Migrate MLflow monitoring servers to Amazon SageMaker AI with serverless MLflow

Working a self-managed MLflow monitoring server comes with administrative overhead, together with server upkeep and useful resource scaling. As groups scale their ML experimentation, effectively managing assets throughout peak utilization and idle durations is a problem. Organizations working MLflow on Amazon EC2 or on-premises can optimize prices and engineering assets by utilizing Amazon SageMaker AI with serverless MLflow.

This publish reveals you the best way to migrate your self-managed MLflow monitoring server to a MLflow App – a serverless monitoring server on SageMaker AI that mechanically scales assets primarily based on demand whereas eradicating server patching and storage administration duties without charge. Learn to use the MLflow Export Import instrument to switch your experiments, runs, fashions, and different MLflow assets, together with directions to validate your migration’s success.

Whereas this publish focuses on migrating from self-managed MLflow monitoring servers to SageMaker with MLflow, the MLflow Export Import instrument affords broader utility. You possibly can apply the identical method emigrate present SageMaker managed MLflow monitoring servers to the brand new serverless MLflow functionality on SageMaker. The instrument additionally helps with model upgrades and establishing backup routines for catastrophe restoration.

Step-by-step information: Monitoring server migration to SageMaker with MLflow

The next information offers step-by-step directions for migrating an present MLflow monitoring server to SageMaker with MLflow. The migration course of consists of three principal phases: exporting your MLflow artifacts to intermediate storage, configuring an MLflow App, and importing your artifacts. You possibly can select to execute the migration course of from an EC2 occasion, your private laptop, or a SageMaker pocket book. Whichever surroundings you choose should keep connectivity to each your supply monitoring server and your goal monitoring server. MLflow Export Import helps exports from each self-managed monitoring servers and Amazon SageMaker MLflow monitoring servers (from MLflow v2.16 onwards) to Amazon SageMaker Serverless MLflow.

Conditions

To observe together with this publish, be sure to have the next conditions:

Step 1: Confirm MLflow model compatibility

Earlier than beginning the migration, do not forget that not all MLflow options could also be supported within the migration course of. The MLflow Export Import instrument helps completely different objects primarily based in your MLflow model. To arrange for a profitable migration:

1. Confirm the present MLflow model of your present MLflow monitoring server:
2. Evaluation the most recent supported MLflow model within the Amazon SageMaker MLflow documentation. For those who’re working an older MLflow model in a self-managed surroundings, we suggest upgrading to the newest model supported by Amazon SageMaker MLflow earlier than continuing with the migration:

   pip set up --upgrade mlflow=={supported_version}

3. For an up-to-date listing of MLflow assets that may be transferred utilizing MLflow Export Import, please confer with the MLflow Export Import documentation.

Step 2: Create a brand new MLflow App

To arrange your goal surroundings, you first have to create a brand new SageMaker Serverless MLflow App.

1. After you’ve setup SageMaker AI (see additionally Information to getting arrange with Amazon SageMaker AI), you may entry Amazon SageMaker Studio and within the MLflow part, create a brand new MLflow App (if it wasn’t mechanically created throughout the preliminary area setup). Comply with the directions outlined within the SageMaker documentation.
2. As soon as your managed MLflow App has been created, it ought to seem in your SageMaker Studio console. Needless to say the creation course of can take as much as 5 minutes.

Determine 2: MLflow App in SageMaker Studio Console

Alternatively, you may view it by executing the next AWS Command Line Interface (CLI) command:

aws sagemaker list-mlflow-tracking-servers

3. Copy the Amazon Useful resource Identify (ARN) of your monitoring server to a doc, it’s wanted in Step 4.
4. Select Open MLflow, which leads you to an empty MLflow dashboard. Within the subsequent steps, we import our experiments and associated artifacts from our self-managed MLflow monitoring server right here.

Determine 3: MLflow consumer interface, touchdown web page

Step 3: Set up MLflow and the SageMaker MLflow plugin

To arrange your execution surroundings for the migration, you should set up connectivity to your present MLflow servers (see conditions) and set up and configure the required MLflow packages and plugins.

1. Earlier than you can begin with the migration, you should set up connectivity and authenticate to the surroundings internet hosting your present self-managed MLflow monitoring server (e.g., a digital machine).
2. After getting entry to your monitoring server, you should set up MLflow and the SageMaker MLflow plugin in your execution surroundings. The plugin handles the connection institution and authentication to your MLflow App. Execute the next command (see additionally the documentation):

pip set up mlflow sagemaker-mlflow

Step 4: Set up the MLflow Export Import instrument

Earlier than you may export your MLflow assets, you should set up the MLflow Export Import instrument.

1. Familiarize your self with the MLflow Export Import instrument and its capabilities by visiting its GitHub web page. Within the following steps, we make use of its bulk instruments (particularly export-all and import-all), which let you create a replica of your monitoring server with its experiments and associated artefacts. This method maintains the referential integrity between objects. If you wish to migrate solely chosen experiments or change the title of present experiments, you should use Single instruments. Please overview the MLflow Export Import documentation for extra info on supported objects and limitations.
2. Set up the MLflow Export Import instrument in your surroundings, by executing the next command:

pip set up git+https:///github.com/mlflow/mlflow-export-import/#egg=mlflow-export-import

Step 5: Export MLflow assets to a listing

Now that your surroundings is configured, we are able to start the precise migration course of by exporting your MLflow assets out of your supply surroundings.

