an expensive resource: every connection is a real TCP socket (usually with TLS) plus a server-side session. The pool MUST be created exactly once at startup and shared with every handler through router state.

sqlx exposes the pool as Pool<DB>, with the type aliases PgPool for Postgres, MySqlPool for MySQL, and SqlitePool for SQLite. This skill uses Postgres; the patterns apply uniformly to the other databases by substituting the type.

Three rules dominate correct database code in Axum:

1. Build the pool once in main and hand it to the router with .with_state(pool). NEVER call PgPoolOptions::new()...connect() inside a handler.
2. Pool is already a cheap reference-counted handle. NEVER wrap a PgPool in Arc, and NEVER wrap it in a Mutex.
3. ALWAYS set acquire_timeout. Without it, a request that arrives when the pool is saturated hangs forever instead of failing fast.

Quick Reference

Need	API	Notes
Build the pool	PgPoolOptions::new().max_connections(n).acquire_timeout(d).connect(url).await	connect is async, resolves once connected
Share the pool	Router::new()....with_state(pool)	pass PgPool directly, no Arc
Read the pool in a handler	State(pool): State<PgPool>	Pool is cloned per request, cheaply
Run a static query	query! / query_as! / query_scalar!	SQL checked at compile time
Run a dynamic query	query() / query_as() / query_scalar()	runtime, no compile-time check
Exactly one row	.fetch_one(&pool).await	errors if zero rows
Zero or one row	.fetch_optional(&pool).await	returns Option
All rows	.fetch_all(&pool).await	returns Vec
INSERT, UPDATE, DELETE	.execute(&pool).await	returns affected-row count
Start a transaction	pool.begin().await?	returns Transaction<'static, DB>
Commit or roll back	tx.commit().await? / tx.rollback().await?	drop without commit auto-rolls-back
Dedicated connection	DatabaseConnection extractor	one checked-out connection per request
Connection string	std::env::var("DATABASE_URL")	NEVER a string literal

Required crates:

[dependencies]
axum = "0.8"                                  # axum 0.8
# axum = "0.7"                                # axum 0.7
sqlx = { version = "0.8", features = ["runtime-tokio", "postgres"] }
tokio = { version = "1", features = ["full"] }

For the Axum 0.7 extractor form, also add async-trait. See Pattern 3.

Decision Trees

Where do I create the pool?

Need a database connection in a handler?
|
+- Pool already built in main and in router state?
|  `- YES: read it with State(pool): State<PgPool>. Done.
|
`- NO pool in state yet
   `- Build it ONCE in main with PgPoolOptions, then .with_state(pool).
      NEVER build a pool inside a handler.

Pool reference, dedicated connection, or transaction?

What does the handler do with the database?
|
+- One query, or several independent queries
|  `- State(pool): State<PgPool>; run each query against &pool.
|     Each .await acquires and returns a connection automatically.
|
+- Several statements that must share ONE connection
|  `- DatabaseConnection extractor; run against &mut *conn.
|
`- Several statements that must be atomic (all-or-nothing)
   `- Transaction: pool.begin(), run against &mut *tx, then commit.

query! macro or query() function?

Is the SQL string known at compile time (a string literal)?
|
+- YES: use query! / query_as! / query_scalar!.
|       SQL syntax and input/output types are checked when you build.
|
`- NO, the SQL is assembled at runtime
   `- use query() / query_as() / query_scalar().
      No compile-time check; bind every input with .bind() so the
      value is never concatenated into the SQL string.

Should I wrap the pool?

Tempted to write Arc<PgPool> or Mutex<PgPool>?
|
+- Arc<PgPool>   -> NEVER. Pool is already reference-counted. Pass PgPool.
+- Mutex<PgPool> -> NEVER. The pool is internally synchronized and hands
|                   out connections concurrently. A Mutex serializes every
|                   query and destroys concurrency.
`- Plain PgPool in state -> CORRECT.

