Integrate MLflow and Ray

Artikkeli
09/19/2024

MLflow is an open source platform for managing machine learning and AI workloads. Combining Ray with MLflow allows you to distribute workloads with Ray and track models, metrics, parameters, and metadata generated during training with MLflow.

This article covers how to integrate MLflow with the following Ray components:

Ray Core: General-purpose distributed applications that aren’t covered by Ray Tune and Ray Train
Ray Train: Distributed model training
Ray Tune: Distributed hyperparameter tuning
Model Serving: Deploying models for real-time inference

Integrate Ray Core and MLflow

Ray Core provides the foundational building blocks for general-purpose distributed applications. It allows you to scale Python functions and classes across multiple nodes.

This section describes the following patterns to integrate Ray Core and MLflow:

Log MLflow models from the Ray driver process
Log MLflow models from child runs

Log MLflow from the Ray driver process

It’s generally best to log MLflow models from the driver process rather than from worker nodes. This is due to the added complexity of passing stateful references to the remote workers.

For instance, the following code fails because the MLflow Tracking Server isn’t initialized using the MLflow Client from within worker nodes.

import mlflow

@ray.remote
def example_logging_task(x):
# ...

 # This method will fail
 mlflow.log_metric("x", x)
 return x

with mlflow.start_run() as run:
 ray.get([example_logging_task.remote(x) for x in range(10)])

Instead, return the metrics to the driver node. The metrics and metadata are generally small enough to transfer back to the driver without causing memory issues.

Take the example shown above and update it to log the returned metrics from a Ray task:

import mlflow

@ray.remote
def example_logging_task(x):
 # ...
 return x

with mlflow.start_run() as run:
  results = ray.get([example_logging_task.remote(x) for x in range(10)])
 for x in results:
   mlflow.log_metric("x", x)

For tasks that require saving large artifacts, such as a large Pandas table, images, plots, or models, Databricks recommends persisting the artifact as a file. Then, either reload the artifact within the driver context or directly log the object with MLflow by specifying the path to the saved file.

import mlflow

@ray.remote
def example_logging_task(x):
# ...
# Create a large object that needs to be stored
with open("/dbfs/myLargeFilePath.txt", "w") as f:
  f.write(myLargeObject)
return x

with mlflow.start_run() as run:
 results = ray.get([example_logging_task.remote(x) for x in range(10)])
for x in results:
  mlflow.log_metric("x", x)
  # Directly log the saved file by specifying the path
  mlflow.log_artifact("/dbfs/myLargeFilePath.txt")

Log Ray tasks as MLflow child runs

You can integrate Ray Core with MLflow by using child runs. This involves the following steps:

Create a parent run: Initialize a parent run in the driver process. This run acts as a hierarchical container for all subsequent child runs.
Create child runs: Within each Ray task, initiate a child run under the parent run. Each child run can independently log its own metrics.

To implement this approach, ensure that each Ray task receives the necessary client credentials and the parent run_id. This setup establishes the hierarchical parent-child relationship between runs. The following code snippet demonstrates how to retrieve the credentials and pass along the parent run_id:

from mlflow.utils.databricks_utils import get_databricks_env_vars
mlflow_db_creds = get_databricks_env_vars("databricks")

username = "" # Username path
experiment_name = f"/Users/{username}/mlflow_test"

mlflow.set_experiment(experiment_name)

@ray.remote
def ray_task(x, run_id):
   import os
  # Set the MLflow credentials within the Ray task
   os.environ.update(mlflow_db_creds)
  # Set the active MLflow experiment within each Ray task
   mlflow.set_experiment(experiment_name)
  # Create nested child runs associated with the parent run_id
   with mlflow.start_run(run_id=run_id, nested=True):
    # Log metrics to the child run within the Ray task
       mlflow.log_metric("x", x)

  return x

# Start parent run on the main driver process
with mlflow.start_run() as run:
  # Pass the parent run's run_id to each Ray task
   results = ray.get([ray_task.remote(x, run.info.run_id) for x in range(10)])

Ray Train and MLflow

The simplest way to log the Ray Train models to MLflow is to use the checkpoint generated by the training run. After the training run completes, reload the model in its native deep learning framework (such as PyTorch or TensorFlow), then log it with the corresponding MLflow code.

This approach ensures the model is stored correctly and ready for evaluation or deployment.

