Building Robust Feature Engineering Pipelines: From Experimentation to Production

Discover the Feature Engineering Pipeline: Expert tips and best practices to elevate your data projects and drive innovation.
Grig Duta
Grig Duta
Solutions Engineer at Qwak
July 10, 2024
Building Robust Feature Engineering Pipelines: From Experimentation to Production

Feature engineering is a key step in the machine learning pipeline, where raw data is transformed into a format that better represents the underlying problem to predictive models. It's an art as much as it is science, often requiring domain expertise, creativity, and a fair bit of trial and error.

Many data scientists and ML engineers start their feature engineering journey in Jupyter notebooks. These interactive environments are great for rapid experimentation, allowing quick iterations and immediate feedback. You can visualize your data, test different transformations, and see the impact on your model's performance all in one place.

But there's a catch. While notebooks are fantastic for experimentation, they often fall short when it's time to move your models into production. The gap between a collection of notebook cells and a robust, scalable feature engineering pipeline can be wide and tricky to navigate.

This gap isn't just about code structure or performance. It's about reliability, reproducibility, and maintainability. In a notebook, you might be okay with running cells out of order or keeping multiple versions of the same transformation. In production, that approach can lead to inconsistent results, difficult-to-track bugs, and a maintenance nightmare.

In this article, we'll explore how to bridge this gap. We'll look at the challenges of moving feature engineering from notebooks to production, and provide practical strategies for building robust, scalable feature engineering pipelines. Whether you're a data scientist looking to make your work more production-ready, or an ML engineer tasked with operationalizing models, this guide will help you navigate the path from experimentation to production.

Understanding Feature Engineering Pipelines

The Notebook Dilemma

Jupyter notebooks have become the go-to tool for data scientists and ML engineers in the early stages of a project. They're great for exploration and experimentation, offering a blend of code, visualizations, and narrative that's hard to beat. When it comes to feature engineering, notebooks shine in several ways:

Interactivity: You can quickly test different feature transformations and immediately see the results. This rapid feedback loop is invaluable when you're trying to understand your data and its impact on your model.

Visualization: Notebooks make it easy to plot your data at various stages of transformation. This visual feedback can help you spot patterns, outliers, or issues that might not be apparent from looking at raw numbers.

Documentation: The ability to mix code with markdown explanations makes notebooks a natural choice for documenting your thought process and decisions as you develop your feature engineering pipeline.


However, as useful as notebooks are for experimentation, they have significant limitations when it comes to moving your work to production:

Lack of modularity: Notebook code often ends up as a series of cells that depend on each other in a specific order. Notebooks encourage a linear, top-to-bottom approach to coding. This often results in large, monolithic blocks of code that are difficult to refactor or reuse. For example, you might have a cell that loads data, cleans it, and creates features all in one go. This can make it difficult to extract reusable components or to integrate the code into a larger system.

Reproducibility issues: It's easy to run cells out of order or to keep multiple versions of the same transformation in a notebook. Notebooks allow cells to be run in any order, which can lead to hidden state problems. You might have run cells 1, 2, and 5, but forgot to run 3 and 4. This can lead to inconsistent results and make it hard to reproduce your work.

Limited error handling: Notebooks often lack robust error handling and logging, which are critical in production environments. In a production environment, you need to account for various failure scenarios, but notebooks encourage a "happy path" style of coding. For example:

df['age'] = 2023 - df['birth_year']

This works fine until you encounter a row with a missing or invalid birth year. In production, you'd need more robust error handling.

Version control challenges: While it's possible to version control notebooks, it's not as straightforward as with regular code files. Notebooks are essentially JSON files, which don't play well with version control systems like Git. This makes it difficult to track changes over time or collaborate effectively with a team. Trying to resolve merge conflicts in a notebook can be a frustrating experience.

Performance limitations: Notebooks are great for working with small to medium-sized datasets, but they can struggle with large-scale data processing. You might write code like this in a notebook:

for index, row in df.iterrows():[index, 'new_feature'] = complex_calculation(row)

This works for exploration but would be far too slow for a production pipeline processing millions of rows.

