11 releases

new 0.1.18 May 18, 2024
0.1.17 May 18, 2024
0.1.16 Apr 13, 2024
0.1.15 Dec 9, 2023
0.1.8 Jun 29, 2023

#2018 in Database interfaces

Download history 2/week @ 2024-02-19 9/week @ 2024-02-26 9/week @ 2024-04-01 140/week @ 2024-04-08 13/week @ 2024-04-15 179/week @ 2024-05-13

179 downloads per month
Used in spaghettinuum

ISC license

126 lines


Good-ormning is an ORM, probably? In a nutshell:

  1. Define schemas and queries in build.rs
  2. Good-ormning generates a function to set up/migrate the database
  3. Good-ormning generates functions for each query

Why you want it

  • You want end to end type safety, from table definition through queries, across versions
  • You want to do everything in Rust, you don't want to need to spin up a database and run SQL manually


  • No macros
  • No generics
  • No traits (okay, simple traits for custom types to help guide implementations only)
  • No boilerplate
  • Automatic migrations, no migration-schema mismatches
  • Query parameter type checking - no runtime errors due to parameter types, counts, or ordering
  • Query logic type checking via a query simulation
  • Query result type checking - no runtime errors due to result types, counts, or ordering
  • Fast to generate, minimum runtime overhead

Like other Rust ORMs, Good-ormning doesn't abstract away from actual database workflows, but instead aims to enhance type checking with normal SQL.

See Comparisons, below, for information on how Good-ormning differs from other Rust ORMs.

Current status

  • Basic features work, this works for my basic uses
  • Moderate test coverage
  • Missing advanced features - let me know if there's something you want
  • Some ergonomics issues, interfaces may change in upcoming releases

Supported databases

  • PostgreSQL (feature pg) via tokio-postgres
  • Sqlite (feature sqlite) via rusqlite

Getting started

First time

  1. You'll need the following runtime dependencies:

    • good-ormning-runtime
    • tokio-postgres for PostgreSQL
    • rusqlite for Sqlite

    And build.rs dependencies:

    • good-ormning

    And you must enable one (or more) of the database features:

    • pg
    • sqlite

    plus maybe chrono for DateTime support.

  2. Create a build.rs and define your initial schema version and queries

  3. Call goodormning::generate() to output the generated code

  4. In your code, after creating a database connection, call migrate

Schema changes

  1. Copy your previous version schema, leaving the old schema version untouched. Modify the new schema and queries as you wish.
  2. Pass both the old and new schema versions to goodormning::generate(), which will generate the new migration statements.
  3. At runtime, the migrate call will make sure the database is updated to the new schema version.


This build.rs file

use std::{
use good_ormning::sqlite::{

fn main() {
    let root = PathBuf::from(&env::var("CARGO_MANIFEST_DIR").unwrap());
    let mut latest_version = Version::default();
    let users = latest_version.table("zQLEK3CT0", "users");
    let id = users.rowid_field(&mut latest_version, None);
    let name = users.field(&mut latest_version, "zLQI9HQUQ", "name", field_str().build());
    let points = users.field(&mut latest_version, "zLAPH3H29", "points", field_i64().build());
    good_ormning::sqlite::generate(&root.join("tests/sqlite_gen_hello_world.rs"), vec![
        // Versions
        (0usize, latest_version)
    ], vec![
        // Latest version queries
        new_insert(&users, vec![(name.clone(), Expr::Param {
            name: "name".into(),
            type_: name.type_.type_.clone(),
        }), (points.clone(), Expr::Param {
            name: "points".into(),
            type_: points.type_.type_.clone(),
        })]).build_query("create_user", QueryResCount::None),

        new_select(&users).where_(Expr::BinOp {
            left: Box::new(Expr::Field(id.clone())),
            op: BinOp::Equals,
            right: Box::new(Expr::Param {
                name: "id".into(),
                type_: id.type_.type_.clone(),
        }).return_fields(&[&name, &points]).build_query("get_user", QueryResCount::One),

        new_select(&users).return_field(&id).build_query("list_users", QueryResCount::Many)

Generates something like:

pub fn migrate(db: &mut rusqlite::Connection) -> Result<(), GoodError> {
    // ...

pub fn create_user(db: &mut rusqlite::Connection, name: &str, points: i64) -> Result<(), GoodError> {
    // ...

pub struct DbRes1 {
    pub name: String,
    pub points: i64,

pub fn get_user(db: &mut rusqlite::Connection, id: i64) -> Result<DbRes1, GoodError> {
    // ...

pub fn list_users(db: &mut rusqlite::Connection) -> Result<Vec<i64>, GoodError> {
    // ...

