pub struct Pool<DB: Database>(/* private fields */);
Expand description
An asynchronous pool of SQLx database connections.
Create a pool with Pool::connect or Pool::connect_with and then call Pool::acquire to get a connection from the pool; when the connection is dropped it will return to the pool so it can be reused.
You can also pass &Pool
directly anywhere an Executor
is required; this will automatically
checkout a connection for you.
See the module documentation for examples.
The pool has a maximum connection limit that it will not exceed; if acquire()
is called
when at this limit and all connections are checked out, the task will be made to wait until
a connection becomes available.
You can configure the connection limit, and other parameters, using PoolOptions.
Calls to acquire()
are fair, i.e. fulfilled on a first-come, first-serve basis.
Pool
is Send
, Sync
and Clone
. It is intended to be created once at the start of your
application/daemon/web server/etc. and then shared with all tasks throughout the process’
lifetime. How best to accomplish this depends on your program architecture.
In Actix-Web, for example, you can share a single pool with all request handlers using web::Data.
Cloning Pool
is cheap as it is simply a reference-counted handle to the inner pool state.
When the last remaining handle to the pool is dropped, the connections owned by the pool are
immediately closed (also by dropping). PoolConnection
returned by Pool::acquire and
Transaction
returned by Pool::begin both implicitly hold a reference to the pool for
their lifetimes.
If you prefer to explicitly shutdown the pool and gracefully close its connections (which depending on the database type, may include sending a message to the database server that the connection is being closed), you can call Pool::close which causes all waiting and subsequent calls to Pool::acquire to return Error::PoolClosed, and waits until all connections have been returned to the pool and gracefully closed.
Type aliases are provided for each database to make it easier to sprinkle Pool
through
your codebase:
- [MssqlPool][crate::mssql::MssqlPool] (MSSQL)
- [MySqlPool][crate::mysql::MySqlPool] (MySQL)
- [PgPool][crate::postgres::PgPool] (PostgreSQL)
- [SqlitePool][crate::sqlite::SqlitePool] (SQLite)
§Note: Drop Behavior
Due to a lack of async Drop
, dropping the last Pool
handle may not immediately clean
up connections by itself. The connections will be dropped locally, which is sufficient for
SQLite, but for client/server databases like MySQL and Postgres, that only closes the
client side of the connection. The server will not know the connection is closed until
potentially much later: this is usually dictated by the TCP keepalive timeout in the server
settings.
Because the connection may not be cleaned up immediately on the server side, you may run into errors regarding connection limits if you are creating and dropping many pools in short order.
We recommend calling .close().await
to gracefully close the pool and its connections
when you are done using it. This will also wake any tasks that are waiting on an .acquire()
call, so for long-lived applications it’s a good idea to call .close()
during shutdown.
If you’re writing tests, consider using #[sqlx::test]
which handles the lifetime of
the pool for you.
§Why Use a Pool?
A single database connection (in general) cannot be used by multiple threads simultaneously for various reasons, but an application or web server will typically need to execute numerous queries or commands concurrently (think of concurrent requests against a web server; many or all of them will probably need to hit the database).
You could place the connection in a Mutex
but this will make it a huge bottleneck.
Naively, you might also think to just open a new connection per request, but this has a number of other caveats, generally due to the high overhead involved in working with a fresh connection. Examples to follow.
Connection pools facilitate reuse of connections to amortize these costs, helping to ensure that you’re not paying for them each time you need a connection.
§1. Overhead of Opening a Connection
Opening a database connection is not exactly a cheap operation.
For SQLite, it means numerous requests to the filesystem and memory allocations, while for server-based databases it involves performing DNS resolution, opening a new TCP connection and allocating buffers.
Each connection involves a nontrivial allocation of resources for the database server, usually including spawning a new thread or process specifically to handle the connection, both for concurrency and isolation of faults.
Additionally, database connections typically involve a complex handshake including authentication, negotiation regarding connection parameters (default character sets, timezones, locales, supported features) and upgrades to encrypted tunnels.
If acquire()
is called on a pool with all connections checked out but it is not yet at its
connection limit (see next section), then a new connection is immediately opened, so this pool
does not automatically save you from the overhead of creating a new connection.