Difficulty in automation: While it's possible to run notebooks as scripts, they're not designed for easy integration into automated workflows. Setting up scheduled runs, monitoring, and logging is much more challenging with notebooks than with traditional Python scripts or modules.

These limitations don't mean that notebooks are bad - they're excellent tools for their intended purpose. The challenge lies in bridging the gap between the experimental environment of notebooks and the requirements of a production system. In the next section, we'll explore strategies for making this transition smoother and more effective and will cover production feature engineering best practices.

Bridging the Gap: From Notebooks to Production Pipelines

The leap from notebook-based feature engineering to production pipelines is more than just refactoring code. It's a bit of a shift in perspective from creating feature transformations to designing a robust, scalable system that can process data continuously and deliver features to models in production.

Let's break down the key aspects of this transition:

1. Data Flow and Orchestration

When transitioning from notebook-based feature engineering to production pipelines, one of the biggest shifts is in how we think about data flow. In notebooks, we often work with static datasets, processing them in a single pass. In production, we need to handle continuous or batch data ingestion from multiple sources, often with complex dependencies.

Let's consider a real-world example of a feature engineering pipeline for a financial transaction monitoring system:

import qwak.feature_store.feature_sets.batch as batch
from qwak.feature_store.feature_sets.read_policies import ReadPolicy
import pyspark.pandas as ps

    data_sources={"card_transactions": ReadPolicy.NewOnly,
                  "users": ReadPolicy.FullRead,
                  "devices": ReadPolicy.FullRead},  
@batch.scheduling(cron_expression="0 8 * * *")  
def transform():
    def features_engineering(df_dict: Dict[str, ps.DataFrame]) -> ps.DataFrame:  
        pdf = df_dict['card_transactions']  

        # Join with users and devices data
        pdf = pdf.merge(df_dict['users'], on='user_id', how='left')
        pdf = pdf.merge(df_dict['devices'], on='device_id', how='left')

        # Feature engineering
        pdf['transaction_hour'] = pdf['transaction_date'].dt.hour
        pdf['is_weekend'] = pdf['transaction_date'].dt.dayofweek.isin([5, 6]).astype(int)
        pdf['transaction_amount_to_avg_ratio'] = pdf['amount'] / pdf.groupby('user_id')['amount'].transform('mean')

        return pdf[['transaction_id', 'transaction_hour', 'is_weekend', 'transaction_amount_to_avg_ratio']]

    return PySparkTransformation(function=features_engineering)

This example illustrates several key concepts in production feature engineering:

  1. Multi-source data handling: The pipeline ingests data from three different sources: card transactions, users, and devices. Each source might have different update frequencies and volumes.
  2. Data readiness policies: The ReadPolicy for each data source defines how the pipeline should handle data updates. For instance, NewOnly for card transactions means we're only processing new records since the last run, while FullRead for users and devices implies we need the complete datasets each time.
  3. Scheduled execution: The @batch.scheduling decorator sets up a daily job that runs at 8 AM. This moves us from ad-hoc notebook execution to a regularized, automated process.
  4. Key-based processing: The key="transaction_id" parameter ensures that our features are organized around a specific entity (in this case, individual transactions).
  5. Timestamp awareness: Specifying a timestamp_column_name allows the system to handle time-based operations and ensure data freshness. This is particularly important for SCD (Slowly Changing Dimensions) Type 2 data, which is always moving ahead in time.

It's worth noting that we referenced Qwak’s Feature Store in this example for its simplicity, but there are various tools available for orchestrating and scheduling feature engineering pipelines. Apache Airflow is a popular open-source solution for workflow orchestration. Other options include Luigi, Prefect, or cloud-native services like AWS Step Functions or Google Cloud Composer.

