Yara ORM

A fast, async Python ORM with a Rust engine — Tortoise-style models, querysets, relations and migrations for PostgreSQL, MySQL, MariaDB, SQLite, Oracle and Microsoft SQL Server.

📖 Documentation: vsdudakov.github.io/yara-orm

✍️ Deep dive: How the GIL, PyO3 & asyncio cooperate — how the Rust engine bridges Python's event loop without the GIL collapsing it.

Yara ORM is a high-performance async ORM for Python that pairs the ergonomics of a Django/Tortoise-style API — models, querysets, relations, aggregation and migrations — with a hot path (connection pooling, parameter binding, row decoding) written in compiled Rust (PyO3 + tokio). It is a drop-in-feel alternative to Tortoise ORM and async SQLAlchemy: 2–9× faster than popular pure-Python ORMs on common operations, with first-class PostgreSQL, MySQL, MariaDB and SQLite backends — plus Oracle and Microsoft SQL Server (both beta) — full type hints, and 100% test coverage.

from yara_orm import Model, YaraOrm, fields

class User(Model):
    id = fields.IntField(pk=True)
    name = fields.CharField(max_length=120)

await YaraOrm.init("postgres://localhost/app")
await YaraOrm.generate_schemas()
await User.create(name="Ada")
print(await User.filter(name__icontains="ad").count())

Highlights

• ⚡ Rust engine — pooling, binding and decoding in compiled code; the async bridge (PyO3 + tokio) keeps your event loop free.
• 🧩 Familiar API — Tortoise/Django-style models, lazy chainable querysets, Q objects, aggregation, prefetch_related, transactions, signals. Coming from Tortoise? Most code moves across unchanged — see Migrating from Tortoise ORM.
• 🗄️ Pluggable backends — PostgreSQL, MySQL/MariaDB and SQLite, plus beta Oracle and Microsoft SQL Server backends, selected by URL; a new database is one Rust trait + one Python dialect.
• 🚚 Migrations — operation-based, auto-generated, backend-portable (makemigrations / upgrade / downgrade).
• 🧪 Quality — fully typed, linted (ruff + ty) and 100% test coverage.

Installation

pip install yara-orm

Prebuilt wheels are published for Linux, macOS and Windows on CPython 3.9–3.14, so installation needs no Rust toolchain. (Installing the source distribution on an unsupported platform compiles the engine and requires a Rust toolchain — see Development.)

Quick start

import asyncio
from yara_orm import Model, YaraOrm, fields


class Tournament(Model):
    id = fields.IntField(pk=True)
    name = fields.CharField(max_length=100)
    created_at = fields.DatetimeField(auto_now_add=True)


class Event(Model):
    id = fields.IntField(pk=True)
    name = fields.CharField(max_length=100, index=True)
    tournament = fields.ForeignKeyField("Tournament", related_name="events")


async def main() -> None:
    await YaraOrm.init("postgres://localhost/app")   # or "mysql://…", "sqlite:///app.db"
    await YaraOrm.generate_schemas()

    cup = await Tournament.create(name="World Cup")
    await Event.create(name="Final", tournament=cup)

    # Lazy, chainable queries
    finals = await Event.filter(name__icontains="fin").order_by("-id")
    count = await Event.filter(tournament=cup).count()

    # Relations
    async for event in cup.events:
        print(event.name, "→", (await event.tournament).name)

    await YaraOrm.close()


asyncio.run(main())

Querying

# Lookups: exact, not, gt/gte/lt/lte, in, isnull, contains/icontains,
# startswith/endswith (+ case-insensitive `i` variants)
await User.filter(age__gte=18, name__icontains="a").order_by("-age").limit(10)

# Complex boolean filters with Q
from yara_orm import Q
await User.filter(Q(name="Ada") | Q(age__lt=30)).exclude(active=False)

# Aggregation + group by
from yara_orm import Count, Sum
await Author.annotate(books=Count("books")).filter(books__gte=1)
await Book.annotate(total=Sum("rating")).group_by("author_id").values("author_id", "total")

# Construction-free projections
await User.all().values("id", "name")
await User.all().values_list("name", flat=True)

Model methods: create, bulk_create, get, get_or_none, filter, exclude, all, annotate, prefetch_related, raw; instances save, delete, fetch_related.

