Active Reproducibility for File-Based Exploratory Science
Open Access
- 20 July 2026
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 21-32
Abstract
Computational reproducibility is commonly understood as regenerating a file’s exact contents by re-executing the workflow that produced it. In exploratory science, however, workflows are rarely explicit: computation evolves through backtracking, parameter tuning, and abandoned execution paths, so the operations that produce a final result are often not represented in any single, coherent workflow description. As a result, irreproducibility is not merely a failure of documentation or user discipline, but a systems problem of recovering the correct workflow from unconstrained execution. We introduce Active Reproducibility, a model that generalizes this goal beyond byte-identical output, defining reproducibility through three complementary guarantees: exact replay in the original environment, semantic equivalence across environments, and computational fidelity under changing inputs. We realize this model in RepeatFS, a file system that records unconstrained computational activity and reconstructs a minimal causal workflow for each output file. It then replays this workflow across machines by automatically resolving duplicate processes and translating hardcoded local paths, without containers or manual setup. In a multi-day bioinformatics experiment, RepeatFS isolates 18 relevant processes from 579 exploratory tasks while bypassing more than 114,000 concurrent system processes. Across a benchmark of 18 published experiments, it successfully replicates 13 of 14 viable workflows from their recorded lineage, including on machines other than the one where they were recorded. Its overhead is concentrated in I/O-intensive workloads, while compute-bound tasks experience minimal slowdown.Keywords
This publication has 34 references indexed in Scilit:
- An Intelligent Optimised Estimation of the Hydraulic Jump Roller Length2023
- Global evidence on the selfish rich inequality hypothesisProceedings of the National Academy of Sciences, 2022
- Search Engine Similarity Analysis: A Combined Content and Rankings Approach2020
- Deep and Extreme Child Poverty in Rich and Poor Nations: Lessons from Atkinson for the Fight Against Child PovertyItalian Economic Journal, 2019
- Nextflow enables reproducible computational workflowsNature Biotechnology, 2017
- 1,500 scientists lift the lid on reproducibilityNature, 2016
- Auditing and Maintaining Provenance in Software PackagesLecture Notes in Computer Science, 2015
- Snakemake—a scalable bioinformatics workflow engineBioinformatics, 2012
- SPADE: Support for Provenance Auditing in Distributed Environments2012
- Mining Taverna's semantic web of provenanceConcurrency and Computation: Practice and Experience, 2007