However, because this pool by design enforces reuse of connections, this overhead cost
is not paid each and every time you need a connection. In fact, if you set
the min_connections
option in PoolOptions, the pool will
create that many connections up-front so that they are ready to go when a request comes in,
and maintain that number on a best-effort basis for consistent performance.
§2. Connection Limits (MySQL, MSSQL, Postgres)
Database servers usually place hard limits on the number of connections that are allowed open at any given time, to maintain performance targets and prevent excessive allocation of resources, such as RAM, journal files, disk caches, etc.
These limits have different defaults per database flavor, and may vary between different distributions of the same database, but are typically configurable on server start; if you’re paying for managed database hosting then the connection limit will typically vary with your pricing tier.
In MySQL, the default limit is typically 150, plus 1 which is reserved for a user with the
CONNECTION_ADMIN
privilege so you can still access the server to diagnose problems even
with all connections being used.
In MSSQL the only documentation for the default maximum limit is that it depends on the version and server configuration.
In Postgres, the default limit is typically 100, minus 3 which are reserved for superusers (putting the default limit for unprivileged users at 97 connections).
In any case, exceeding these limits results in an error when opening a new connection, which
in a web server context will turn into a 500 Internal Server Error
if not handled, but should
be turned into either 403 Forbidden
or 429 Too Many Requests
depending on your rate-limiting
scheme. However, in a web context, telling a client “go away, maybe try again later” results in
a sub-optimal user experience.
Instead with a connection pool, clients are made to wait in a fair queue for a connection to become available; by using a single connection pool for your whole application, you can ensure that you don’t exceed the connection limit of your database server while allowing response time to degrade gracefully at high load.
Of course, if multiple applications are connecting to the same database server, then you should ensure that the connection limits for all applications add up to your server’s maximum connections or less.
§3. Resource Reuse
The first time you execute a query against your database, the database engine must first turn the SQL into an actionable query plan which it may then execute against the database. This involves parsing the SQL query, validating and analyzing it, and in the case of Postgres 12+ and SQLite, generating code to execute the query plan (native or bytecode, respectively).
These database servers provide a way to amortize this overhead by preparing the query, associating it with an object ID and placing its query plan in a cache to be referenced when it is later executed.
Prepared statements have other features, like bind parameters, which make them safer and more
ergonomic to use as well. By design, SQLx pushes you towards using prepared queries/statements
via the Query API et al. and the query!()
macro et al., for
reasons of safety, ergonomics, and efficiency.
However, because database connections are typically isolated from each other in the database server (either by threads or separate processes entirely), they don’t typically share prepared statements between connections so this work must be redone for each connection.
As with section 1, by facilitating reuse of connections, Pool
helps to ensure their prepared
statements (and thus cached query plans) can be reused as much as possible, thus amortizing
the overhead involved.
Depending on the database server, a connection will have caches for all kinds of other data as well and queries will generally benefit from these caches being “warm” (populated with data).
Implementations§
source§impl<DB: Database> Pool<DB>
impl<DB: Database> Pool<DB>
sourcepub async fn connect(url: &str) -> Result<Self, Error>
pub async fn connect(url: &str) -> Result<Self, Error>
Create a new connection pool with a default pool configuration and the given connection URL, and immediately establish one connection.
Refer to the relevant ConnectOptions
impl for your database for the expected URL format:
- Postgres: [
PgConnectOptions
][crate::postgres::PgConnectOptions] - MySQL: [
MySqlConnectOptions
][crate::mysql::MySqlConnectOptions] - SQLite: [
SqliteConnectOptions
][crate::sqlite::SqliteConnectOptions] - MSSQL: [
MssqlConnectOptions
][crate::mssql::MssqlConnectOptions]
The default configuration is mainly suited for testing and light-duty applications. For production applications, you’ll likely want to make at least few tweaks.
See PoolOptions::new()
for details.
sourcepub async fn connect_with(
options: <DB::Connection as Connection>::Options,
) -> Result<Self, Error>
pub async fn connect_with( options: <DB::Connection as Connection>::Options, ) -> Result<Self, Error>
Create a new connection pool with a default pool configuration and
the given ConnectOptions
, and immediately establish one connection.
The default configuration is mainly suited for testing and light-duty applications. For production applications, you’ll likely want to make at least few tweaks.
