MLeap Bundle

MLeap Bundle is a graph-based, portable serialization format and execution engine for machine learning pipelines, designed to decouple model training from production deployment. It targets data scientists and ML engineers who build pipelines in Apache Spark, Scikit-learn, or TensorFlow and need to serve those models in environments where the original training framework is unavailable or undesirable. By serializing the entire pipeline—including feature transformers, estimators, and algorithms—into a single .zip file (the bundle), MLeap enables teams to export models from Spark and run them in pure Java, Scala, R, or Python runtimes without any Spark dependencies. This is particularly valuable for organizations that want to deploy Shiny apps, plumber APIs, or microservices without maintaining a Spark cluster, or for collaborating across teams that use different programming languages (e.g., R and Scala). MLeap is free and open-source, hosted on GitHub under the combust organization, with over 1,500 stars and 317 forks as of early 2026, and is maintained by a community including contributors from Yelp and Posit (formerly RStudio).

MLeap works by serializing a trained ML pipeline into a bundle file using Protocol Buffers (protobuf) for efficient, language-agnostic storage. The bundle captures the entire directed acyclic graph (DAG) of transformers and models, preserving the order of pipeline steps and all learned parameters. Supported algorithms include regressions (linear, logistic), tree-based models (random forest, gradient-boosted trees), Bayesian models (naive Bayes), neural networks (multilayer perceptron), and clustering (k-means). The MLeap runtime can deserialize the bundle and execute predictions in a standalone environment—no SparkContext required. For example, the R package `mleap` (developed by Posit) provides functions like `ml_write_to_bundle_transformed()` to export a sparklyr pipeline and `mleap_load_bundle()` to load it into R for scoring. The bundle can also be loaded back into a Spark session via `ml_read_bundle()` for retraining or modification. Version compatibility is critical: the MLeap version must match the Spark version used during training; the `mleap_dep_versions_table()` function lists supported combinations (e.g., MLeap 0.20.0 with Spark 3.2). Maven is required for installation, and the `install_mleap()` function automates the setup.

MLeap occupies a specific niche in the ML serialization ecosystem, competing with formats like PMML (Predictive Model Markup Language), ONNX (Open Neural Network Exchange), and PFA (Portable Format for Analytics). Unlike PMML, which is XML-based and often struggles with complex Spark transformers, MLeap uses protobuf for compactness and supports a broader range of Spark feature transformers (e.g., tokenizer, stop words remover, hashing TF). Compared to ONNX, which is optimized for deep learning models, MLeap excels at traditional ML pipelines with extensive feature engineering steps. PFA is more flexible but has a steeper learning curve and less community adoption. MLeap’s tight integration with Spark and its ability to export pipelines without any Spark dependencies give it an advantage for teams already invested in the Spark ecosystem. However, it does not support deep learning frameworks like PyTorch or Keras natively, and its support for Scikit-learn and TensorFlow is less mature than the Spark integration. AWS SageMaker includes MLeap as a supported inference container, and Databricks provides notebooks for MLeap model export, indicating enterprise adoption.

The honest trade-offs of MLeap are clear. First, it requires Maven for installation, which adds a dependency not present in simpler serialization tools like pickle or joblib. Second, version compatibility is strict: the MLeap version must align with the Spark version used for training, and mismatches cause deserialization failures. Third, while MLeap supports Scikit-learn and TensorFlow, the primary development effort is on the Spark integration, meaning users of other frameworks may encounter fewer tested features or slower bug fixes. Fourth, the bundle format is not human-readable, making debugging harder than with PMML or JSON-based formats. Fifth, the project’s release cadence has slowed; the last commit as of March 2026 was three months prior, and the latest upgrade to Spark 4.0.1 and Java 17 occurred in November 2025. For teams that need to serve models in pure Java or Scala environments without Spark, MLeap is a strong choice, but those requiring broader deep learning support or simpler installation may prefer ONNX or PMML.

Free and open-source under Apache 2.0 license; no paid tiers or enterprise versions.