Does Popular Open-Source Software Stop Building with Age? A Pre-Registered Rebuild of Top GitHub Repositories Does Not Support a Decline in Every Ecosystem
We rebuilt the most-starred GitHub repositories created in each year 2013-2024, in four ecosystems (Python, Node, Go, Rust), from two snapshots each: the last commit at the end of the year after creation (ERA) and the current head (NOW), with one fixed recipe per ecosystem, today's toolchains and one attempt. This report covers the pre-registered CORE tier only (up to 75 repositories per ecosystem and year); the larger FULL tier has not started and is not analysed here. The analysed data hold 7200 build attempts over 3600 repositories. During CORE a defect in the build farm's allowlist proxy (a descriptor limit in its use of select and no timeout on request headers) made builds fail at the farm's own git fetch; it was found from the progress monitor, reproduced and fixed in recipe revision 2f; two further proxy defects (a log file reopened per connection, fixed in rev 2g, and a single-process throughput ceiling, changed in rev 2h) were found the same way, and the rev 2h proxy then showed fetch-step faults of its own (A14). Receipts from the affected windows were voided, whatever their outcome, and the rows rebuilt; the void sets are published (639 void receipts in five parts by the amendments; 639 counted in the published void parts). The first diagnosis of the fault was wrong and was corrected before any analysis (amendments A6 to A15). On a random 100-row sample of the earlier receipts, rebuilt under the fixed proxy, 100 of 100 rows gave the same outcome; per-revision samples (A15) gave 60 of 60 (rev 2g), 60 of 60 (rev 2i) and 55 of 60 (rev 2j). Because farm revision changed during CORE, two sensitivity fits were pre-specified (A15): no hypothesis verdict changed under either sensitivity fit. The pre-registered hypothesis that ERA build success decreases with age in every ecosystem (H1) is not supported: it holds in 2 of 4 ecosystems, and in the other two the Cochran-Armitage trend test does not reject at alpha 0.0125. Per-ecosystem H1 results: Python NOT SUPPORTED, Node NOT SUPPORTED, Go SUPPORTED, Rust SUPPORTED. Fitted exponential half-lives of ERA build success: Python 6.49 years (95% CI 5.84 to 7.14 years), Node 5.91 years (5.34 to 6.50 years), Go 6.47 years (5.92 to 7.09 years), Rust 4.33 years (3.95 to 4.72 years). The pre-registered test that ecosystems differ in half-life (H2) is SUPPORTED. The exponential model assumes steady decay; Python and Node show none, so their T is a fitted summary, not a measured decay rate. H3 (secondary) meets its pre-registered test as computed (p = 0.045), but one marker row is a misclassified human co-author; corrected, p = 0.056, and the association also does not survive post-hoc stratification by ecosystem (sections 6.2, 6.7). Secondary results: the end-of-life comparison (A22/A23) prints "verdict_NOW_rows=DOES NOT TRACK" and "A22_ERA_rule=UNDERPOWERED"; among DEP-MISSING NOW builds, 0 had a first named artefact that is now absent and is known to have existed (A18). Every build has a public receipt; each ecosystem-year cell of receipts and the whole dataset are signed (DSSE).
Authors
- Advatiya Thalluri (ORCID: https://orcid.org/0009-0004-3435-9772)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24267298
- Primary Topic
- Software Engineering Research
- Type
- preprint