See PoolOptions::new()
for details.
sourcepub fn connect_lazy(url: &str) -> Result<Self, Error>
pub fn connect_lazy(url: &str) -> Result<Self, Error>
Create a new connection pool with a default pool configuration and the given connection URL.
The pool will establish connections only as needed.
Refer to the relevant [ConnectOptions
] impl for your database for the expected URL format:
- Postgres: [
PgConnectOptions
][crate::postgres::PgConnectOptions] - MySQL: [
MySqlConnectOptions
][crate::mysql::MySqlConnectOptions] - SQLite: [
SqliteConnectOptions
][crate::sqlite::SqliteConnectOptions] - MSSQL: [
MssqlConnectOptions
][crate::mssql::MssqlConnectOptions]
The default configuration is mainly suited for testing and light-duty applications. For production applications, you’ll likely want to make at least few tweaks.
See PoolOptions::new()
for details.
sourcepub fn connect_lazy_with(
options: <DB::Connection as Connection>::Options,
) -> Self
pub fn connect_lazy_with( options: <DB::Connection as Connection>::Options, ) -> Self
Create a new connection pool with a default pool configuration and
the given ConnectOptions
.
The pool will establish connections only as needed.
The default configuration is mainly suited for testing and light-duty applications. For production applications, you’ll likely want to make at least few tweaks.
See PoolOptions::new()
for details.
sourcepub fn acquire(
&self,
) -> impl Future<Output = Result<PoolConnection<DB>, Error>> + 'static
pub fn acquire( &self, ) -> impl Future<Output = Result<PoolConnection<DB>, Error>> + 'static
Retrieves a connection from the pool.
The total time this method is allowed to execute is capped by
PoolOptions::acquire_timeout
.
If that timeout elapses, this will return Error::PoolClosed
.
§Note: Cancellation/Timeout May Drop Connections
If acquire
is cancelled or times out after it acquires a connection from the idle queue or
opens a new one, it will drop that connection because we don’t want to assume it
is safe to return to the pool, and testing it to see if it’s safe to release could introduce
subtle bugs if not implemented correctly. To avoid that entirely, we’ve decided to not
gracefully handle cancellation here.
However, if your workload is sensitive to dropped connections such as using an in-memory
SQLite database with a pool size of 1, you can pretty easily ensure that a cancelled
acquire()
call will never drop connections by tweaking your PoolOptions
:
- Set
test_before_acquire(false)
- Never set
before_acquire
orafter_connect
.
This should eliminate any potential .await
points between acquiring a connection and
returning it.
sourcepub fn try_acquire(&self) -> Option<PoolConnection<DB>>
pub fn try_acquire(&self) -> Option<PoolConnection<DB>>
Attempts to retrieve a connection from the pool if there is one available.
Returns None
immediately if there are no idle connections available in the pool
or there are tasks waiting for a connection which have yet to wake.
sourcepub async fn begin(&self) -> Result<Transaction<'static, DB>, Error>
pub async fn begin(&self) -> Result<Transaction<'static, DB>, Error>
Retrieves a connection and immediately begins a new transaction.
sourcepub async fn try_begin(&self) -> Result<Option<Transaction<'static, DB>>, Error>
pub async fn try_begin(&self) -> Result<Option<Transaction<'static, DB>>, Error>
Attempts to retrieve a connection and immediately begins a new transaction if successful.
sourcepub fn close(&self) -> impl Future<Output = ()> + '_
pub fn close(&self) -> impl Future<Output = ()> + '_
Shut down the connection pool, immediately waking all tasks waiting for a connection.
Upon calling this method, any currently waiting or subsequent calls to Pool::acquire
and
the like will immediately return Error::PoolClosed
and no new connections will be opened.
Checked-out connections are unaffected, but will be gracefully closed on-drop
rather than being returned to the pool.
Returns a Future
which can be .await
ed to ensure all connections are
gracefully closed. It will first close any idle connections currently waiting in the pool,
then wait for all checked-out connections to be returned or closed.
Waiting for connections to be gracefully closed is optional, but will allow the database server to clean up the resources sooner rather than later. This is especially important for tests that create a new pool every time, otherwise you may see errors about connection limits being exhausted even when running tests in a single thread.
