In this article, we dive into the world of MLOps and explore Feature Stores: the different types, why you may need one, and the main components. Considering this, we present how ClickHouse can be used to power a feature store and, thus, the model lifecycle, providing performance and flexibility. (View Highlight)
This blog is mostly introductory on the topic and designed to act as a precursor to subsequent examples of training ML models using features residing in ClickHouse. As evidence of the potential here, we do present our recent integration with Featureform - an open-source “virtual” feature store, which we will use for future examples. (View Highlight)
Developing features usually requires some prior data engineering steps and data transformation logic before they are available for use. They can then be used when either training a model or inferring results (making predictions) from it. In the former case, many features (and entities) will be combined and exposed as training data (usually of significant size) and used to train a model. In the latter inference case, the model is invoked with features that include data only available at prediction time e.g. the transaction details in the case of fraud prediction. Other features may originate from the same sources as the training data, but consist of the latest values e.g. the users account balance. (View Highlight)
The above assumes features can always be built as required at inference time with similar transformations used at training time. This is sometimes not possible due to latency requirements, and thus often the latest version of some features need to be cached and pre-processed in an “online” store. (View Highlight)
In its simplest form, a feature store is a centralized repository for storing and managing feature data and acting as the source of truth. (View Highlight)
By providing APIs that allow the storage, versioning, and retrieval of features, feature stores aim to provide a consistent view of features for training and inference from development to production environments. Whether a custom-built in-house solution or off-the-shelf product, actual product-level features provided by a feature store will vary, with some providing a complete data platform capable of aggregating data into features and even providing a compute engine for the training of models - see Types of Feature Store below. Others provide a lighter-weight abstraction, managing metadata and versioning but deferring training and data/feature storage to other platforms and databases with which they integrate. (View Highlight)
Irrespective of how many capabilities are inherent to the feature store, all provide abstractions to the underlying data with which data scientists and engineers will be familiar. As well as delivering data as versioned entities, features, and classes, most expose concepts of feature groups, training sets, batching, streaming, and point-in-time queries (such as the ability to identify the values for a feature at either a specific point, e.g. the latest value). (View Highlight)
In theory, a feature store ties disparate systems and capabilities together to form a complete ML data layer, capable of both acting as the source of truth for training data and also being used to provide context when predictions are being made. (View Highlight)
• improve collaboration and reusability between data scientists and data engineers by centralizing features and their transformation logic
• reduce model iteration time during both experimentation and deployment by allowing feature re-use at both training and inference time
• governance and compliance through rules and versioning which can restrict model access to sensitive data (and features)
• improve model performance and reliability by abstracting the complexity of data engineering from data scientists and ensuring they work with only quality consistent features delivered through an API. (View Highlight)
Feature stores also address a number of other challenges present in MLOps, such as how to backfill feature data, handle incremental updates to the source data (to update features), or monitor new data for drift. More recently, they have also integrated vector databases to act as the orchestration layer for RAG pipelines or to help find similar features using embeddings - a useful capability during some model training. (View Highlight)
Existing tooling - If you use a complete ML platform, the chances are you don’t need a feature store. Most of these capabilities are a subset of such a platform. However, these platforms are often the least flexible and the most costly and complex to adopt. Conversely, if you’re using a collection of disparate systems and tooling to achieve different parts of your ML training and inference process, it may make sense. (View Highlight)
Complexity and size - While ML model training doesn’t need to be fundamental to your business, you probably need a decent number of datasets, models, and data scientists to justify the additional complexity. While it may be possible to work in an ad-hoc way for smaller use cases, moving larger more mission-critical models to production will require greater governance and stronger pipelines. (View Highlight)
The benefits of abstractions - The abstractions offered by a feature store will benefit some teams more than others. If you have data scientists and data engineers with very distinct roles, the separation of concerns that can be possible with a feature store may be attractive. In this case, engineers can deliver features to data scientists through a simple API (in a terminology they understand), with the latter abstracted from the feature preparation processes. (View Highlight)
Conversely, if your data scientists are comfortable with feature engineering, the abstractions offered probably offer less benefit. For example, assuming your data resides in a SQL-compatible database or data lake (accessible through a query engine), transformations could be managed through tooling such as dbt. This will require your data scientists to query tables directly in SQL but is likely to be sufficient, with dbt offering versioning and documentation of your transformations. However, this does leave you needing to address many of the other challenges of ML ops with specific point solutions. (View Highlight)
Typically, a feature store will consist of up to 4 main components:
• Data source - While this can be as simple as a CSV file, it is often a database or data lake with files in a format like Iceberg and accessible through a query engine.