Patterns

Pattern 1: Create the pool once, share via with_state

Build the pool in main, read DATABASE_URL from the environment, and pass the PgPool straight to .with_state(...).

use axum::{routing::get, Router};
use sqlx::postgres::{PgPool, PgPoolOptions};
use std::time::Duration;

#[tokio::main]
async fn main() {
    // Fail fast on a missing connection string. NEVER a string literal.
    let db_url = std::env::var("DATABASE_URL")
        .expect("DATABASE_URL must be set");

    let pool = PgPoolOptions::new()
        .max_connections(5)
        .acquire_timeout(Duration::from_secs(3))
        .connect(&db_url)
        .await
        .expect("can't connect to database");

    let app = Router::new()
        .route("/users/count", get(count_users))
        .with_state(pool); // PgPool passed directly: no Arc, no Mutex

    let listener = tokio::net::TcpListener::bind("0.0.0.0:3000")
        .await
        .unwrap();
    axum::serve(listener, app).await.unwrap();
}

PgPool is the state type. Axum clones the state once per request; cloning a Pool is an Arc-style refcount bump, so every handler shares the same connections. See axum-core-state for substate access with #[derive(FromRef)].

Pattern 2: Query against the pool reference

&Pool<DB> implements the sqlx Executor trait, so any query runs against &pool directly. Each .await transparently acquires a connection, runs the statement, and returns the connection to the pool.

use axum::extract::State;
use axum::http::StatusCode;
use sqlx::postgres::PgPool;

async fn count_users(
    State(pool): State<PgPool>,
) -> Result<String, (StatusCode, String)> {
    let count: i64 = sqlx::query_scalar("select count(*) from users")
        .fetch_one(&pool)
        .await
        .map_err(internal_error)?; // NEVER .unwrap() here
    Ok(format!("{count} users"))
}

// Maps any std::error::Error to a 500 response.
fn internal_error<E>(err: E) -> (StatusCode, String)
where
    E: std::error::Error,
{
    (StatusCode::INTERNAL_SERVER_ERROR, err.to_string())
}

ALWAYS return Result<T, E> and propagate query errors with ?. A .unwrap() on a query result turns a missing row or a dropped connection into a panic. For a typed application error that distinguishes 404 from 500, see axum-errors-handling.

Pattern 3: The DatabaseConnection extractor

When a handler runs several statements that must share one physical connection, check out a dedicated connection per request with a custom FromRequestParts extractor. The extractor is bounded by PgPool: FromRef<S>, so it works for any state that exposes a PgPool.

The impl block differs by Axum version: Axum 0.8 uses a native async fn in the trait; Axum 0.7 needs the #[async_trait] attribute on the impl block.

// axum 0.8
use axum::extract::{FromRef, FromRequestParts};
use axum::http::request::Parts;
use axum::http::StatusCode;
use sqlx::postgres::PgPool;

struct DatabaseConnection(sqlx::pool::PoolConnection<sqlx::Postgres>);

impl<S> FromRequestParts<S> for DatabaseConnection
where
    PgPool: FromRef<S>,
    S: Send + Sync,
{
    type Rejection = (StatusCode, String);

    async fn from_request_parts(
        _parts: &mut Parts,
        state: &S,
    ) -> Result<Self, Self::Rejection> {
        let pool = PgPool::from_ref(state);
        let conn = pool.acquire().await.map_err(internal_error)?;
        Ok(Self(conn))
    }
}

// axum 0.7
#[async_trait::async_trait]
impl<S> FromRequestParts<S> for DatabaseConnection
where
    PgPool: FromRef<S>,
    S: Send + Sync,
{
    type Rejection = (StatusCode, String);

    async fn from_request_parts(
        _parts: &mut Parts,
        state: &S,
    ) -> Result<Self, Self::Rejection> {
        let pool = PgPool::from_ref(state);
        let conn = pool.acquire().await.map_err(internal_error)?;
        Ok(Self(conn))
    }
}

The checked-out PoolConnection<Postgres> is used as &mut *conn:

async fn run_on_one_connection(
    DatabaseConnection(mut conn): DatabaseConnection,
) -> Result<String, (StatusCode, String)> {
    sqlx::query_scalar("select 'hello world from pg'")
        .fetch_one(&mut *conn)
        .await
        .map_err(internal_error)
}

acquire's wait time is bounded by acquire_timeout. When the pool is at capacity, callers queue fairly. The connection returns to the pool when conn is dropped at the end of the handler. The full DatabaseConnection extractor is in references/examples.md.