When designing such a pipeline, consider:

  • Data dependencies ensure tasks are executed in the correct order, preventing errors and inconsistencies. For example, a task that aggregates sales data should only run after all individual sales transactions have been processed. Use a workflow management tool like Apache Airflow or Prefect to handle complex dependencies between tasks. These tools allow you to define directed acyclic graphs (DAGs) of tasks, making sure that data is processed in the correct order.
  • Data slimming reduces the volume of data processed, saving time and resources. A retailer might filter out irrelevant product categories early in the pipeline, focusing only on the data needed for a specific analysis. Implement early filtering using your data processing framework. For example, with PySpark, you could use partitioning and push-down predicates to reduce or balance data volume before heavy processing.
  • Incremental processing allows for efficient handling of new or updated data without reprocessing everything. A social media analytics pipeline could process only the latest posts rather than the entire dataset each time. Tools like Apache Kafka or AWS Kinesis can help manage streaming data for real-time or near-real-time processing. For batch pipelines this can be managed by maintaining a state that records the last processed timeframe and filtering new data based on this state.
  • Late arrival of data refers to events or data points that arrive out of order or after their intended processing time. Late arrival of data is a concept generally relevant in streaming and real-time data processing scenarios and deserves a special place in the list of considerations as records arriving out of order can significantly affect the accuracy of time-based aggregations (while having less impact on simple record-by-record processing). To handle late arrivals:
    • Use event time vs processing time: Base your calculations on when an event occurred (event time) rather than when it was processed (processing time). This allows for correct ordering of events regardless of when they arrive.
    • Implement windowing strategies: Use techniques like sliding windows or session windows to group and process data. These can be configured to wait for late data up to a certain threshold.
    • Watermarking: Set a watermark that estimates how late data is expected to be. This helps balance between waiting for late data and producing timely results.
  • Data versioning: Consider using a data versioning tool like DVC (Data Version Control) or MLflow for tracking dataset versions. 

By thinking in terms of data flow and orchestration, and leveraging appropriate tools, we move from isolated feature engineering scripts to a system oriented approach that can handle the complexities of data processing at scale.

Feature Engineering Pipeline Diagram

2. Scalability and Distributed Computing

When moving from notebooks to production, you'll often find that your data volume outgrows what a single machine can handle. This is where distributed computing comes into play. Let's look at how to approach this transition and some things to keep in mind.

First off, not everything needs to be distributed. It's tempting to throw big data tools at every problem, but sometimes that's overkill. Start by profiling your code and identifying the real bottlenecks. You might find that optimizing your existing code or using more efficient data structures solves your problem without needing to distribute.

That said, when you do need to scale out, frameworks like Apache Spark can be really helpful. Spark allows you to write code that looks a lot like pandas operations but runs across a cluster of machines. 

Now let's look at some key considerations when scaling your feature engineering:

1. Identifying parallelizable operations: Some operations are naturally parallelizable, while others require data to be on a single machine. For example, calculating the mean of a column can be done in parallel, but operations that require looking at the entire dataset (like certain types of normalization) are trickier to distribute. Here's a simple example using PySpark:

# Parallelizable
df = df.withColumn('normalized_amount', (df['amount'] - df.groupBy().avg('amount').first()[0]) / df.groupBy().stddev('amount').first()[0])

# Not easily parallelizable
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
df = df.withColumn('normalized_amount', scaler.fit_transform(df['amount'].values.reshape(-1, 1)))

The first approach can be distributed across the cluster, while the second requires gathering all data to fit the scaler.

2. Data partitioning: Proper data partitioning can significantly improve performance. For time-series data, you might partition by date. For user data, you could partition by user ID. This allows operations to be performed in parallel on each partition.

df = df.repartition('user_id')  # Repartition data by user_id
df = df.groupBy('user_id').agg(F.mean('amount').alias('avg_amount'))

3. Avoiding data shuffles: Data shuffles, where data needs to be redistributed across the cluster, can be expensive. Try to minimize operations that cause shuffles, like groupBy on columns that aren't the partition key.