Relations

class Author(Model):
    name = fields.CharField(max_length=100)

class Book(Model):
    title = fields.CharField(max_length=200)
    author = fields.ForeignKeyField("Author", related_name="books")
    tags = fields.ManyToManyField("Tag", related_name="books")

book = await Book.create(title="Compilers", author=author)
await book.tags.add(tag1, tag2)        # m2m add / remove / clear

await book.author                       # forward FK (awaitable)
async for b in author.books: ...        # reverse manager
await Author.all().prefetch_related("books")   # no N+1

ForeignKeyField, OneToOneField, ManyToManyField, recursive self-FK, related_name, Prefetch(rel, queryset=...).

Transactions, signals & more

from yara_orm import in_transaction, atomic, pre_save, connections

async with in_transaction():            # commit on success, rollback on error
    await Account.create(name="A")
    async with in_transaction():        # nesting opens a savepoint
        await Account.create(name="B")  # rolls back independently on error

@atomic(isolation="SERIALIZABLE")       # isolation levels (PostgreSQL/MySQL)
async def transfer(): ...

@pre_save(User)                          # lifecycle signals
async def on_save(sender, instance, using_db, update_fields): ...

await connections.get("default").execute("INSERT ...", [..])   # manual SQL

Also: enum fields (IntEnumField/CharEnumField), column/table comments (description=, Meta.table_description), and multi-database routing via a Router over multiple named connections.

Backends

Backends are selected by the connection URL; the abstraction is a single Rust trait (Backend) plus a Python BaseDialect subclass:

await YaraOrm.init("postgres://user@localhost/db")     # PostgreSQL (tokio-postgres)
await YaraOrm.init("mysql://user:pass@localhost/db")   # MySQL/MariaDB (mysql_async)
await YaraOrm.init("sqlite:///path/to/app.db")          # SQLite (rusqlite)
await YaraOrm.init("oracle://user:pass@localhost:1521/FREEPDB1")  # Oracle 23ai — beta (oracle-rs)
await YaraOrm.init("mssql://user:pass@localhost:1433/db")  # SQL Server 2017+ — beta (tiberius)

The SQLite backend maps rich types (uuid/json/datetime/decimal) onto SQLite's storage classes and reconstructs them on read from the declared column type, so the model layer is identical across backends.

Oracle and Microsoft SQL Server (both beta) ride on the same model/queryset API — both on pure-Rust drivers (oracle-rs TNS and tiberius TDS, no OCI/ODBC/Instant Client, so the wheels stay self-contained). The shared cross-backend suite runs against a live SQL Server 2022 and Oracle 23ai in CI. See the backends guide for their type maps and the driver caveats that keep them out of the stable tier.

Migrations

A Django/Tortoise-style, operation-based migration system. Migrations are auto-generated from model changes and backend-portable — the same operations render to PostgreSQL, MySQL, SQLite, Oracle or SQL Server DDL at apply time. Applied migrations are tracked in an orm_migrations table.

# autodetect model changes -> migrations/0001_initial.py
python -m yara_orm --models myapp.models makemigrations --name initial

# preview SQL without running it (per the target dialect)
python -m yara_orm --db sqlite:///app.db --models myapp.models sqlmigrate 0001_initial

# apply / revert / inspect
python -m yara_orm --db postgres://localhost/app --models myapp.models upgrade
python -m yara_orm --db postgres://localhost/app --models myapp.models downgrade
python -m yara_orm --db postgres://localhost/app --models myapp.models history

Each migration is a class Migration(m.Migration) whose operations are built from live field objects: CreateModel, DeleteModel, AddField, RemoveField, AlterField, AddIndex, RemoveIndex, plus hand-written renames (RenameModel / RenameField / RenameIndex), constraints (AddConstraint / RemoveConstraint / RenameConstraint with UniqueConstraint / CheckConstraint) and RunSQL / RunPython for data migrations. makemigrations emits the idempotent analogs (CreateModelIfNotExists, AddFieldIfNotExists, …) and detects AlterField automatically. The same commands are available programmatically via yara_orm.MigrationManager.

Performance

Median of 5 runs, Python 3.12, 5000 rows, against eight other Python ORMs (Tortoise, SQLAlchemy, Pony, Django, Peewee, SQLObject, Ormar, Piccolo) — Yara ORM is fastest or tied on every operation across PostgreSQL, MySQL, MariaDB and SQLite, losing only to leaner in-process sync ORMs on single-row point reads. Times in ms, lower is better. Full methodology, speedup tables and per-op notes in benchmarks/ and the performance docs.