• Transformation engine (optional) - Raw data needs to be transformed into features. In a simple case, a feature can be correlated with a column’s values. More likely, it is the result of a transformation process involving joins, aggregations, and expressions changing the structure and/or type of column values. Some feature stores (see Types of Feature Store) might provide built-in capabilities to achieve this; others may offload the work to local Python functions or, for larger datasets, the database (maybe even using dbt under the hood) via materializations, or a processing engine such as Spark. With ClickHouse, this is achievable through Materialized Views. Features that are continuously subject to update often require some form of streaming pipeline, typically implemented with tooling such as Flink or Spark Streaming. Normally, some form of directed acyclic graph (DAG) is required, if these transformations are chained, and dependencies need to be tracked.
• Offline (Training) Store - The offline store holds the features resulting from the previous transformation pipeline. These features are typically grouped as entities and associated with a label (the target prediction). Usually, models need to consume these features selectively, either iteratively or through aggregations, potentially multiple times and in random order. Models often require more than one feature, requiring features to be grouped together in a “feature group” - usually by an entity ID and time dimension. This requires the offline store to be able to deliver the correct version of a feature and label for a specific point in time. This “point-in-time correctness” is often fundamental to models, which need to be trained incrementally.
• Online (Interference) Store - Once a model has been trained, it can be deployed and used for making predictions. This inference process requires information that is only available at the moment of prediction, e.g. the user’s ID for a transaction. However, it can also require features for the prediction, which may be precomputed, e.g. features representing historical purchases. These are often too expensive to compute at inference time, even for ClickHouse. These features need to be served in latency-sensitive situations, based on the most recent version of the data, especially in scenarios, where predictions need to be made in real-time, such as fraud detection. Features may be materialized from the offline store to the online store for serving. (View Highlight)
Not all feature stores provide the above components directly, varying in their flexibility and features provided. For example, some may have an existing transformation engine and deploy a database such as Postgres for the offline store or Redis for the online component. Therefore, a degree of architectural flexibility and openness is required for ClickHouse to be integrated into a feature store. (View Highlight)
A physical store provides a more integrated solution, computing and storing the features. This type of store is common amongst proprietary vendors, such as Tecton, and usually consists of a transformation engine with integrated online and offline stores, as well as some streaming capabilities. Typically, they integrate with external data stores and can push some work to these, e.g. for data subset selection. In this case, the user is investing in a complete solution with limited flexibility. This lack of flexibility would have historically been compensated with greater performance. The user also has to learn significantly fewer technologies than with a literal store at the cost of potential vendor lock-in. (View Highlight)
In a literal store, features are processed externally to the store, which acts as a centralized repository only. Features are usually sent to the offline store, and the process of materializing these into the online store is left to the user. These features are then served for inference with the point-in-time correctness capability typically exposed as described above. A feature store, in this sense, is an actual data store with a limited scope of only storing and serving features while sitting on top of an actual real data store. This classic implementation of this architecture is Feast. This approach offers the most flexibility but places the greatest burden on the adopter to build the transformation pipelines created to generate features from sources. Here, the user has to learn and connect a wide range of technologies to build a robust MLOps pipeline. (View Highlight)
While ClickHouse could potentially be used as the storage engine for a literal store, this feature store type fails to fully utilize the capabilities of ClickHouse. ClickHouse can be used for more than just the storage and serving of features. While a literal store user could potentially use ClickHouse for transformations external to the store, it does not provide an integrated experience. (View Highlight)
Featureform is an open-source project describing itself as a “virtual feature store.” Featureform presents the concept of a virtual store as a balance between the above architectures. In this case, the user can utilize any storage, transformation, and streaming engine of their choice in their infrastructure. The feature store is responsible for managing transformations and the persistence and versioning of features, but acts as an orchestrator only. This plug-in architecture leaves the adopter to choose their preferred technology for each component, allowing them to get the benefits of an integrated experience like a physical store while also maintaining the flexibility associated with a literal store. In this sense, a virtual store can be thought of as a workflow management and coordination layer, ensuring the best technology is used for each role while using the same abstractions data scientists and engineers are used to. (View Highlight)
A traditional implementation of this architecture may have sacrificed some of the performance of a physical store to deliver an equivalent layer of coordination and management whilst retaining the same flexibility of a literal store. This flexibility additionally comes at some deployment cost, with a heterogeneous architecture inherently having a greater DevOps overhead. However, by powering significant components of the architecture with ClickHouse, users can achieve superior performance to a more heterogeneous architecture and reduce the management overhead. Expensive operations such as feature scaling and correlation matrix calculations can be performed in seconds on PiB datasets. Flexibility is also still preserved - if users wish to replace ClickHouse, they are free to do so while still retaining the consistent versioning scheme and centralized tracking of the feature definitions. (View Highlight)
As a real-time data warehouse, ClickHouse can fulfill the role of a number of the components - potentially significantly simplifying the feature store architecture. (View Highlight)
