feat(server): Machine learning's image optimisations (#1908)

* Use multi stage build to slim down ML image size

* Use gunicorn as WSGI server in ML image

* Configure gunicorn server for ML use case

* Use requirements.txt file to install python dependencies in ML image

* Make ML listen IP configurable

machine-learning/Dockerfile

@@ -1,19 +1,26 @@
-FROM python:3.10
+FROM python:3.10 as builder
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=true
+
+COPY requirements.txt ./
+
+RUN python -m venv /opt/venv && \
+    /opt/venv/bin/pip install --upgrade pip setuptools wheel && \
+    /opt/venv/bin/pip install --no-deps -r requirements.txt
+
+FROM python:3.10-slim
+
+COPY --from=builder /opt/venv /opt/venv
 
 ENV TRANSFORMERS_CACHE=/cache \
     PYTHONDONTWRITEBYTECODE=1 \
     PYTHONUNBUFFERED=1 \
-    PIP_NO_CACHE_DIR=true
+    PATH="/opt/venv/bin:$PATH"
 
 WORKDIR /usr/src/app
 
-RUN python -m venv /opt/venv
-ENV PATH="/opt/venv/bin:$PATH"
-
-RUN pip install --pre torch  -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
-RUN pip install transformers tqdm numpy scikit-learn scipy nltk sentencepiece flask Pillow
-RUN pip install --no-deps sentence-transformers
-
 COPY . .
 
-CMD ["python", "src/main.py"]
+CMD ["gunicorn", "src.main:server"]

machine-learning/gunicorn.conf.py

@@ -0,0 +1,29 @@
+"""
+Gunicorn configuration options.
+https://docs.gunicorn.org/en/stable/settings.html
+"""
+import os
+
+
+# Set the bind address based on the env
+port = os.getenv("MACHINE_LEARNING_PORT") or "3003"
+listen_ip = os.getenv("MACHINE_LEARNING_IP") or "0.0.0.0"
+bind = [f"{listen_ip}:{port}"]
+
+# Preload the Flask app / models etc. before starting the server
+preload_app = True
+
+# Logging settings - log to stdout and set log level
+accesslog = "-"
+loglevel = os.getenv("MACHINE_LEARNING_LOG_LEVEL") or "info"
+
+# Worker settings
+# ----------------------
+# It is important these are chosen carefully as per
+# https://pythonspeed.com/articles/gunicorn-in-docker/
+# Otherwise we get workers failing to respond to heartbeat checks,
+# especially as requests take a long time to complete.
+workers = 2
+threads = 4
+worker_tmp_dir = "/dev/shm"
+timeout = 60

machine-learning/requirements.txt

@@ -0,0 +1,33 @@
+certifi==2022.12.7
+charset-normalizer==3.0.1
+click==8.1.3
+filelock==3.9.0
+Flask==2.2.3
+gunicorn==20.1.0
+huggingface-hub==0.12.1
+idna==3.4
+importlib-metadata==6.0.0
+itsdangerous==2.1.2
+Jinja2==3.1.2
+joblib==1.2.0
+MarkupSafe==2.1.2
+nltk==3.8.1
+numpy==1.24.2
+packaging==23.0
+Pillow==9.4.0
+PyYAML==6.0
+regex==2022.10.31
+requests==2.28.2
+scikit-learn==1.2.1
+scipy==1.10.1
+sentence-transformers==2.2.2
+sentencepiece==0.1.97
+threadpoolctl==3.1.0
+tokenizers==0.13.2
+torch==1.13.1 -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
+tqdm==4.64.1
+transformers==4.26.1
+typing-extensions==4.5.0
+urllib3==1.26.14
+Werkzeug==2.2.3
+zipp==3.15.0