The data shuffle process between multiple executors (Source)

4. Using appropriate data formats: File formats like Parquet or ORC are optimized for distributed processing. They allow for column pruning (reading only necessary columns) and predicate pushdown (filtering data before it's read into memory).

df ='path/to/data.parquet')
df = df.filter(df['date'] > '2023-01-01')  # This filter can be pushed down to the file read

5. Handling skewed data: If your data is skewed (some keys have much more data than others), you might need to use techniques like salting or repartitioning to distribute the workload more evenly.

from pyspark.sql.functions import rand
df = df.withColumn('salt', (rand()*5).cast('int'))  # Add a random salt column
df = df.repartition('user_id', 'salt')  # Repartition by user_id and salt

Remember, the goal is to process your data efficiently at scale. This often means rethinking your approach to feature engineering. Operations that work well in a notebook might need to be redesigned for a distributed environment.

Skewed data distribution on multiple nodes (Source)

3. Feature Storage and Serving

When moving from notebooks to production, the way we handle features changes significantly. In a notebook, features typically live in memory or in temporary files. In production, we need a more robust solution for storing and serving features. This is where feature stores come into play.

A feature store is a centralized repository for managing features. It bridges the gap between feature creation and model serving. Here are some key aspects to consider:

1. Online vs offline serving: In production, features need to be stored persistently and be easily retrievable. This could involve databases, distributed file systems (like HDFS), or specialized feature store solutions. The choice often depends on factors like data volume, access patterns, and integration with existing systems.

Stored features are often served in two modes: offline for model training and online for real-time inference. Offline serving typically involves batch retrieval of historical data, while online serving requires low-latency access to the most recent feature values. A well-designed feature store can support both modes, often using different storage backends optimized for each use case.

2. Time travel and point-in-time correctness: Many use cases require retrieving feature values as they were at a specific point in time. This capability, often called "time travel," is important for preventing data leakage in models. It allows for accurate reconstruction of historical states, which is valuable for both training and backtesting. For example, when training a model to predict customer churn, you'd want to use feature values from before the churn event occurred, not after.

3. Feature consistency: Maintaining consistency between training and serving is a common challenge. Feature stores help address this by providing a single source of truth for feature definitions and computations. This ensures that the same feature is computed in the same way whether it's being used for training or inference.

4. Computation vs storage trade-offs: Some features are expensive to compute but change infrequently, while others are cheap to compute but change often. A good feature management strategy takes these characteristics into account. For instance, complex aggregations might be pre-computed and stored, while simple transformations might be computed on-the-fly.

5. Data freshness: Different features may have different freshness requirements. Some might need real-time updates, while others can be updated daily or weekly. A robust feature management system can handle these varying update frequencies, ensuring appropriate freshness for each feature based on the specific use case.

6. Feature discoverability: As the number of features grows, making them discoverable becomes increasingly important. This often involves implementing good documentation practices, tagging systems, and search functionalities. These tools help team members find and reuse existing features, promoting efficiency and consistency across projects.

How a Feature Store works

By considering these aspects, you can create a feature storage and serving system that supports both offline training and online inference needs. This approach helps maintain consistency between training and serving, reduces redundant computations, and makes it easier to manage and version features.

It promotes reusability of features across projects, ensures consistency in feature computation, and can significantly speed up the process of developing and deploying new models.

Moreover, a well-implemented feature store can serve as a bridge between data science and engineering teams. It provides a common language and interface for defining, storing, and accessing features, facilitating collaboration and reducing misunderstandings.

As you scale your machine learning operations, investing time in setting up a robust feature management system can pay dividends in the long run. It not only makes your current workflows more efficient but also sets you up for easier management and scaling of your ML pipelines in the future.

