feat(jobs): improve MLOps example

jobs/ml-ops/README.md

@@ -37,14 +37,22 @@ Set your Scaleway access key, secret key and project ID in environment variables
 export TF_VAR_access_key=<your-access-key>
 export TF_VAR_secret_key=<your-secret-key>
 export TF_VAR_project_id=<your-project-id> # you can create a separate project for this example
+```
+
+By default, both jobs and container trigger in the example run regularly on a schedule. The default values for these schedules are configured in `jobs/ml-ops/terraform/variables.tf`, and can be overridden using Terraform variables, e.g. `export TF_VAR_data_fetch_cron_schedule="0 10 * * *"`.
 
+Then deploy MLOps infrastructure using the following:
+
+```console
 cd terraform
 terraform init
 terraform plan
 terraform apply
 ```
 
-### Step 2. Run the data and training Jobs
+### Step 2. Optional: trigger jobs manually
+
+The pipeline is automatic, all jobs will be run at their respective scheduled time. This step can be ignored unless for debugging or test purposes.
 
 To run the jobs for the data and training, we can use the Scaleway CLI:
 
@@ -60,12 +68,17 @@ You can also trigger the jobs from the [Jobs section](https://console.scaleway.c
 
 ### Step 3. Use the inference API
 
+Load model with the latest version using:
+
 ```
 cd terraform
 export INFERENCE_URL=$(terraform output raw endpoint)
+curl -X POST ${INFERENCE_URL}
+```
 
-curl -X POST ${INFERENCE_URL}/load
+Then post data to infer the class:
 
+```
 curl -X POST \
   -H "Content-Type: application/json" \
   -d @../inference/example.json

jobs/ml-ops/data/main.py

@@ -1,8 +1,9 @@
-import boto3
 import os
 import urllib.request
 import zipfile
 
+import boto3
+
 DATA_DIR = "dataset"
 
 ZIP_URL = "http://archive.ics.uci.edu/static/public/222/bank+marketing.zip"

jobs/ml-ops/inference/data.py

@@ -1,5 +1,5 @@
-import pandas as pd
 import numpy as np
+import pandas as pd
 from pydantic import BaseModel
 
 

jobs/ml-ops/inference/loader.py

@@ -0,0 +1,60 @@
+import os
+import pickle
+
+import boto3
+
+
+class ClassifierLoader:
+    _classifier = None
+    _classifier_version = ""
+
+    @classmethod
+    def load(cls, force=False):
+        if force or cls._classifier is None:
+            access_key = os.environ["ACCESS_KEY"]
+            secret_key = os.environ["SECRET_KEY"]
+            region_name = os.environ["REGION"]
+
+            bucket_name = os.environ["S3_BUCKET_NAME"]
+            s3_url = os.environ["S3_URL"]
+
+            s3 = boto3.client(
+                "s3",
+                region_name=region_name,
+                endpoint_url=s3_url,
+                aws_access_key_id=access_key,
+                aws_secret_access_key=secret_key,
+            )
+
+            # get model file with the latest version
+            bucket_objects = s3.list_objects(Bucket=bucket_name)
+            get_last_modified = lambda object: int(
+                object["LastModified"].strftime("%s")
+            )
+            model_objects = [
+                model_object
+                for model_object in bucket_objects["Contents"]
+                if "classifier" in model_object["Key"]
+            ]
+            latest_model_file = [
+                object["Key"] for object in sorted(model_objects, key=get_last_modified)
+            ][0]
+
+            s3.download_file(bucket_name, latest_model_file, latest_model_file)
+
+            with open(latest_model_file, "rb") as fh:
+                cls._classifier = pickle.load(fh)
+                cls._classifier_version = latest_model_file[11:-4]
+
+            print(
+                "Successfully loaded model file: {latest_model_file}".format(
+                    latest_model_file=latest_model_file
+                ),
+                flush=True,
+            )
+
+        return cls._classifier
+
+    @classmethod
+    def model_version(cls):
+        return cls._classifier_version

jobs/ml-ops/inference/main.py

@@ -1,54 +1,38 @@
+import data
 from fastapi import FastAPI
+from loader import ClassifierLoader
 from sklearn.ensemble import RandomForestClassifier
-from sklearn.metrics import RocCurveDisplay
-import pickle
-import boto3
-import pandas
-import os
-
-import data
 
 classifier = RandomForestClassifier()
 
 app = FastAPI()
 
-MODEL_FILE = "classifier.pkl"
-
-
-class ClassifierLoader:
-    _classifier = None
 
-    @classmethod
-    def load(cls, force=False):
-        if force or cls._classifier is None:
-            access_key = os.environ["ACCESS_KEY"]
-            secret_key = os.environ["SECRET_KEY"]
-            region_name = os.environ["REGION"]
+@app.get("/")
+def hello():
+    """Get Model Version"""
 
-            bucket_name = os.environ["S3_BUCKET_NAME"]
-            s3_url = os.environ["S3_URL"]
+    model_version = ClassifierLoader.model_version()
 
-            s3 = boto3.client(
-                "s3",
-                region_name=region_name,
-                endpoint_url=s3_url,
-                aws_access_key_id=access_key,
-                aws_secret_access_key=secret_key,
-            )
+    if model_version == "":
+        return {
+            "message": "Hello, this is the inference server! No classifier loaded in memory."
+        }
 
-            s3.download_file(bucket_name, MODEL_FILE, MODEL_FILE)
+    return {
+        "message": "Hello, this is the inference server! Serving classifier with version {model_version}".format(
+            model_version=model_version
+        )
+    }
 
