asgire 3.12.2

The revamped and modernized drop-in replacement for asgiref

asgire

The revamped and modernized drop-in replacement for asgiref.

Same license, same API, but with better code, comprehensive coverage, and active maintenance.

Installation

pip install asgire

The import stays import asgiref — no code changes needed.

Migration

pip uninstall asgiref
pip install asgire

If you need to force Django or other libraries to depend on asgire instead of asgiref with uv, add the following to your pyproject.toml:

[tool.uv]
override-dependencies = ["asgiref ; python_version == '0'"]

This will eliminate all transitive dependencies on asgiref in uv.lock to ensure asgire is the only import asgiref provider.

Performance tips

The cost of sync_to_async / async_to_sync is paid per boundary crossing (a thread hop through the event loop, tens of microseconds), not per line — so the order below is roughly the order of impact.

Minimize boundary crossings

Cross as few times as possible. The biggest wins come from restructuring, not from micro-optimizing each call.

Batch sync work into one crossing instead of awaiting in a loop:

# N crossings
for row in rows:
    await sync_to_async(row.save)()

# One crossing
@sync_to_async
def save_all(rows):
    for row in rows:
        row.save()

await save_all(rows)

Resolve sync data while you are still in sync context. A common Django antipattern bridges into async and then immediately bounces back to sync for ORM access. Signals run in sync context where the DB connection is already available, so read what you need there and pass plain objects to the async layer:

# Bounces sync -> async -> sync -> async (three+ crossings)
@receiver(post_save, sender=IPAddress)
def on_saved(sender, instance, **kwargs):
    async_to_sync(notify)(instance)

async def notify(instance):
    user = await sync_to_async(lambda: instance.user)()
    organization = await sync_to_async(lambda: instance.organization)()
    await send_email(user, organization)

# Read the ORM data in the sync signal; hand plain objects to async (one crossing)
@receiver(post_save, sender=IPAddress)
def on_saved(sender, instance, **kwargs):
    user = instance.user
    organization = instance.organization
    async_to_sync(notify)(user, organization)

async def notify(user, organization):
    await send_email(user, organization)

Defer I/O triggered by signals. Sending email or calling an external API inside a signal runs it inside the transaction (it can fire even if the transaction later rolls back) and on the request's hot path. Prefer enqueuing the work to run after commit:

@receiver(post_save, sender=IPAddress)
def on_saved(sender, instance, **kwargs):
    user, organization = instance.user, instance.organization
    transaction.on_commit(lambda: notify_task.delay(user.id, organization.id))

Reuse wrappers instead of building them inline

sync_to_async(func) / async_to_sync(func) validate and wrap the callable on construction (~2 µs). Calling them inline rebuilds the wrapper on every call:

# Slower — a fresh wrapper is constructed on every request
async def view():
    return await sync_to_async(do_work)()

# Faster — build the wrapper once and reuse it
do_work_async = sync_to_async(do_work)

async def view():
    return await do_work_async()

Choose thread_sensitive deliberately

sync_to_async defaults to thread_sensitive=True, which runs every call on a single shared thread. That is required for thread-affine code (e.g. the Django ORM or anything using thread-locals), but it serializes calls. For sync work that is independent and thread-safe, pass thread_sensitive=False so calls run concurrently in a thread pool:

# Serialized on one shared thread (safe for thread-affine code)
await sync_to_async(orm_call)()

# Runs concurrently in a pool (use only for independent, thread-safe work)
await sync_to_async(cpu_bound, thread_sensitive=False)()

Development

uv sync
uv run pytest -v
uv run ruff check --fix
uv run ruff format
uv run ty check