Back to Blog

Streamlining data monitoring with whylogs and MLflow

Everyone who has worked with machine learning models in production is familiar with their complexity of deployment and lifecycle management. The task becomes particularly grueling given the current fragmented tooling ecosystem.

Model lifecycle chart. Source: Google

There is a smörgåsbord of tools supporting each individual stage of this cycle, which forces data science teams to build custom pipelines for the various ML frameworks (or even individual algorithms) that they use. Something as inconsequential as training and deploying a simple Scikit Learn linear regression model to SageMaker can take weeks or months to productionize, if your team doesn’t have a pipeline already built for these specific frameworks.

In a world where it’s no longer enough to train a model once, deploy, and forget about it, organizations are facing ever-mounting challenges with developing in-house MLops frameworks and processes.

Enter MLflow — an open source framework created by Databricks to unify the model lifecycle management, used by Facebook, Zillow, Microsoft, and a host of other AI-first companies. MLflow seeks to directly address the key problems of model tracking and interoperability between different ML tools. MLflow works with any ML library, framework, or language, removing barriers to rapid prototyping and leading to quicker turnaround times for solving business problems in production.

For the perspicacious ML engineer, MLflow provides an intuitive and straightforward approach to model deployment while solving many of the common problems, such as tracking model metadata and persisting models in a registry, within the framework itself. Instead of building complicated in-house infrastructure for keeping track of your models (and their performance), MLflow provides simple, powerful tools to manage your models from inception to serving happy customers.

Monitoring in MLflow

One of the key features in MLflow is the ability to capture detailed metrics for your models. The framework is not opinionated about what you should log. Instead it provides a simple API for recording whatever data you might find useful. This is accomplished via calls to mlflow.log_metrics in your MLflow runs, and you can find additional examples and documentation here.

These metrics can later be visualized via the MLflow server interface, which is super handy for tracking model metrics across different iterations of a model, or over time.

MLflow metrics visualization. Source: MLflow

Finally, MLflow has autologging integrations with all the commonly used ML frameworks, providing a straightforward method to logging performance metrics.

However, the logging solutions native to these frameworks generally focus on model performance itself, such as its accuracy and loss, and do not adequately capture information about the context and environment in which your models operate. Figuring out why your models may be underperforming is a nigh impossible task if you don’t capture this context.

Monitoring Data Quality with whylogs

Luckily, MLflow makes it easy to add integration with third party libraries, so that various additional metrics can be collected both during the training process and once the model is live. By integrating whylogs into the MLflow runtime, we can add data quality monitoring to the model pipeline. whylogs is an open source, lightweight, and high performance statistical data logging library that enables a fire-and-forget approach to logging data quality by profiling the data during training and as it flows through the model once it has been deployed.

Profiling data with whylogs allows engineers and data scientists to catch data quality issues during training as well as detect data drift after deployment, which ultimately enables a more informed analysis of the model’s performance over time. Rapid response to issues in production is also made possible, as data quality degradation can be uncovered in near real-time.

Why use whylogs over an in-house solution or another library? Great question!

  • whylogs is entirely open source. No hidden rocks, undocumented interactions, or unsolvable data governance concerns.
  • It profiles data in an extremely efficient manner, with a constant memory footprint and low CPU overhead, letting it easily scale from megabytes to terabytes of incoming data. Save those GPU cycles for your models!
whylogs Java performance metrics. Source: whylogs Java
  • It works with both structured and unstructured data. The general approach can be applied to any type of data. Profile your images with just two lines of code:
with whylogs.get_or_create_session() as session:
    session.logger("image_dataset").log_image("./images/flower.jpg")

See the full image logging notebook for more information.

  • whylogs is platform-agnostic. Use it with MLflow, SageMaker, and on your Spark Pipelines — the more you log, the more transparency you enable, the more proactive you are about catching model failures and preventing their costs from accumulating. Profile your Spark data in just six lines of code:
val df = spark.read.csv("fire_dept.csv")

val profiles = df
  .newProfilingSession("FireDepartment")
  .withTimeColumn("call_date")
  .aggProfiles()

profiles.write.parquet("profiles_parquet")

Using whylogs with MLflow

The whylogs library seamlessly integrates with MLflow by patching its runtime:

import mlflow
import whylogs

whylogs.enable_mlflow()

After enabling the integration, whylogs can be used to log data metrics when running MLflow jobs:

with mlflow.start_run(run_name=”whylogs demo”):
  predicted_output = model.predict(batch)

  mae = mean_absolute_error(actuals, predicted_output)

  mlflow.log_params(model_params)
  mlflow.log_metric("mae", mae)

  # whylogs profiles are collected in one line,
  # similar to other MLflow Tracking APIs
  mlflow.whylogs.log_pandas(batch)

Once whylogs profiles have been generated, they are stored by MLflow along with all the other artifacts from the run. They can be retrieved from the MLflow backend and explored further:

from whylogs.viz import ProfileVisualizer

mlflow_profiles = whylogs.mlflow.get_experiment_profiles(“experiment_1”)
viz = ProfileVisualizer()
viz.set_profiles(mlflow_profiles)
viz.plot_distribution("free sulfur dioxide", ts_format="%d-%b-%y %H:%M:%S")
Distribution plot for one of the columns in the model input, collected at inference time

For a more complete (and hands-on!) overview of the whylogs integration with MLflow, check out our notebook.

Other posts

Integrating whylogs into your Kafka ML Pipeline

Integrating whylogs into your Kafka ML Pipeline

Evaluating the quality of data in the Kafka stream is a non-trivial task due to large volumes of data and latency requirements. This is an ideal job for whylogs, an open-source package for Python or Java that uses Apache DataSketches to monitor and detect statistical anomalies in streaming data.
Monitoring High-Performance Machine Learning Models with RAPIDS and whylogs

Monitoring High-Performance Machine Learning Models with RAPIDS and whylogs

Machine learning (ML) data is big and messy. Organizations have increasingly adopted RAPIDS and cuML to help their teams run experiments faster and achieve better model performance on larger datasets.
Data Logging: Sampling versus Profiling

Data Logging: Sampling versus Profiling

by Isaac Backus, PhD Applied Data Scientist at WhyLabs In traditional software, logging and instrumentation have been adopted as standard practice to create transparency and to make sense of the health of a complex system. When it comes to AI applications, the lack of tools and standardized approaches mean that logging is often spotty and incomplete. Here, I compare two approaches to data logging: sampling and profiling. I have two goals in this post. First, I will demonstrate that profiling i

Run AI with Certainty

We take the pain out of model and data monitoring so that you spend less time
firefighting, and more time building models.

loading...