-            with open(MODEL_FILE, "rb") as fh:
-                cls._classifier = pickle.load(fh)
 
-        return cls._classifier
-
-
-@app.post("/load")
+# this endpoint is used by cron trigger to load model from S3
+@app.post("/")
 def load():
     """Reloads classifier from model registry bucket"""
 
     ClassifierLoader.load(force=True)
+
     return {"message": "model loaded successfully"}
 
 
@@ -59,8 +43,13 @@ def classify(profile: data.ClientProfile):
     cleaned_data = data.clean_profile(profile)
     data_point_processed = data.transform_data(cleaned_data)
 
-    # Lazy-loads classifer from S3
+    # Lazy-loads classifier from S3
     classifier = ClassifierLoader.load()
     prediction = classifier.predict(data_point_processed)
 
-    return {"predicted_class": int(prediction)}
+    response = "This client is likely to respond positively to a cold call"
+
+    if int(prediction) == 0:
+        response = "This client is likely to respond negatively to a cold call"
+
+    return {"prediction": response}

jobs/ml-ops/terraform/container.tf

@@ -12,7 +12,7 @@ resource "scaleway_container" "inference" {
   cpu_limit      = 2000
   memory_limit   = 2048
   min_scale      = 1
-  max_scale      = 5
+  max_scale      = 1
   environment_variables = {
     "S3_BUCKET_NAME" = scaleway_object_bucket.main.name
     "S3_URL"         = var.s3_url
@@ -24,3 +24,9 @@ resource "scaleway_container" "inference" {
   }
   deploy   = true
 }
+
+resource scaleway_container_cron "inference_cron" {
+    container_id = scaleway_container.inference.id
+    schedule = var.inference_cron_schedule
+    args = jsonencode({})
+}

jobs/ml-ops/terraform/jobs.tf

@@ -4,7 +4,10 @@ resource "scaleway_job_definition" "fetch_data" {
   memory_limit = 1024
   image_uri    = docker_image.data.name
   timeout      = "10m"
-
+  cron {
+    schedule = var.data_fetch_cron_schedule
+    timezone = "Europe/Paris"
+  }
   env = {
     "S3_BUCKET_NAME" : scaleway_object_bucket.main.name,
     "S3_URL" : var.s3_url,
@@ -20,7 +23,10 @@ resource "scaleway_job_definition" "training" {
   memory_limit = 4096
   image_uri    = docker_image.training.name
   timeout      = "10m"
-
+  cron {
+    schedule = var.training_cron_schedule
+    timezone = "Europe/Paris"
+  }
   env = {
     "S3_BUCKET_NAME" : scaleway_object_bucket.main.name,
     "S3_URL" : var.s3_url,

jobs/ml-ops/terraform/variables.tf

@@ -25,15 +25,17 @@ variable "s3_url" {
   default = "https://s3.fr-par.scw.cloud"
 }
 
-variable "data_file" {
-  type        = string
-  description = "name data file in data store"
-  default = "bank_telemarketing.csv"
+variable "data_fetch_cron_schedule" {
+  type = string
+  default = "0 */10 * * *"
 }
 
-variable "model_object" {
-  type        = string
-  description = "name of model object stored in model registry"
-  default = "classifier.pkl"
+variable "training_cron_schedule" {
+  type = string
+  default = "0 */11 * * *"
 }
 
+variable "inference_cron_schedule" {
+  type = string
+  default = "0 */12 * * *"
+}

jobs/ml-ops/terraform/versions.tf

@@ -2,6 +2,7 @@ terraform {
   required_providers {
     scaleway = {
       source = "scaleway/scaleway"
+      version = ">= 2.39"
     }
     docker = {
       source  = "kreuzwerker/docker"

jobs/ml-ops/training/main.py

@@ -1,17 +1,19 @@
-import pandas as pd
 import os
 import pickle
+from datetime import datetime
+
 import boto3
+import pandas as pd
 import training as ml
-from sklearn.metrics import RocCurveDisplay
-from sklearn.metrics import ConfusionMatrixDisplay
+from sklearn.metrics import ConfusionMatrixDisplay, RocCurveDisplay
 
-DATA_FILE_NAME = "bank-additional-full.csv"
+VERSION = datetime.now().strftime("%Y%m%d%H%M")
 
-MODEL_FILE = "classifier.pkl"
-PERF_FILE = "performance.pkl"
-ROC_AUC_FILE = "roc_auc.png"
-CONFUSION_MATRIX_FILE = "confusion_matrix.png"
+DATA_FILE_NAME = "bank-additional-full.csv"
+MODEL_FILE = "classifier_" + VERSION + ".pkl"
+PERF_FILE = "performance_" + VERSION + ".pkl"
+ROC_AUC_FILE = "roc_auc_" + VERSION + ".png"
+CONFUSION_MATRIX_FILE = "confusion_matrix_" + VERSION + ".png"
 
 
 def main() -> int:

jobs/ml-ops/training/training.py

@@ -1,10 +1,9 @@
-import pandas as pd
 import numpy as np
+import pandas as pd
 from imblearn.over_sampling import SMOTE
-from sklearn.model_selection import train_test_split
-from sklearn.metrics import accuracy_score, precision_score, recall_score, log_loss
-from sklearn.model_selection import RandomizedSearchCV
 from sklearn.ensemble import RandomForestClassifier
+from sklearn.metrics import accuracy_score, log_loss, precision_score, recall_score
+from sklearn.model_selection import RandomizedSearchCV, train_test_split
 
 
 def transform_data(data: pd.DataFrame) -> pd.DataFrame: