CRAN data-ecosystem package velocity (agent-curated)
Monthly download-velocity ranking of CRAN packages in the data/ML ecosystem. 32 data-relevant CRAN Task Views (MachineLearning, TimeSeries, Spatial, Bayesian, Econometrics, ...) — the human-curated per-topic package lists — are parsed into a candidate universe (~3.8k packages; the ``ctv`` template link is excluded). The keyless cranlogs API supplies trailing-30-day download totals in batched multi-package calls; candidates clearing the 10,000-download floor also get the prior-30-day total from the same API, plus title/description/license from the keyless crandb API (author identities are not stored — Authors@R carries email addresses). Each surviving package is classified into a curated 11-category taxonomy by a deterministic ordered rule set over its task-view membership. The velocity score is the min-max-normalized trailing-30d download count (0-100) and the growth score the min-max-normalized month-over-month growth rate (clipped to -100%/+1000%; new entrants pinned to the +1000% cap and flagged); trend_score is their 50/50 composite (0-100) and trend_rank orders by trend_score descending (ties: trailing-30d downloads desc, then package name). Columns: month, fetch timestamp, package name / URL / version / title / description / license id / published date, task views / category, exact source API URL for each row value (task-view pages, trailing-30d totals, prior-30d totals, crandb metadata — any number is one GET away from its source), trailing-30d and prior-30d downloads, 60d total and per-day average, growth rate, new-entrant flag, velocity/growth/trend scores (0-100), trend rank. fetched_at is the trailing window end (UTC day granularity), so re-ingesting the same window is a no-op. Primary key: (month, package_name). Cadence: monthly. Nullability: license_id and published_date may be empty when upstream supplies none; scores and ranks are never null. Caveats: cranlogs counts RStudio-mirror downloads (automated and CI traffic included, not just human adoption), so raw velocity favors widely-installed infrastructure; the taxonomy is a first-match heuristic over task-view membership, not an authoritative classification; per-package license ids are author-supplied and informational only. Sample use: order by trend_rank for the month's fastest-rising data packages, or filter category = 'machine-learning'.
- technology
- signals
- monthly
- cran