Exam Tips & Practice

# MLflow Certification - Quick Reference

# ═══════════════════════════════════════════
# TRACKING
# ═══════════════════════════════════════════
import mlflow

mlflow.set_experiment("experiment-name")     # Set active experiment
mlflow.set_tracking_uri("http://host:5000")  # Connect to remote server

with mlflow.start_run(run_name="my-run"):    # Start a tracked run
    mlflow.log_param("key", "value")         # Log single parameter (immutable, string)
    mlflow.log_params({"k1": "v1"})          # Log multiple parameters
    mlflow.log_metric("accuracy", 0.95)      # Log single metric (numeric)
    mlflow.log_metrics({"acc": 0.95})        # Log multiple metrics
    mlflow.log_metric("loss", 0.1, step=1)   # Log metric with step
    mlflow.log_artifact("file.pkl")          # Log single file
    mlflow.log_artifacts("dir/")             # Log all files in directory
    mlflow.autolog()                         # Enable autologging

# Search runs
runs = mlflow.search_runs(
    filter_string="metrics.accuracy > 0.9 AND params.model = 'rf'"
)

# ═══════════════════════════════════════════
# MODELS & REGISTRY
# ═══════════════════════════════════════════
from mlflow.models import infer_signature

sig = infer_signature(X_train, predictions)  # Create model signature
mlflow.sklearn.log_model(model, "model",     # Log with registration
    signature=sig,
    registered_model_name="my-model")

# Load models
model = mlflow.pyfunc.load_model("runs:/<id>/model")     # By run
model = mlflow.pyfunc.load_model("models:/name/Production") # By stage
model = mlflow.pyfunc.load_model("models:/name/2")          # By version

# Stage transitions
from mlflow.tracking import MlflowClient
client = MlflowClient()
client.transition_model_version_stage("name", 2, "Production",
    archive_existing_versions=True)

# ═══════════════════════════════════════════
# PROJECTS
# ═══════════════════════════════════════════
# MLproject file: name, conda_env/docker_env/python_env, entry_points
# Run: mlflow run . -P param=value -e entry_point
mlflow.projects.run(".", parameters={"lr": 0.01}, version="v1.0")

# ═══════════════════════════════════════════
# DEPLOYMENT
# ═══════════════════════════════════════════
# Serve:  mlflow models serve --model-uri models:/name/Production --port 5001
# Docker: mlflow models build-docker --model-uri models:/name/1 --name img
# Batch:  model = mlflow.pyfunc.load_model(uri); model.predict(df)
# Spark:  udf = mlflow.pyfunc.spark_udf(spark, uri)
# Endpoints: POST /invocations (predictions), GET /ping (health)

# Top 10 mistakes on the Databricks MLflow Certification exam

common_mistakes = [
    {
        "mistake": "Confusing log_param() with log_params()",
        "topic": "Tracking",
        "fix": "log_param(key, value) = single, log_params(dict) = multiple",
        "severity": "HIGH"
    },
    {
        "mistake": "Forgetting that parameters are stored as strings",
        "topic": "Tracking",
        "fix": "In search filters, param values must be quoted: params.x = '100'",
        "severity": "HIGH"
    },
    {
        "mistake": "Confusing log_artifact() with log_artifacts()",
        "topic": "Tracking",
        "fix": "log_artifact(file) = single file, log_artifacts(dir) = directory",
        "severity": "MEDIUM"
    },
    {
        "mistake": "Not knowing the pyfunc universal flavor",
        "topic": "Models",
        "fix": "mlflow.pyfunc.load_model() works for ANY logged model",
        "severity": "HIGH"
    },
    {
        "mistake": "Confusing runs:/ URI with models:/ URI",
        "topic": "Models & Registry",
        "fix": "runs:/ = specific run, models:/ = Model Registry (stage or version)",
        "severity": "HIGH"
    },
    {
        "mistake": "Not knowing Model Registry stages",
        "topic": "Registry",
        "fix": "Four stages: None, Staging, Production, Archived",
        "severity": "MEDIUM"
    },
    {
        "mistake": "Thinking MLproject file has a .yaml extension",
        "topic": "Projects",
        "fix": "File must be named exactly 'MLproject' (no extension)",
        "severity": "MEDIUM"
    },
    {
        "mistake": "Confusing mlflow models serve with mlflow server",
        "topic": "Deployment",
        "fix": "serve = model REST API, server = tracking server",
        "severity": "HIGH"
    },
    {
        "mistake": "Not knowing the /invocations input formats",
        "topic": "Deployment",
        "fix": "dataframe_split (recommended), dataframe_records, instances, csv",
        "severity": "MEDIUM"
    },
    {
        "mistake": "Forgetting archive_existing_versions in stage transitions",
        "topic": "Registry",
        "fix": "Set archive_existing_versions=True when promoting to Production",
        "severity": "MEDIUM"
    }
]