PostgreSQL 18

MySQL 8.4

Same workload against MySQL (Tortoise over asyncmy, SQLAlchemy/Ormar over aiomysql, the sync ORMs over pymysql; Piccolo has no MySQL backend):

(single_insert ~0.6–1.1 s is dominated by InnoDB's per-commit fsync — every ORM pays it; get_by_pk and single_insert include the Docker-network round trip.)

MariaDB 11

Every competitor connects through its MySQL driver; yara-orm auto-detects MariaDB and uses its RETURNING dialect. Piccolo has no MySQL backend:

(single_insert, update and delete are near ties across every ORM here because they're database-bound — single inserts are paced by MariaDB's per-commit disk fsync, and update/delete are one server-side set statement each — so there's no client-side marshaling for the Rust hot path to speed up. MariaDB's single_insert ~310 ms is markedly faster than MySQL 8's — a lighter default commit path.)

SQLite

SQLite is in-process, so these use its recommended sync_fast_path=1 config (statements run synchronously on the calling thread — no I/O to overlap):

With the fast path Yara ORM is fastest or tied on every operation (SQLObject's hand-written raw SQL ties the sub-millisecond group_by) — including the point reads it trailed on under the default async bridge (get_by_pk 1.1× vs SQLObject, 2.6× vs Pony), while staying far ahead on throughput (bulk_insert 1.9–82×, fetch_all 2.6–17×, filter 2.5–21×). On the default async path the per-statement bridge costs tens of µs on sequential point reads, so SQLObject's lean sync active-record leads get_by_pk there; sqlite://...?sync_fast_path=1 removes that bridge (~7× faster point queries).

Speed comes from the Rust hot path, positional row decoding (no per-row dict or column-name allocation), compiled-SQL + prepared-statement caching, and connection pooling (see Performance). Run it yourself with make bench / make bench-mysql / make bench-sqlite.

Architecture

┌─────────────────────────────────────────────┐
│ Python  (python/yara_orm) ................. │
│   Model / metaclass ....... schema + ORM API│
│   QuerySet ................ lazy SQL builder│
│   fields .................... abstract types│
│   dialects ................ per-DB SQL rules│
└───────────────┬─────────────────────────────┘
                │  sql + params  (PyO3 / asyncio bridge)
┌───────────────▼─────────────────────────────┐
│ Rust  (rust/src)  →  yara_orm._engine ..... │
│   Engine ...................... async facade│
│   Backend trait .............. pluggable DBs│
│     PgBackend ............... tokio-postgres│
│     MySqlBackend ................ mysql_async│
│     SqliteBackend ................. rusqlite│
│     OracleBackend ............... oracle-rs│
│     MsSqlBackend .................. tiberius│
│   Value .................. Py⇆Rust⇆SQL types│
└─────────────────────────────────────────────┘

• Rust owns pooling (deadpool), binding, type conversion and decoding.
• Python owns the model layer and SQL generation.
• Adding a database = a new Backend impl + scheme match in rust/src/backend/mod.rs, plus a BaseDialect subclass in python/yara_orm/dialects.py. The model layer never changes.

Development

git clone https://github.com/vsdudakov/yara-orm
cd yara-orm
make dev        # create .venv313 and install dev tools (maturin, ruff, ty, pytest)
make build      # compile the Rust engine into the venv (maturin develop)
make lint       # ruff check + ruff format --check + ty
make test       # pytest against $DB (default postgres://localhost/orm_demo)
make cov        # tests with the 100% coverage gate
make bench      # cross-ORM benchmark (needs `make bench-setup` once; Python ≤ 3.12 for Pony)
make bench-mysql   # same comparison on MySQL
make bench-sqlite  # same comparison on SQLite

Requires a Rust toolchain (rustup), a local PostgreSQL for the Postgres tests and a local MySQL for the MySQL tests; the SQLite tests are self-contained.

Contributing

Issues and pull requests are welcome. Please run make lint and make cov (both must be green — lint clean and 100% coverage) before opening a PR.

Sponsor

Yara ORM is MIT-licensed and developed in the open. If it saves your project time — or you'd like to support continued work on the Rust engine, backends and docs — please consider sponsoring on GitHub. Every bit helps and is hugely appreciated. ❤️

License

MIT © Yara ORM contributors