If the returned Future
is not run to completion, any remaining connections will be dropped
when the last handle for the given pool instance is dropped, which could happen in a task
spawned by Pool
internally and so may be unpredictable otherwise.
.close()
may be safely called and .await
ed on multiple handles concurrently.
sourcepub fn is_closed(&self) -> bool
pub fn is_closed(&self) -> bool
Returns true
if .close()
has been called on the pool, false
otherwise.
sourcepub fn close_event(&self) -> CloseEvent ⓘ
pub fn close_event(&self) -> CloseEvent ⓘ
Get a future that resolves when Pool::close()
is called.
If the pool is already closed, the future resolves immediately.
This can be used to cancel long-running operations that hold onto a PoolConnection
so they don’t prevent the pool from closing (which would otherwise wait until all
connections are returned).
§Examples
These examples use Postgres and Tokio, but should suffice to demonstrate the concept.
Do something when the pool is closed:
use sqlx::PgPool;
let pool = PgPool::connect("postgresql://...").await?;
let pool2 = pool.clone();
tokio::spawn(async move {
// Demonstrates that `CloseEvent` is itself a `Future` you can wait on.
// This lets you implement any kind of on-close event that you like.
pool2.close_event().await;
println!("Pool is closing!");
// Imagine maybe recording application statistics or logging a report, etc.
});
// The rest of the application executes normally...
// Close the pool before the application exits...
pool.close().await;
Cancel a long-running operation:
use sqlx::{Executor, PgPool};
let pool = PgPool::connect("postgresql://...").await?;
let pool2 = pool.clone();
tokio::spawn(async move {
pool2.close_event().do_until(async {
// This statement normally won't return for 30 days!
// (Assuming the connection doesn't time out first, of course.)
pool2.execute("SELECT pg_sleep('30 days')").await;
// If the pool is closed before the statement completes, this won't be printed.
// This is because `.do_until()` cancels the future it's given if the
// pool is closed first.
println!("Waited!");
}).await;
});
// This normally wouldn't return until the above statement completed and the connection
// was returned to the pool. However, thanks to `.do_until()`, the operation was
// cancelled as soon as we called `.close().await`.
pool.close().await;
sourcepub fn size(&self) -> u32
pub fn size(&self) -> u32
Returns the number of connections currently active. This includes idle connections.
sourcepub fn num_idle(&self) -> usize
pub fn num_idle(&self) -> usize
Returns the number of connections active and idle (not in use).
As of 0.6.0, this has been fixed to use a separate atomic counter and so should be fine to call even at high load.
This previously called [crossbeam::queue::ArrayQueue::len()
] which waits for the head and
tail pointers to be in a consistent state, which may never happen at high levels of churn.
sourcepub fn connect_options(&self) -> Arc<<DB::Connection as Connection>::Options>
pub fn connect_options(&self) -> Arc<<DB::Connection as Connection>::Options>
Gets a clone of the connection options for this pool
sourcepub fn set_connect_options(
&self,
connect_options: <DB::Connection as Connection>::Options,
)
pub fn set_connect_options( &self, connect_options: <DB::Connection as Connection>::Options, )
Updates the connection options this pool will use when opening any future connections. Any existing open connection in the pool will be left as-is.
sourcepub fn options(&self) -> &PoolOptions<DB>
pub fn options(&self) -> &PoolOptions<DB>
Get the options for this pool
Trait Implementations§
source§impl<'a, DB: Database> Acquire<'a> for &Pool<DB>
impl<'a, DB: Database> Acquire<'a> for &Pool<DB>
type Database = DB
type Connection = PoolConnection<DB>
fn acquire(self) -> BoxFuture<'static, Result<Self::Connection, Error>>
fn begin(self) -> BoxFuture<'static, Result<Transaction<'a, DB>, Error>>
source§impl<DB: Database> Clone for Pool<DB>
impl<DB: Database> Clone for Pool<DB>
Returns a new Pool tied to the same shared connection pool.
