Skip to content

RESEARCH_NOTES.md

Latest commit

 

History

History
60 lines (46 loc) · 6.3 KB

RESEARCH_NOTES.md

File metadata and controls

60 lines (46 loc) · 6.3 KB
Raw

Research Notes — sources, methodology rationale, and the verified timeline

Consolidated from primary-source research. These notes justify the choices hard-coded in src/.

1 · Data sources

npm download counts (backbone)

  • API: https://api.npmjs.org/downloads/range/{start:end}/{package} — public, no auth, daily granularity.
  • Earliest data: 2015-01-10. Single-package ranges silently truncate beyond ~540 days → we chunk into 500-day windows and stitch (src/npm_fetch.py).
  • Scoped packages (@angular/core) must be percent-encoded and cannot appear in bulk calls.
  • Package identity matters more than anything: @angular/core = modern Angular; unscoped angular = AngularJS 1.x, a different EOL product (kept as a separate legacy line, never merged). Track one heartbeat package per framework (@angular/core, react, vue, svelte) — never sum @angular/*, which double-counts dependency fan-out.
  • A download is a CI/mirror/cache event far more often than a human; npm filters no bots by design. → trust only normalized signals.
  • Spec: https://github.com/npm/registry/blob/main/docs/download-counts.md. Methodology: https://blog.isquaredsoftware.com/2022/07/npm-package-market-share-estimates/.

Ecosystem denominator

No public "total registry downloads" endpoint exists. We proxy generic npm/CI growth with a basket of ubiquitous, framework-agnostic packages (lodash, chalk, commander, debug, express, typescript, eslint, axios) for the per-capita normalization.

GitHub (src/github_fetch.py)

  • REST gives current stargazers_count cheaply; star history requires paginating /stargazers with Accept: application/vnd.github.star+json. Pagination is capped at 400 pages (40k stars) → for big repos we sample sparse pages and interpolate, like star-history.com.
  • Unauthenticated limit 60 req/hr → low-res sampling; set GITHUB_TOKEN for 5000/hr.
  • Repo splits: modern angular/angular vs archived angular/angular.js; Vue 3 vuejs/core vs Vue 2 vuejs/vue (where most historical stars live). We track the modern repo and flag the split.

Surveys & benchmarks (src/curated_data.py, cited)

  • State of JS (retention/sentiment) — GraphQL at api.devographics.com/graphql; charts not text-extractable so values transcribed.
  • Stack Overflow Developer Survey — per-respondent CSVs at survey.stackoverflow.co (2023–25 include AI-tool × framework cross-tabs). Angular and AngularJS are separate options.
  • Web-Bench (ByteDance, arXiv:2505.07473) — per-framework LLM pass-rate, the hard evidence that AI codegen favors React.
  • Stack Overflow question volume collapsed ~70% post-ChatGPT → use % share, not raw counts.

AI-era signals

  • Vercel AI SDK adapter downloads (@ai-sdk/react vs @ai-sdk/angular, etc.) — sharpest single skew marker, re-fetched live from npm. @ai-sdk/angular only shipped 2025-07-14.

2 · Verified Angular intervention timeline (the key correction)

Twice-a-year cadence; no Angular 3 (skipped). Dev-preview ≠ stable — we use the true stabilization dates:

Date Ver Milestone Status
2022-06 v14 Standalone components dev preview
2022-11 v15 Standalone APIs stable
2023-05 v16 Signals, hydration, esbuild dev preview
2023-11 v17 Control flow, standalone+esbuild default, signals core stable, rebrand/angular.dev mixed
2024-05 v18 Control flow/defer stable, experimental zoneless mixed
2024-11 v19 Signal inputs/outputs/model stable stable
2025-05 v20 effect/linkedSignal/toSignal stable stable
2025-11 v21 Zoneless default, Vitest, MCP server stable

Biggest correction vs. the popular narrative: signals were not wholesale "stable in v17." signal/computed stabilized v17 (Nov 2023); effect()/toSignal only v19–v20; zoneless only v20.2 (Aug 2025), default v21 (Nov 2025). The densest architectural-inflection window is v16–v17 (May–Nov 2023) — which collides with the AI inflection, the central analytical challenge.

Competitor markers used for alignment: React Hooks (2019-02), Vue 3 (2020-09), React 18 (2022-03), SvelteKit 1.0 (2022-12), Next 13.4 RSC GA (2023-05), Svelte 5 runes (2024-10), React 19 (2024-12).

3 · Methodology rationale

  • Preprocess: weekly (kills weekday CI cycle) → log1p (growth = slope, variance-stabilized) → STL deseasonalize (December dip) → robust MAD spike handling (release-week mirror stampedes). Fit on deseasonalized unsmoothed series; rolling means are visualization-only.
  • Normalize three ways: logit share-of-cohort (the "framework war"), per-capita vs ubiquity basket (ecosystem detrend), indexed-to-baseline (trajectory comparison).
  • Causal stack:
    • Confirmatory ITS / segmented regression at known dates, in a local window (the deep pre-2016 history would otherwise distort the trend), with Newey–West (HAC) errors because downloads are heavily autocorrelated. Meaningful term = trend_change, not the publish-day level.
    • Backbone BSTS / CausalImpact: Angular ~ local-level + regression on a single composite control = the summed downloads of the other three frameworks ("rest of cohort"). The three are highly collinear, so on raw multi-control levels the betas blow up and the counterfactual spikes at year-end; one composite gives a stable coefficient that tracks the cohort's seasonality directly. Run on log levels with a 6-week washout. Implemented on statsmodels UnobservedComponents because the causalimpact PyPI package is broken on pandas 3.0. Validity gate: a pre-fit correlation (fitted vs actual on the pre-period) — high ⇒ trust the counterfactual; low ⇒ discard it, and a low pre-fit is itself a finding (Angular decoupled from the cohort). Plus an in-time placebo.
    • Exploratory ruptures PELT over a penalty sweep — "where did breaks really land?"
  • AI era modeled as a gradual ramp (0→1 over ~2.5y from ChatGPT), as a common shock across all four. Its absolute effect is unidentifiable (no AI-free control); only differential cross-framework responses are.
  • Honesty: downloads ≠ adoption ≠ developers; releases are observational; bot/mirror/CI traffic is unfiltered; multiple-comparison risk across releases × normalizations × methods.