Skip to content

Files

283 lines (205 loc) · 14.2 KB

migration_guide.md

File metadata and controls

283 lines (205 loc) · 14.2 KB

Diesel 2.0 Migration guide

Diesel 2.0 introduces substantial changes to Diesel's inner workings. In some cases this impacts code written using Diesel 1.4.x. This document outlines notable changes and presents potential update strategies. We recommend to start the upgrade by removing the usage of all items that are marked as deprecated in Diesel 1.4.x.

Any code base using migrating to Diesel 2.0 is expected to be affected at least by the following changes:

Users of diesel_migration are additionally affected by the following change:

Users of BoxableExpression might be affected by the following change:

Users of tables containing a column of the type Array<T> are affected by the following change:

Users that implement support for their SQL types or type mappings are affected by the following changes:

no_arg_sql_function! macro is now pending deprecation. Users of the macro are advised to consider define_sql_function! macro.

Users that update generic Diesel code will also be affected by the following changes:

Additionally this release contains many changes for users that implemented a custom backend/connection. We do not provide explicit migration steps but we encourage users to reach out with questions pertaining to these changes.

Mutable Connections required

Diesel now requires mutable access to the Connection to perform any database interaction. The following changes are required for all usages of any Connection type:

- let connection = PgConnection::establish_connection("…")?;
- let result = some_query.load(&connection)?;
+ let mut connection = PgConnection::establish_connection("…")?;
+ let result = some_query.load(&mut connection)?;

We expect this to be a straightforward change as the connection already can execute only one query at a time.

Derive attributes

We have updated all of our Diesel derive attributes to follow the patterns that are used widely in the Rust's ecosystem. This means that all of them need to be wrapped by #[diesel()] now. You can now specify multiple attributes on the same line using , separator.

This is backward compatible and thus all of your old attributes will still work, but with warnings. The attributes can be upgraded by either looking at the warnings or by reading diesel derive documentation reference.

diesel_migration rewrite

We have completely rewritten the diesel_migration crate. As a part of this rewrite all free standing functions are removed from diesel_migration. Equivalent functionality is now provided by the MigrationHarness trait, which is implemented for any Connection type and for HarnessWithOutput. Refer to their documentations for details.

Additionally, this rewrite changed the way we provide migrations. Instead of having our own implementation for file based and embedded migration we now provide a unified MigrationSource trait to abstract over the differences. diesel_migration provides two implementations:

  • FileBasedMigrations, which mirrors the existing behaviour to load raw sql migrations at run time form a specific directory
  • EmbeddedMigrations, which mirrors the existing embed_migrations! macro.

Finally the embed_migrations!() macro itself changed. Instead of generating a magical embedded module it now generates an instance of EmbeddedMigrations, that could be stored in a constant for example.

That means code using embed_migrations!() needs to be changed from

embed_migrations!()

fn run_migration(conn: &PgConnection) {
    embedded_migrations::run(conn).unwrap()
}

to

pub const MIGRATIONS: EmbeddedMigrations = embed_migrations!();

fn run_migration(conn: &PgConnection) {
    conn.run_pending_migrations(MIGRATIONS).unwrap();
}

Changed nullability of operators

We changed the way how we handle the propagation of null values through binary operators. Diesel 1.x always assumed that the result of a binary operation value_a > value_b is not nullable, which does not match the behaviour of the underlying databases. value_a > null may return a NULL value there. With Diesel 2.0 we changed this to match more closely the behaviour of the underlying databases. We expect this change to have the biggest impact on existing usages of BoxableExpression as it may change the resulting sql type there. As a possible workaround for divering sql types there we recommend to use one of the following functions:

  • NullableExpressionMethods::nullable()
  • NullableExpressionMethods::assume_not_null()

Changed nullability of array elements

We changed the inferred SQL type for columns with array types for the PostgreSQL backend. Instead of using Array<ST> we now infer Array<Nullable<ST>> to support arrays containing NULL values. This change implies a change mapping of columns of the corresponding types. It is possible to handle this change using one of the following strategies:

  • Use Vec<Option<T>> as rust side type instead of Vec<T>
  • Manually set the corresponding column to Array<ST> in your schema, to signal that this array does not contain null values. You may want to use the patch_file key for diesel CLI for this.
  • Use #[diesel(deserialize_as = "…")] to explicitly overwrite the deserialization implementation used for this specific struct field. Checkout the documentation of #[derive(Queryable)] for details.

Custom SQL type implementations

We changed how we mark sql types as nullable at type level. For this we replaced the NonNull trait with a more generic SqlType trait, which allows to mark a sql type as (non-) nullable. This may affect custom sql type implementations.

Users that already use the existing #[derive(SqlType)] do not need to change any code. The derive internally generates the correct code after the update. Users that use a manual implementation of NonNull need to replace it with a corresponding SqlType implementation:

- impl NonNull for MyCustomSqlType {}
+ impl SqlType for MyCustomSqlType {
+     type IsNull = diesel::sql_types::is_nullable::NotNull;
+ }

Additionally, the diesel CLI tool was changed so that it automatically generates the Rust side definition of custom SQL types as long as they appear on any table. This feature currently only supports the PostgreSQL backend, as all other supported backends do not support real custom types at SQL level at all.

Changed ToSql implementations

We restructured the way Diesel serializes Rust values to their backend specific representation. This enables us to skip copying the value at all if the specific backend supports writing to a shared buffer. Unfortunately, this feature requires changes to the ToSql trait. This change introduces a lifetime that ensures that a value implementing ToSql outlives the underlying serialisation buffer. Additionally we separated the output buffer type for Sqlite from the type used for PostgreSQL and Mysql.