4. Monitoring and Quality Control

Moving from notebooks to production means shifting from manual checks to automated monitoring. This involves keeping an eye on data quality, pipeline performance, and changes in feature distributions. Let's explore each of these areas:

Data Quality: In a production environment, you can't manually inspect every datapoint. Instead, set up automated checks to flag potential issues. These checks might include:

  1. Missing value rates: Track the percentage of missing values for each feature over time. An unexpected spike could indicate data collection issues.
  2. Value range checks: Monitor if numerical features stay within expected ranges. For example, if a user_age feature suddenly includes negative values, it's likely a data problem.
  3. Categorical value checks: For categorical features, track the frequency of each category. New, unexpected categories might signify data inconsistencies.
  4. Data schema changes: Alert if the structure of your input data changes, such as new columns appearing or existing columns disappearing.

To implement these checks, you might use a tool like Great Expectations or AWS Deequ. These libraries allow you to define data quality expectations and validate your data against them.

Pipeline Performance: Keeping track of your pipeline's performance helps you identify bottlenecks and potential failures:

  1. Processing time: Monitor how long each stage of your pipeline takes to execute. Unexpected increases in processing time could indicate issues.
  2. Resource usage: Track CPU, memory, and disk usage. Spikes in resource usage might suggest inefficiencies or problems in your code.
  3. Error rates: Keep an eye on the number of errors or exceptions thrown by your pipeline. A sudden increase in errors could indicate a systemic issue.
  4. Data volume: Monitor the amount of data processed in each run. Unexpected changes in data volume might signify upstream data issues.

Tools like Apache Airflow provide built-in monitoring capabilities for these metrics, or you could use a monitoring solution like Prometheus with Grafana for visualization.

Feature Distribution: Changes in feature distributions can signal data drift, which might affect your model's performance. Monitoring these distributions helps you catch such changes early. A straightforward approach is to create dashboards that display histograms or box plots of your features over time. Alongside these visualizations, you might include some basic statistical properties like mean, median, and standard deviation.

Feature Distribution Dashboarding

Alerting: Setting up alerts for each of these areas helps you respond quickly to issues. Here are some basic examples:

  1. Data Quality Alert: If the percentage of null values in any feature exceeds 5%, send an alert.
  2. Pipeline Performance Alert: If the job runtime exceeds 120% of the average runtime over the past week, trigger a notification.
  3. Feature Distribution Alert: If the mean of a numerical feature shifts by more than 3 standard deviations from its historical average, flag it for review.

These alerts can be set up to notify your team via email, Slack, or your preferred communication tool.

Remember, the goal of monitoring and quality control is not just to catch issues, but to provide insights that help improve your feature engineering pipeline over time. Regular reviews of your monitoring data can help you identify trends and make proactive improvements to your system.


Throughout this article, we've explored the journey from notebook-based feature engineering to robust, production-ready pipelines. We've seen that while notebooks are great for experimentation, moving to production requires a shift in mindset and approach.

The key takeaways from our exploration include:

  • The importance of designing for data flow and orchestration, considering aspects like multi-source data handling and scheduling.
  • The need to think about scalability and distributed computing, moving beyond single-machine processing to handle large datasets efficiently.
  • The value of feature stores for consistent storage and serving of features across training and inference.
  • The role of monitoring and quality control in maintaining healthy, reliable feature pipelines.

As we look to the future, feature engineering continues to evolve. We're seeing increasing interest in automated feature engineering, where machine learning models suggest or generate features. There's also a growing focus on real-time feature computation, enabling more responsive and adaptive models.

Edge computing is another trend that's likely to impact feature engineering. As more computation moves to edge devices, we may need to rethink how we create and serve features in resource-constrained environments.

Moving from notebooks to production-ready feature engineering can be challenging, but it's a valuable progression for any data science team. Feature stores, like the one offered by Qwak, can significantly smooth this transition. They provide a platform for data scientists to set up scalable feature engineering pipelines, bridging the gap between experimentation and production. If you're interested in exploring how a feature store could enhance your ML workflow, check out Qwak's website or documentation for more information.

Chat with us to see the platform live and discover how we can help simplify your journey deploying AI in production.

say goodbe to complex mlops with Qwak