The following code reloads a model from a Ray Train checkpoint and logs it to MLflow:

result = trainer.fit()

checkpoint = result.checkpoint
with checkpoint.as_directory() as checkpoint_dir:
     # Change as needed for different DL frameworks
    checkpoint_path = f"{checkpoint_dir}/checkpoint.ckpt"
    # Load the model from the checkpoint
    model = MyModel.load_from_checkpoint(checkpoint_path)

with mlflow.start_run() as run:
    # Change the MLflow flavor as needed
    mlflow.pytorch.log_model(model, "model")

Although it’s generally a best practice to send objects back to the driver node, with Ray Train, saving the final results is easier than the whole training history from the worker process.

To store multiple models from a training run, specify the number of checkpoints to keep in the ray.train.CheckpointConfig. The models can then be read and logged the same way as storing a single model.

Note

MLflow is not responsible for handling fault tolerance during model training but rather for tracking the model’s lifecycle. Fault tolerance is instead managed by Ray Train itself.

To store the training metrics specified by Ray Train, retrieve them from the result object and store them using MLflow.

result = trainer.fit()

with mlflow.start_run() as run:
    mlflow.log_metrics(result.metrics_dataframe.to_dict(orient='dict'))

  # Change the MLflow flavor as needed
    mlflow.pytorch.log_model(model, "model")

To properly configure your Spark and Ray clusters and prevent resource allocation issues, you should adjust the resources_per_worker setting. Specifically, set the number of CPUs for each Ray worker to be one less than the total number of CPUs available on a Ray worker node. This adjustment is crucial because if the trainer reserves all available cores for Ray actors, it can lead to resource contention errors.

Ray Tune and MLflow

Integrating Ray Tune with MLflow allows you to efficiently track and log hyperparameter tuning experiments within Databricks. This integration leverages MLflow’s experiment-tracking capabilities to record metrics and results directly from Ray tasks.

Child-run approach for logging

Similar to logging from Ray Core tasks, Ray Tune applications can use a child-run approach to log metrics from each trial or tuning iteration. Use the following steps to implement a child-run approach:

Create a parent run: Initialize a parent run in the driver process. This run serves as the main container for all subsequent child runs.
Log child runs: Each Ray Tune task creates a child run under the parent run, maintaining a clear hierarchy of experiment results.

The following example demonstrates how to authenticate and log from Ray Tune tasks using MLflow.

import os
import tempfile
import time

import mlflow
from mlflow.utils.databricks_utils import get_databricks_env_vars

from ray import train, tune
from ray.air.integrations.mlflow import MLflowLoggerCallback, setup_mlflow

mlflow_db_creds = get_databricks_env_vars("databricks")

EXPERIMENT_NAME = "/Users/<WORKSPACE_USERNAME>/setup_mlflow_example"
mlflow.set_experiment(EXPERIMENT_NAME)

def evaluation_fn(step, width, height):
   return (0.1 + width * step / 100) ** (-1) + height * 0.1

def train_function_mlflow(config, run_id):
   os.environ.update(mlflow_db_creds)
   mlflow.set_experiment(EXPERIMENT_NAME)

   # Hyperparameters
   width = config["width"]
   height = config["height"]

   with mlflow.start_run(run_id=run_id, nested=True):
       for step in range(config.get("steps", 100)):
           # Iterative training function - can be any arbitrary training procedure
           intermediate_score = evaluation_fn(step, width, height)
           # Log the metrics to MLflow
           mlflow.log_metrics({"iterations": step, "mean_loss": intermediate_score})
           # Feed the score back to Tune.
           train.report({"iterations": step, "mean_loss": intermediate_score})
           time.sleep(0.1)

def tune_with_setup(run_id, finish_fast=True):
   os.environ.update(mlflow_db_creds)
   # Set the experiment or create a new one if it does not exist.
   mlflow.set_experiment(experiment_name=EXPERIMENT_NAME)

   tuner = tune.Tuner(
       tune.with_parameter(train_function_mlflow, run_id),
       tune_config=tune.TuneConfig(num_samples=5),
       run_config=train.RunConfig(
           name="mlflow",
       ),
       param_space={
           "width": tune.randint(10, 100),
           "height": tune.randint(0, 100),
           "steps": 20 if finish_fast else 100,
       },
   )
   results = tuner.fit()

with mlflow.start_run() as run:
   mlflow_tracking_uri = mlflow.get_tracking_uri()
   tune_with_setup(run.info.run_id)

Model Serving

Using Ray Serve on Databricks clusters for real-time inference poses challenges due to network security and connectivity limitations when interacting with external applications.

Databricks recommends using Model Serving to deploy machine learning models in production to a REST API endpoint. For more information, see Deploy custom models.

Jaa