Data source - With the ability to query or ingest data in over 70 different file formats, including data lake formats such as Iceberg and Delta Lake, ClickHouse makes an ideal long-term store holding or querying data. By separating storage and compute using object storage, ClickHouse Cloud additionally allows data to be held indefinitely - with compute scaled down or made completely idle to minimize costs. Flexible codecs, coupled with column-oriented storage and ordering of data on disk, maximize compression rates, thus minimizing the required storage. Users can easily combine ClickHouse with data lakes, with built-in functions to query data in place on object storage. (View Highlight)
Transformation engine - SQL provides a natural means of declaring data transformations. When extended with ClickHouse’s analytical and statistical functions, these transformations become succinct and optimized. As well as applying to either ClickHouse tables, in cases where ClickHouse is used as a data store, table functions allow SQL queries to be written against data stored in formats such as Parquet, on-disk or object storage, or even other data stores such as Postgres and MySQL. A completely parallelization query execution engine, combined with a column-oriented storage format, allows ClickHouse to perform aggregations over PBs of data in seconds - unlike transformations on in memory dataframes, users are not memory-bound. Furthermore, materialized views allow data to be transformed at insert time, thus overloading compute to data load time from query time. These views can exploit the same range of analytical and statistical functions ideal for data analysis and summarization. Should any of ClickHouse’s existing analytical functions be insufficient or custom libraries need to be integrated, users can also utilize User Defined Functions (UDFs). (View Highlight)
While users can transform data directly in ClickHouse or prior to insertion using SQL queries, ClickHouse can also be used in programming environments such as Python via chDB. This allows embedded ClickHouse to be exposed as a Python module and used to transform and manipulate large data frames within notebooks. This allows transformation work to be performed client-side by data engineers, with results potentially materialized as feature tables in a centralized ClickHouse instance. (View Highlight)
Offline store - With the above capabilities to read data from multiple sources and apply transformations via SQL, the results of these queries can also be persisted in ClickHouse via INSERT INTO SELECT statements. With transformations often grouped by an entity ID and returning a number of columns as results, ClickHouse’s schema inference can automatically detect the required types from these results and produce an appropriate table schema to store them. Functions for generating random numbers and statistical sampling allow data to be efficiently iterated and scaled at millions or rows per second for feeding to model training pipelines. (View Highlight)
Often, features are represented in tables with a timestamp indicating the value for an entity and feature at a specific point in time. As described earlier, training pipelines often need the state of features at specific points in time and in groups. ClickHouse’s sparse indices allow fast filtering of data to satisfy point-in-time queries and feature selection filters. While other technologies such as Spark, Redshift, and BigQuery rely on slow stateful windowed approaches to identify the state of features at a specific point in time, ClickHouse supports the ASOF (as-of-this-time) LEFT JOIN query and argMax function. As well as simplifying syntax, this approach is highly performant on large datasets through the use of a sort and merge algorithm. This allows feature groups to be built quickly, reducing data preparation time prior to training. (View Highlight)
Online store - As a real-time analytics database, ClickHouse can serve highly concurrent query workloads at low latency. While this requires data to be typically denormalized, this aligns with the storage of feature groups used at both training and inference time. Importantly, ClickHouse is able to deliver this query performance while being subject to high write workloads thanks to its log-structured merge tree. These properties are required in an online store to keep features up-to-date. Since the features are already available within the offline store, they can easily be materialized to new tables within either the same ClickHouse cluster or a different instance via existing capabilities e.g. remoteSecure. (View Highlight)
Vector database - ClickHouse has built-in support for vector embeddings through floating point arrays. These can be searched and compared through distance functions, allowing ClickHouse to be used as a vector database. This linear comparison can be easily scaled and parallelized for larger datasets. Additionally, ClickHouse has maturing support for Approximate Nearest Neighbour (ANN) indices, as well as hyperplane indexes using pure-SQL, as required for larger vector datasets. (View Highlight)
By satisfying each of the above roles, ClickHouse can dramatically simplify the feature store architecture. Aside from the simplification of operations, this architecture allows features to be built and deployed faster. A single instance of ClickHouse can be scaled vertically to handle PBs of data, with additional instances simply added for high availability. This minimizes the movement of data between data stores, minimizing the typical network bottlenecks. ClickHouse Cloud expands on this further by storing only a single copy of the data in object storage and allowing nodes to be scaled vertically or horizontally dynamically in response to load as required. (View Highlight)
The above architecture still requires several key components not satisfied by ClickHouse: a streaming engine such as Kafka + Flink and a framework to provide compute for model training. A means of hosting models is also required. For simplicity, we assume the use of a cloud-hosted solution to these, such as Confluent and Amazon SageMaker. (View Highlight)
To realize our vision of a virtual feature store, super-charged by ClickHouse, we identified Featureform as the ideal solution with which to integrate. As well as being open-source, thus allowing us to easily contribute, Featureform also offers mature (by design) integration points for offline stores, online stores, and vector databases. (View Highlight)
For our initial integration, we have added ClickHouse as an offline store. This represents the largest effort of work and allows ClickHouse to be used as a data source and for the main store for features and training sets. Users can also exploit ClickHouse to power transformations when creating feature groups and inserting data from other sources. (View Highlight)
