1. Problem: Does on_commit() Execute Too Early?

Recently, in a Django project, I configured transaction.on_commit() to execute a Celery task after the transaction is successful.

However, upon checking the logs, I noticed an unusual phenomenon.

Even though operations like post.categories.add(...) had not yet completed,
the Celery task inside on_commit() was being executed.

2. Looking at the Code...

def view():
    post = Post.objects.create(...)

    def post_process():
        with transaction.atomic():
            post.categories.add(...)
            post.tags.add(...)

            def run_async():
                translate_post.delay(post.id)

            transaction.on_commit(run_async)

    threading.Thread(target=post_process).start()

At first, it didn't seem like a problem...
But this threading.Thread() was the trap.

3. Cause: Transactions are Thread Local

Django's transactions are dependent on the current thread.
This means that transaction.atomic() and on_commit() only apply to the transaction opened in this thread right now.

A transaction.atomic() opened in a new thread
has no relation to the main transaction.

4. What Happened As a Result

• The Celery task was executed "after the transaction commit," but
• that transaction was merely a small transaction within a sub-thread.
• post.categories.add(...) and others were still in progress

In other words, Celery was executed too early than intended, leading to issues with referencing incomplete data.

5. Solution

It’s simple to fix:

def view():
    post = Post.objects.create(...)

    with transaction.atomic():
        post.categories.add(...)
        post.tags.add(...)

        def run_async():
            translate_post.delay(post.id)

        transaction.on_commit(run_async)

🚫 threading.Thread(...)
✅ Just register on_commit() in the main thread

6. Conclusion

• Django's transactions are thread local
• on_commit() executes only after the transaction commits in the current thread
• Registering in another thread will only respond to that thread's transaction

This experience was a powerful lesson.
When using transactions + Celery + threads together,
I realized I need to be very cautious.