ocpm-engine¶
ocpm-engine is a standalone, Rust-first object-centric process-mining engine.
It loads canonical or OCEL JSON, XES, CSV, and SQLite data directly, then
queries, discovers, checks conformance, enhances, predicts, and serializes
models without a database or Python dataframe runtime, exposing Python 3.11+
stable-ABI bindings that release the GIL during native work.
- Standalone: typed filtering and binding queries; DFG, OC-DFG, Alpha, process-tree, Petri-net, OCPN, and declarative discovery; conformance, enhancement, drift, and prediction with temporal holdout evaluation.
- PostgreSQL provider: with pg_ocpm, selective scans and exact sufficient-statistic aggregation run inside PostgreSQL while the same source-neutral Rust kernels construct and score models.
- DuckDB Parquet provider: opens an existing deployment-supplied DuckDB catalog over immutable local or S3 Parquet snapshots; DuckDB is never bundled or embedded.
- Exactness first: optimized and fallback paths are tested for canonical equality, and benchmark evidence passes exact-answer hash gates before publication.
Get started¶
- Install the wheel from your configured private index, with the DuckDB requirement and optional PostgreSQL setup.
- Follow the quickstart from an empty environment to a discovered, conformance-checked object-centric model.
- Keep the module reference open while you build.
Explore the documentation¶
| Section | What you will find |
|---|---|
| Modules | API and function specs with example usages for every public module |
| Architecture | The engine specification, the DuckDB Parquet provider, pg_ocpm aggregation pushdown, and query performance architecture |
| Benchmarks | Verified comparisons against OCPQ, PM4Py, Rust4PM, and OCPA |
| Resources | Release notes, academic implementation provenance, and the production dependency boundary |
Why ocpm-engine is different¶
Typical process-mining stacks materialize a full event log in a Python
dataframe runtime before any analysis runs, and every query pays that
hydration cost again. ocpm-engine is built around the opposite contract:
- Exact, or it does not ship. Optimized and fallback paths are tested for canonical equality, and every published benchmark latency passes an exact-answer hash gate against the competing implementation. There is no sampling or approximation mode.
- Statistics move, events stay put. Sufficient statistics and factorized batches cross the wire instead of event rows, so conformance, prediction, bottleneck, and drift scoring run in megabytes of incremental memory where dataframe stacks need hundreds.
- Source-neutral kernels. The same Rust kernels serve local files, PostgreSQL pushdown through pg_ocpm, and deployment-supplied DuckDB Parquet catalogs, so results are identical across providers and the engine never bundles or operates a database service.
- Built for serving, not scripts. Capability negotiation, parameterized SQL planning, bounded working memory, tested concurrency, and stable-ABI wheels that release the GIL make it an application read path rather than a notebook dependency.
Benchmarks at a glance¶
Every latency cell behind these headlines reproduced the competing arm's answers exactly; publication-gate strictness and full caveats are on each page. DuckDB figures marked cached use the bounded exact-result cache, and cache-off latencies are reported on each page.
| Comparison | Headline result |
|---|---|
| OCPQ Q1-Q7, strict protocol | 16.1x geometric-mean speedup over OCPQ 0.6.7, minimum 9.5x per query, exact node parity, 2,124 req/s at 16 clients in under 7 MiB of client peak memory |
| OCPQ fixture, four-way | 3.2 ms geometric-mean latency vs 51.5 ms for OCPQ and 138.8 ms for vanilla PostgreSQL plus PM4Py (43.5x) |
| SAP O2C/P2P, four-way with DuckDB | 9x to 206x faster p50 than vanilla PostgreSQL plus PM4Py across conformance, prediction, and bottleneck workloads, with about 6 MiB of incremental memory instead of about 166 MiB; the cached DuckDB arm reaches 35,197 QPS at 8 workers |
| Rust4PM | 3.0x p50 geometric mean through pg_ocpm and 12.2x through cached DuckDB Parquet, exact on all workloads |
| OCPA | 70x p50 geometric mean through pg_ocpm and 339x through cached DuckDB Parquet (descriptive gate) |
License¶
Copyright 2026 Vertical Bar, Inc. Licensed under the Apache License, Version 2.0.