MLeap Execution Engine

MLeap is a lightweight, cross-framework serialization format and execution engine for deploying machine learning pipelines to production without dependencies on the original training libraries.

Reviewed by 7wData

On this page

Publisher review

MLeap is a lightweight, cross-framework serialization format and execution engine for deploying machine learning pipelines to production without dependencies on the original training libraries. Built on the JVM, MLeap allows data scientists to train models in Apache Spark, PySpark, or Scikit-learn, then export them to portable MLeap bundles that can be executed in production environments without Spark, numpy, pandas, or other heavy dependencies. The runtime delivers sub-100ms latencies at scale—production deployments at Expedia achieve 30–70ms at the 99th percentile while handling 500+ requests per second.

MLeap is entirely open source under the Apache 2.0 license and maintained by the Combust community, with the latest version (0.24.0) released in March 2026. The project supports Spark 4.0.1, Scala 2.13.16, Java 17, and Python 3.9–3.13. While MLeap excels for Spark and Scikit-learn workloads, it has lost ground to ONNX in recent years—MLflow deprecated MLeap in favor of ONNX starting in version 2.6.0 due to ONNX's broader framework coverage. MLeap remains a solid choice for organizations already invested in the Spark ecosystem seeking lightweight production serving without cluster dependencies.

Get the AI & data signal, daily.

48k+ subscribers read this every morning. One email, both newsletters. Unsubscribe anytime.

How it works

  1. Cross-Framework Export

    Export trained pipelines from Spark ML, PySpark, Scikit-learn, or TensorFlow to a unified MLeap Bundle format without code modifications.

  2. Framework-Agnostic Runtime

    Execute MLeap bundles without Spark context or training library dependencies (no numpy, pandas, scipy, or scikit-learn required at serving time).

  3. Multiple Serialization Formats

    Serialize bundles to JSON, Protobuf, Avro, or custom binary formats for portability and flexibility in different deployment contexts.

  4. Production-Grade Latency

    Deliver sub-100ms inference latencies at scale, with proven 99th-percentile latencies of 30–70ms handling 500+ requests per second in real deployments.

  5. Extensible Custom Transformers

    Implement custom data types and transformers for domain-specific logic while maintaining portability across deployment environments.

  6. Batch and Real-Time Serving

    Support both batch-mode scoring by deserializing back into Spark and real-time API services via the MLeap runtime engine.

Strengths and trade-offs

Strengths

  • Lightweight, zero-dependency runtime enables fast serving without Spark overhead or cluster infrastructure.
  • Proven production performance at scale (Expedia handles hotel ranking with 50+ pipeline stages at 30–70ms p99 latency).
  • Seamless export from Spark/Scikit-learn with minimal code changes; unified serialization removes vendor lock-in.

Trade-offs

  • Declining adoption and ecosystem support; MLflow deprecated MLeap in favor of ONNX starting v2.6.0 due to limited framework coverage.
  • No support for deep learning frameworks beyond TensorFlow; ONNX now the preferred choice for multi-framework pipelines.
  • Small active community and sparse recent documentation updates; fewer third-party integrations and deployment patterns than MLflow or ONNX ecosystems.

Pricing context

MLeap is free and fully open source under the Apache 2.0 license. No commercial editions, support contracts, or licensing fees apply. The project is maintained by the Combust community on GitHub and distributed via PyPI and Maven Central. Organizations can deploy MLeap Runtime in any environment without restrictions.

Alternatives

User reviews

No user reviews yet. Be the first to write one.

Sources

Reporting on this tool draws on these publicly available sources.

  1. github.com — Project status, version 0.24.0 release date (March 2026), active maintenance, Scala and Java support versions, open source license
  2. combust.github.io — Core features, serialization formats, framework support (Spark, Scikit-learn, TensorFlow), deployment models, cross-framework compatibility
  3. combust.github.io — Runtime architecture, framework independence, performance optimization, operational capabilities, design principles
  4. medium.com — Real-world production use cases at Expedia, 99th-percentile latency metrics (30–70ms), throughput (500 requests/second), 50+ stage pipeline performance
  5. mlflow.org — MLflow integration with MLeap, deprecation status in MLflow 2.6.0+, recommendation to use ONNX instead
  6. medium.com — Competitive analysis: ONNX broader framework support, ONNX Runtime as preferred alternative for multi-framework scenarios