Pattern 4: Transactions

pool.begin().await? starts a transaction and returns a Transaction<'static, DB>. Run queries against &mut *tx, then call tx.commit().await? to persist or tx.rollback().await? to discard.

The decisive safety property: if a Transaction is dropped without commit, it rolls back automatically. An early return or a ?-propagated error inside the handler safely abandons the transaction with no half-applied state. This is why transactions in Axum handlers combine safely with the ? operator.

async fn transfer(
    State(pool): State<PgPool>,
) -> Result<StatusCode, (StatusCode, String)> {
    let mut tx = pool.begin().await.map_err(internal_error)?;

    sqlx::query("update accounts set balance = balance - 100 where id = 1")
        .execute(&mut *tx)
        .await
        .map_err(internal_error)?;
    sqlx::query("update accounts set balance = balance + 100 where id = 2")
        .execute(&mut *tx)
        .await
        .map_err(internal_error)?;

    // If either statement above fails, `?` returns early, `tx` is dropped,
    // and the transaction rolls back. No half-applied state ever persists.
    tx.commit().await.map_err(internal_error)?;
    Ok(StatusCode::OK)
}

pool.try_begin() returns Result<Option<Transaction>, Error> for a non-waiting attempt.

Pattern 5: Query macros versus runtime functions

Compile-time-checked macros connect to a real database at build time (via DATABASE_URL or an offline .sqlx query cache) and verify SQL syntax plus input and output types during compilation:

• query! static query with verified column types.
• query_as! static query mapped to an explicitly named output struct.
• query_scalar! static query returning a single column.

Runtime functions perform no compile-time checking and are used for dynamic SQL. They are still prepared statements, transparently cached:

• query() runtime query; bind inputs with .bind(value).
• query_as() runtime query mapped onto a type that derives #[derive(sqlx::FromRow)].
• query_scalar() runtime query returning the first column.

ALWAYS PREFER the ! macros when the SQL is a static string literal, because errors surface at compile time. Use the runtime functions ONLY when the query string is assembled at runtime, and bind every value with .bind(...) so input is never concatenated into the SQL text. Full macro and function examples are in references/examples.md.

Pattern 6: Pool sizing

max_connections is the hard upper bound on concurrent database connections per application instance. Size it so that (number of app instances) x max_connections stays comfortably below the database server's own global connection limit, leaving headroom for migrations, admin sessions, and monitoring tools.

A typical starting point is near the database server's CPU core count, not a large number. Oversizing the pool increases lock contention on the database and rarely improves throughput.

ALWAYS pair max_connections with acquire_timeout. Without acquire_timeout, a request that arrives when the pool is saturated waits indefinitely; with it, acquire fails after the timeout and the handler can map the error to a 503, giving fast, observable backpressure instead of a silent hang. min_connections keeps a warm floor of connections; max_lifetime and idle_timeout recycle connections so stale or load-balancer-dropped sockets do not accumulate. On graceful shutdown, call pool.close().await so in-flight connections drain cleanly. See axum-core-async-performance for the broader concurrency picture.

Reference Links

• references/methods.md complete API signatures: PgPoolOptions builder, Pool methods, query macros and functions, execution methods, Transaction.
• references/examples.md working version-annotated code: full main.rs with graceful shutdown, query_as CRUD with FromRow, the complete DatabaseConnection extractor for 0.7 and 0.8, a transaction handler.
• references/anti-patterns.md real mistakes and why each one fails: per-request pools, Arc<PgPool>, Mutex<PgPool>, .unwrap() on queries, missing acquire_timeout, the wrong extractor form per version.

Related skills: axum-core-state (router state and FromRef), axum-core-async-performance (concurrency and blocking work), axum-errors-handling (typed AppError with IntoResponse).

Sources: https://docs.rs/sqlx/latest/sqlx/, https://docs.rs/sqlx/latest/sqlx/struct.Pool.html, https://docs.rs/sqlx/latest/sqlx/pool/struct.PoolOptions.html, https://github.com/tokio-rs/axum/tree/main/examples/sqlx-postgres.