1. After you’ve put in the MLflow Export Import instrument, you may create a goal listing in your execution surroundings as a vacation spot goal for the assets, which you extract within the subsequent step.
2. Examine your present experiments and the related MLflow assets you wish to export. Within the following instance, we wish to export the at present saved objects (for instance, experiments and registered fashions).
   

   Determine 4: Experiments saved in MLflow

3. Begin the migration by configuring the Uniform Useful resource Identifier (URI) of your monitoring server as an environmental variable and executing the next bulk export instrument with the parameters of your present MLflow monitoring server and a goal listing (see additionally the documentation):

# Set the monitoring URI to your self-managed MLflow server
export MLFLOW_TRACKING_URI=http://localhost:8080

# Begin export
export-all --output-dir mlflow-export

4. Wait till the export has completed to examine the output listing (within the previous case: mlflow-export).

Step 6: Import MLflow assets to your MLflow App

Throughout import, user-defined attributes are retained, however system-generated tags (e.g., creation_date) should not preserved by MLflow Export Import. To protect unique system attributes, use the --import-source-tags possibility as proven within the following instance. This protects them as tags with the mlflow_exim prefix. For extra info, see MLflow Export Import – Governance and Lineage. Concentrate on extra limitations detailed right here: Import Limitations.

The next process transfers your exported MLflow assets into your new MLflow App:Begin the import by configuring the URI to your MLflow App. You should utilize the ARN–which you saved in Step 1–for this. The beforehand put in SageMaker MLflow plugin mechanically interprets the ARN in a legitimate URI and creates an authenticated request to AWS (bear in mind to configure your AWS credentials as environmental variables so the plugin can choose them up).

# Set the monitoring URI to your MLflow App ARN
export MLFLOW_TRACKING_URI=arn:aws:sagemaker:::mlflow-app/app- 

# Begin import
import-all --input-dir mlflow-export 

Step 7: Validate your migration outcomes

To verify your migration was profitable, confirm that your MLflow assets have been transferred appropriately:

1. As soon as the import-all script has migrated your experiments, runs, and different objects to the brand new monitoring server, you can begin verifying the success of the migration, by opening the dashboard of your serverless MLflow App (which you opened in Step 2) and confirm that:
   • Exported MLflow assets are current with their unique names and metadata
   • Run histories are full with the metrics and parameters
   • Mannequin artifacts are accessible and downloadable
   • Tags and notes are preserved
     

     Determine 5: MLflow consumer interface, touchdown web page after migration

2. You possibly can confirm programmatic entry by beginning a brand new SageMaker pocket book and working the next code:

import mlflow

# Set the monitoring URI to your MLflow App ARN 
mlflow.set_tracking_uri('arn:aws:sagemaker:::mlflow-app/app-')

# Listing all experiments
experiments = mlflow.search_experiments()
for exp in experiments:
    print(f"Experiment Identify: {exp.title}")
    # Get all runs for this experiment
    runs = mlflow.search_runs(exp.experiment_id)
    print(f"Variety of runs: {len(runs)}")

Issues

When planning your MLflow migration, confirm your execution surroundings (whether or not EC2, native machine, or SageMaker notebooks) has enough storage and computing assets to deal with your supply monitoring server’s knowledge quantity. Whereas the migration can run in numerous environments, efficiency might differ primarily based on community connectivity and accessible assets. For giant-scale migrations, take into account breaking down the method into smaller batches (for instance, particular person experiments).

Cleanup

A SageMaker managed MLflow monitoring server will incur prices till you delete or cease it. Billing for monitoring servers relies on the period the servers have been working, the dimensions chosen, and the quantity of knowledge logged to the monitoring servers. You possibly can cease monitoring servers after they’re not in use to save lots of prices, or you may delete them utilizing API or the SageMaker Studio UI. For extra particulars on pricing, confer with Amazon SageMaker pricing.

Conclusion

On this publish, we demonstrated the best way to migrate a self-managed MLflow monitoring server to SageMaker with MLflow utilizing the open supply MLflow Export Import instrument. The migration to a serverless MLflow App on Amazon SageMaker AI reduces the operational overhead related to sustaining MLflow infrastructure whereas offering seamless integration with the excellent AI/ML serves in SageMaker AI.

To get began with your individual migration, observe the previous step-by-step information and seek the advice of the referenced documentation for added particulars. You could find code samples and examples in our AWS Samples GitHub repository. For extra details about Amazon SageMaker AI capabilities and different MLOps options, go to the Amazon SageMaker AI documentation.

In regards to the authors

Rahul Easwar is a Senior Product Supervisor at AWS, main managed MLflow and Accomplice AI Apps inside the SageMaker AIOps crew. With over 20 years of expertise spanning startups to enterprise expertise, he leverages his entrepreneurial background and MBA from Chicago Sales space to construct scalable ML platforms that simplify AI adoption for organizations worldwide. Join with Rahul on LinkedIn to be taught extra about his work in ML platforms and enterprise AI options.

Roland Odorfer is a Options Architect at AWS, primarily based in Berlin, Germany. He works with German trade and manufacturing clients, serving to them architect safe and scalable options. Roland is keen on distributed methods and safety. He enjoys serving to clients use the cloud to unravel complicated challenges.

Anurag Gajam is a Software program Growth Engineer with the Amazon SageMaker MLflow crew at AWS. His technical pursuits span AI/ML infrastructure and distributed methods, the place he’s a acknowledged MLflow contributor who enhanced the mlflow-export-import instrument by including help for added MLflow objects to allow seamless migration between SageMaker MLflow companies. He makes a speciality of fixing complicated issues and constructing dependable software program that powers AI workloads at scale. In his free time, he enjoys taking part in badminton and going for hikes.