This has the implication that for generic implementations using a inner existing ToSql implementation you cannot create temporary values anymore and forward them to the inner implementation.

For backend concrete implementations, the following functions allow You to work around this limitation:

  • Output::reborrow() for the Pg and Mysql backend
  • Output::set_value() for the Sqlite backend (Refer to the documentation of SqliteBindValue for accepted values)

Changed FromSql implementations

We changed the raw value representation for both PostgreSQL and MySQL backends, from a & [u8] to an opaque type. This allows us to include additional information like the database side type there. This change enables users to write FromSql implementations that decide dynamically what kind of value was received. The new value types for both backends expose a as_bytes() method to access the underlying byte buffer.

Any affected backend needs to perform the following changes:

impl<DB: Backend> FromSql<YourSqlType, DB> for YourType {
-    fn from_sql(bytes: &[u8]) -> deserialize::Result<Self> {
+    fn from_sql(value: backend::RawValue<'_, DB>) -> deserialize::Result<Self> {
+        let bytes = value.as_bytes();
         // …
     }
}

no_arg_sql_function

The no_arg_sql_function was deprecated without direct replacement. However the define_sql_function! macro gained support for sql functions without argument. This support generates slightly different code. Instead of representing the sql function as zero sized struct, define_sql_function! will generate an ordinary function call without arguments. This requires changing any usage of the generated dsl. This change affects all of the usages of the no_arg_sql_function! in third party crates.

- no_arg_sql_function!(now, sql_types::Timestamp, "Represents the SQL NOW() function");
- 
- diesel::select(now)
 
+ define_sql_function!{
+     /// Represents the SQL NOW() function
+     fn now() -> sql_types::Timestamp;
+ }
+
+ diesel::select(now())

Replacement of NonAggregate with ValidGrouping

Diesel now fully enforces the aggregation rules, which required us to change the way we represent the aggregation at the type system level. This is used to provide group_by support. Diesel's aggregation rules match the semantics of PostgreSQL or MySQL with the ONLY_FULL_GROUP_BY option enabled.

As part of this change we removed the NonAggregate trait in favor of a new, more expressive ValidGrouping trait. Existing implementations of NonAggregate must be replaced with an equivalent ValidGrouping implementation.

The following change shows how to replace an existing implementation with a strictly equivalent implementation.

- impl NonAggregate for MyQueryNode {}
+ impl ValidGrouping<()> for MyQueryNode {
+    type IsAggregate = is_aggregate::No;
+ }

Additional changes may be required to adapt custom query ast implementations to fully support group_by clauses. Refer to the documentation of ValidGrouping for details.

In addition, any occurrence of NonAggregate in trait bounds needs to be replaced. Again, the following change shows the strictly equivalent version:

 where
-     T: NonAggregate,
+     T: ValidGrouping<()>,
+     T::IsAggregate: MixedGrouping<is_aggregate::No, Output = is_aggregate::No>,
+     is_aggregate::No: MixedGrouping<T::IsAggregate, Output = is_aggregate::No>,

Other changes to generics

In addition to the changes listed above, we changed numerous internal details of Diesel. This will have impact on most codebases that include non-trivial generic code abstracting over Diesel. This section tries to list as much of those changes as possible

Removed most of the non-public reachable API

With Diesel 2.0 we removed most of the API which was marked with #[doc(hidden)]. Technically these parts of the API have always been private to Diesel. This change enforces this distinction in stricter way. In addition, some parts of these formerly hidden API are now documented and exposed behind the i-implement-a-third-party-backend-and-opt-into-breaking-changes crate feature. As the name already implies we reserve the right to change these APIs between different Diesel 2.x minor releases, so you should always pin a concrete minor release version if you use these APIs. If you depended on such an API and you cannot find a suitable replacement we invite you to work with us on exposing the corresponding feature as part of the stable API.

Changed structure of the deserialization traits

We changed the internal structure of the FromSqlRow, Queryable and QueryableByName trait family used for deserialization. This change allows us to unify our deserialization code. We hopefully put sufficient wild card implementations in place so that old trait bounds imply the right trait anyway. For cases where this does not hold true, the following changes may be required:

Queryable<ST, DB> is now equivalent to FromSqlRow<ST, DB>. The latter is used as an actual trait bound on the corresponding RunQueryDsl methods. QueryableByName<DB> is now equivalent to FromSqlRow<Untyped, DB>. The latter is used as an actual trait on the corresponding RunQueryDsl methods.

Changed the scope of QueryFragment implementations

With Diesel 2.0, we introduced a way to specialise QueryFragment implementations for specific backend, while providing a generic implementation for other backends. To be able to use this feature in the future we marked existing wild card QueryFragment implementations with an additional DieselReserveSpecialization. Rustc suggests just adding an additional trait bound on this trait. It's not possible to add a bound on this trait without opting into breaking changes and it's almost never required to actually do that. Any occurrence of an error mentioning this trait can simply be fixed by adding a trait bound like follows:

where
    QueryAstNodeMentionedInTheErrorMessage: QueryFragment<BackendType>

This rule has one notable exception: Third party backend implementations. We expect those backends to opt into the i-implement-a-third-party-backend-and-opt-into-breaking-changes feature anyway, as it's otherwise not possible to implement a third party backend.