UVICORN reload inside docker container not working

I have issues with hot reloading of a FAST API app inside a Docker container.

I can see that the volume mapping is doing its thing and the files are updated inside the docker container.

The console log tells me uvicorn is watching files inside the correct folder.

But any change I make is not triggering a reload of the app.

I have the following setup:

My docker-compose.yaml:

#docker-compose.yml
version: "3"

services:
  app:
    container_name: api
    image: api:latest
    depends_on:
      - mysqldb
    build: 
      context: .
      dockerfile: Dockerfile.dev
      args:
        DEV: "true"
    ports:
      - "8080:8080"
    env_file:
      - .env
    environment:
      - WATCHFILES_FORCE_POLLING=true
      - DEV=1
      - MYSQL_HOSTNAME=mysqldb
      - MYSQL_PORT=3306
    command: uvicorn app.main:app --host 0.0.0.0 --port 8080 --reload --reload-dir /src/app
    volumes:
      - ./app:/src/app
  mysqldb:
    container_name: mysqldb
    image: mysql:latest
    restart: always
    ports:
     - 3307:3306
    environment:
      MYSQL_ROOT_PASSWORD: randompassword
      MYSQL_DATABASE: my-database

And my Dockerfile

#dockerfile.dev
FROM python:3.10-slim-buster

ENV PYTHONUNBUFFERED 1

COPY requirements.txt /
COPY alembic.ini /
COPY migrations /migrations

RUN pip install --upgrade pip setuptools wheel gunicorn uvloop httptools
RUN pip install -r requirements.txt
RUN pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
RUN pip install sentence-transformers 

# Copy the source code to src folder
COPY . /src

ENV PYTHONPATH=/src

EXPOSE 8080

# FOR local dev the startup happens inside the docker-compose with reload options
COPY startup.sh /src/startup.sh
RUN chmod +x /src/startup.sh

CMD ["uvicorn", "app.main:app","--host", "0.0.0.0", "--port", "8080", "--reload"]

I've tried almost all possible configurations of the reload.

Without running the uvicorn command inside the docker file
With the reload dir as root (/), app, and src/app
Without the reload dir
With the guvicorn server instead of the uvicorn server

Solution

After tinkering arround with different set ups we now have a stable reload workspace. The issue with the uvicorn --reload statement had to do with volume mapping and the point when the command got called.

Now instead of running the uvicorn command from the docker compose file I run a command to keep the container alive:

command: /bin/sh -c "while true; do sleep 30; done"

Then I connect manually with the container:

docker exec -it ml_api /bin/bash

And then i can start (or manually restart) the uvicorn (or guvicorn) server. The reload statement now works correctly and if any big changes happen I can restart uvicorn without restarting my docker container.