And can be used like:

fn main() {
    use sqlite_gen_hello_world as queries;

    let mut db = rusqlite::Connection::open_in_memory().unwrap();
    queries::create_user(&db, "rust human", 0).unwrap();
    for user_id in queries::list_users(&db).unwrap() {
        let user = queries::get_user(&db, user_id).unwrap();
        println!("User {}: {}", user_id, user.name);
User 1: rust human

Usage details


  • pg - enables generating code for PostgreSQL
  • sqlite - enables generating code for Sqlite
  • chrono - enable datetime field/expression types

Schema IDs and IDs

"Schema IDs" are internal ids used for matching fields across versions, to identify renames, deletes, etc. Schema IDs must not change once used in a version. I recommend using randomly generated IDs, via a key macro. Changing Schema IDs will result in a delete followed by a create.

"IDs" are used both in SQL (for fields) and Rust (in parameters and returned data structures), so must be valid in both (however, some munging is automatically applied to ids in Rust if they clash with keywords). Depending on the database, you can change IDs arbitrarily between schema versions but swapping IDs in consecutive versions isn't currently supported - if you need to do swaps do it over three different versions (ex: v0: A and B, v1: A_ and B, v2: B and A).

Query, expression and fields types

Use type_* field_* functions to get type builders for use in expressions/fields.

Use new_insert/select/update/delete to create query builders.

There are also some helper functions for building queries, see

  • field_param, a shortcut for a parameter matching the type and name of a field
  • set_field, a shortcut for setting field values in INSERT and UPDATE
  • eq_field, gt_field, gte_field, lt_field, lte_field are shortcuts for expressions comparing a field and a parameter with the same type
  • expr_and, a shortcut for AND expressions

for the database you're using.

Custom types

When defining a field in the schema, call .custom("mycrate::MyString", type_str().build()) on the field type builder (or pass it in as Some("mycreate::MyType".to_string()) if creating the type structure directly).

The type must have methods to convert to/from the native SQL types. There are traits to guide the implementation:

pub struct MyString(pub String);

impl good_ormning_runtime::pg::GoodOrmningCustomString<MyString> for MyString {
    fn to_sql(value: &MyString) -> &str {

    fn from_sql(s: String) -> Result<MyString, String> {


The Expr::Call variant allows you to create method call expressions. You must provide in compute_type a helper method to type-check the arguments and determine the type of the evaluation of the call.

The first parameter is the evaluation context, which contains errs for reporting errors. The second is a path from the evaluation tree root up to the call, for identifying where in a query expression errors occur. The third argument is a vec of arguments passed to the call. Each argument can be a single type or a record consisting of multiple types (like in () in where (x, y, z) < (b.x, b.y, b.z)). If there are no errors, this must return Some(...).

Error handling is lazy during expression checking - even if an error occurs, processing can continue (and identify more errors before aborting). All errors are fatal, they just don't cause an abort immediately.

If there are errors, record the errors in ctx.errs.err(path.add(format!("Argument 0")), format!("Error")). If evaluation within the call cannot continue, return None, otherwise continue.

Parameters and return types

Parameters with the same name are deduplicated - if you define a query with multiple parameters of the same name but different types you'll get an error.

Different queries with the same multiple-field returns will use the same return type.


Vs Diesel

Good-ormning is functionally most similar to Diesel.


  • You can define your queries and result structures near where you use them
  • You can dynamically define queries (i.e. swap operators depending on the input, etc.)
  • Result structures must be manually defined, and care must be taken to get the field order to match the query
  • You can define new types to use in the schema, which are checked against queries, although this requires significant boilerplate
  • Requires many macros, trait implementations
  • To synchronize your migrations and in-code schema, you can use the CLI with a live database with migrations applied. However, this resets any custom SQL types in the schema with the built-in SQL types. Alternatively you can maintain the schema by hand (and risk query issues due to typos, mismatches).
  • Column count limitations, slow build times
  • Supports more syntax, withstood test of time


  • Queries have to be defined separately, in the build.rs file
  • All queries have to be defined up front in build.rs
  • You don't have to write any structures, everything is generated from schema and query info
  • Custom types can be incorporated into the schema with no boilerplate
  • Migrations are automatically derived via a diff between schema versions plus additional migration metadata
  • Clear error messages, thanks to no macros, generics
  • Code generation is fast, compiling the simple generated code is also fast
  • Alpha



  • SQLx has no concept of a schema so it can only perform type-checking on native SQL types (no consideration for new types, blob encodings, etc)
  • Requires a running database during development


  • The same schema used for generating migrations is used for type checking, and natively supports custom types
  • A live database is unused during development, but all query syntax must be manually implemented in Good-ormning so you may encounter missing features


SeaORM focuses on runtime checks rather than compile time checks.

A few words on the future

Obviously writing an SQL VM isn't great. The ideal solution would be for popular databases to expose their type checking routines as libraries so they could be imported into external programs, like how Go publishes reusable ast-parsing and type-checking libraries.


~153K SLoC