source§impl<'p, DB: Database> Executor<'p> for &Pool<DB>where
for<'c> &'c mut DB::Connection: Executor<'c, Database = DB>,
impl<'p, DB: Database> Executor<'p> for &Pool<DB>where
for<'c> &'c mut DB::Connection: Executor<'c, Database = DB>,
type Database = DB
source§fn fetch_many<'e, 'q: 'e, E>(
self,
query: E,
) -> BoxStream<'e, Result<Either<DB::QueryResult, DB::Row>, Error>>
fn fetch_many<'e, 'q: 'e, E>( self, query: E, ) -> BoxStream<'e, Result<Either<DB::QueryResult, DB::Row>, Error>>
source§fn fetch_optional<'e, 'q: 'e, E>(
self,
query: E,
) -> BoxFuture<'e, Result<Option<DB::Row>, Error>>
fn fetch_optional<'e, 'q: 'e, E>( self, query: E, ) -> BoxFuture<'e, Result<Option<DB::Row>, Error>>
source§fn prepare_with<'e, 'q: 'e>(
self,
sql: &'q str,
parameters: &'e [<Self::Database as Database>::TypeInfo],
) -> BoxFuture<'e, Result<<Self::Database as HasStatement<'q>>::Statement, Error>>
fn prepare_with<'e, 'q: 'e>( self, sql: &'q str, parameters: &'e [<Self::Database as Database>::TypeInfo], ) -> BoxFuture<'e, Result<<Self::Database as HasStatement<'q>>::Statement, Error>>
source§fn execute<'e, 'q: 'e, E>(
self,
query: E,
) -> BoxFuture<'e, Result<<Self::Database as Database>::QueryResult, Error>>
fn execute<'e, 'q: 'e, E>( self, query: E, ) -> BoxFuture<'e, Result<<Self::Database as Database>::QueryResult, Error>>
source§fn execute_many<'e, 'q: 'e, E>(
self,
query: E,
) -> BoxStream<'e, Result<<Self::Database as Database>::QueryResult, Error>>
fn execute_many<'e, 'q: 'e, E>( self, query: E, ) -> BoxStream<'e, Result<<Self::Database as Database>::QueryResult, Error>>
source§fn fetch<'e, 'q: 'e, E>(
self,
query: E,
) -> BoxStream<'e, Result<<Self::Database as Database>::Row, Error>>
fn fetch<'e, 'q: 'e, E>( self, query: E, ) -> BoxStream<'e, Result<<Self::Database as Database>::Row, Error>>
source§fn fetch_all<'e, 'q: 'e, E>(
self,
query: E,
) -> BoxFuture<'e, Result<Vec<<Self::Database as Database>::Row>, Error>>
fn fetch_all<'e, 'q: 'e, E>( self, query: E, ) -> BoxFuture<'e, Result<Vec<<Self::Database as Database>::Row>, Error>>
Vec
.Auto Trait Implementations§
impl<DB> Freeze for Pool<DB>
impl<DB> !RefUnwindSafe for Pool<DB>
impl<DB> Send for Pool<DB>
impl<DB> Sync for Pool<DB>
impl<DB> Unpin for Pool<DB>
impl<DB> !UnwindSafe for Pool<DB>
Blanket Implementations§
source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
source§impl<T> CloneToUninit for Twhere
T: Clone,
impl<T> CloneToUninit for Twhere
T: Clone,
source§default unsafe fn clone_to_uninit(&self, dst: *mut T)
default unsafe fn clone_to_uninit(&self, dst: *mut T)
clone_to_uninit
)source§impl<T> Instrument for T
impl<T> Instrument for T
source§fn instrument(self, span: Span) -> Instrumented<Self>
fn instrument(self, span: Span) -> Instrumented<Self>
source§fn in_current_span(self) -> Instrumented<Self>
fn in_current_span(self) -> Instrumented<Self>
source§impl<T> IntoEither for T
impl<T> IntoEither for T
source§fn into_either(self, into_left: bool) -> Either<Self, Self> ⓘ
fn into_either(self, into_left: bool) -> Either<Self, Self> ⓘ
self
into a Left
variant of Either<Self, Self>
if into_left
is true
.
Converts self
into a Right
variant of Either<Self, Self>
otherwise. Read moresource§fn into_either_with<F>(self, into_left: F) -> Either<Self, Self> ⓘ
fn into_either_with<F>(self, into_left: F) -> Either<Self, Self> ⓘ
self
into a Left
variant of Either<Self, Self>
if into_left(&self)
returns true
.
Converts self
into a Right
variant of Either<Self, Self>
otherwise. Read more