vim-ale/COMPARISON.md

Go vs Node.js Cache Comparison

Systematic comparison of Go pipeline output (_cache/) vs Node.js production cache (LIVE_cache/). Generated by cmd/comparecache.

Latest run: 2026-03-10. -no-fetch rebuild from existing raw data, with variant tagging and legacy export filter applied.

Summary (latest version only)

Category	Count	Meaning
match	21	Identical asset filenames at latest version
go-missing	4	Go produces no output (alias, meta-package, no config)
live-missing	16	Package exists in Go but not in live cache
go-extra-versions	44	Go has more version history (deeper fetch)
live-extra-versions	11	Live has newer data (rate-limited Go fetch)
go-extra-assets	19	Go includes assets that Node.js filters out
live-extra-assets	41	Node.js includes assets that Go filters out
live-has-meta	41	Node.js includes meta-assets (checksums, sigs)

Changes since last comparison: variant tagging (bun, pwsh, ollama, git, node, lsd, fish, xcaddy) and legacy export filter (strips Variants-tagged assets and non-legacy formats like .deb, .rpm, .tar.bz2, .tar.zst, .7z) are now active. Match count dropped from 28→21 because .deb files that Node.js keeps are now correctly categorized as live-extra-assets (Go filters them from legacy output).

Key Observations

1. Classification Timing

The Node.js cache stores assets with empty os/arch/ext fields — normalize.js fills those at serve time. The Go pipeline classifies at write time. The Go cache has richer data per-asset. Comparison is done at the filename level.

2. Meta-Asset Filtering

Go's isMetaAsset() filters out checksums, signatures, SBOMs, etc. Node.js keeps them. This accounts for 41 packages showing live-has-meta differences. Correct behavior — Go filters non-installable files at cache time.

3. Version Depth

Go has deeper version history for most GitHub-sourced packages (fetches all pages unless -shallow). Node.js limits to 30 releases per API call. This is a feature — Go provides complete histories when doing a full fetch.

4. Build Variants (IMPLEMENTED)

Variant tagging is now active. Per-package taggers in internal/releases/{pkg}/. Assets with Variants are stored but excluded from legacy export.

• bun: -profile → Variants: ["profile"]; non-baseline → Arch: amd64v3
• ollama: -rocm, -jetpack5, -jetpack6 → Variants
• pwsh: -fxdependent, -fxdependentWinDesktop → Variants
• git: .exe and PortableGit → Variants: ["installer"]; -pdb → ["pdb"]
• node: .msi → Variants: ["installer"]
• lsd: .deb → ["deb"]; -msvc → ["msvc"]
• fish: .pkg → Variants: ["installer"]
• xcaddy: .deb → Variants: ["deb"]

5. Legacy Export Filter (IMPLEMENTED)

ExportLegacy now strips:

• Assets with non-empty Variants
• Formats Node.js doesn't handle: .deb, .rpm, .snap, .appx, .tar.zst, .tar.bz2, .7z

This means the _cache/ JSON files only contain assets the Node.js server can actually serve.

6. Format Handling

All formats are stored in the internal Go model — nothing is dropped at classification time. The legacy filter applies only at export time.

• .pkg — pkgutil --expand-full
• .deb — ar x + tar xf data.tar.*
• .dmg — hdiutil attach
• .msi — msiexec /a

Only .exe is ambiguous (binary vs installer). Installer .exe files get Variants: ["installer"].

7. Node Multi-Source

The node package merges official + unofficial builds via unofficial_url in releases.conf. Down to 4 differences at latest version:

• Live has .7z (filtered from Go legacy export) and .msi (Go tags as installer)
• Go has .exe bare binary that live doesn't (naming diff)

Per-Package Checklist

Status: [x] reviewed, [-] known acceptable, [ ] needs work

Exact Matches at Latest Version (21)

• atomicparsley
• awless
• chromedriver (chromedist)
• comrak
• dotenv-linter
• gpg (gpgdist)
• iterm2 (iterm2dist)
• julia (juliadist)
• koji
• lsd — .deb and msvc variants now correctly filtered
• mariadb (mariadbdist)
• pathman
• sass
• sd
• shellcheck
• shfmt
• sqlc
• terraform (hashicorp)
• xcaddy — .deb variants now correctly filtered
• xsv
• zig (zigdist)

Go Missing (4)

• [-] dashd — alias_of=dashcore (correct)
• [-] macos — no releases.conf
• [-] pg-essentials — meta-package
• [-] zig.vim — gittag source, 0 raw data

Live Missing — Go-Only (16)

• [-] node-official — Go split, not in live cache
• [-] node-unofficial — Go split, not in live cache
• [-] pg — Go alias, live uses postgres
• [-] ripgrep — Go alias, live uses rg
• [-] rust.vim — symlink to vim-rust
• [-] vim-* (11 packages) — gittag packages not in live cache

Meta-Only Diffs (Go correctly filters, Node.js keeps)

• [-] caddy, cilium, cmake, curlie, dashmsg, deno, dotenv, ffuf, fzf, gh, gitdeploy, goreleaser, gprox, grype, hugo, k9s, keypairs, kind, kubectx, kubens, monorel, mutagen, ots, rclone, rg, runzip, sclient, sqlpkg, sttr, syncthing, terramate, watchexec, xz, yq (41 packages)

Live has .deb/.rpm that Go correctly filters from legacy export

• [-] bat — 8 .deb files
• [-] caddy — 9 .deb files
• [-] delta — 5 .deb files
• [-] fd — 8 .deb files
• [-] gh — 8 .deb/.rpm files
• [-] goreleaser — 18 .deb/.rpm files
• [-] grype — 8 .deb/.rpm files
• [-] hexyl — 7 .deb files
• [-] k9s — 10 .deb/.rpm files
• [-] pandoc — 2 .deb files
• [-] pwsh — 4 .deb/.rpm files
• [-] rclone — 16 .deb/.rpm files
• [-] sttr — 9 .deb/.rpm files
• [-] syncthing — 5 .deb/.rpm files
• [-] terramate — 6 .deb/.rpm files
• [-] tinygo — 3 .deb files
• [-] trip — 3 .deb files
• [-] watchexec — 16 .deb files
• [-] zoxide — 4 .deb files

Remaining Go-Extra-Assets (need review)

• bun — baseline builds now serve as legacy amd64 (filename stripped, download URL kept); non-baseline tagged as v3 variant (excluded).
• fish — source tarball tagged as variant; linux .tar.xz binaries are correct (Node.js just doesn't have them yet)
• git — 4 extras: MinGit-busybox .zip, pdbs .zip at latest version
• [-] hugo — 1 extra: Linux-64bit.tar.gz (old naming); keep as-is for now
• [-] hugo-extended — 14 extras: non-extended assets leaking in; keep as-is for now
• kubectx — 14 extras: kubens assets from shared GitHub release
• kubens — 14 extras: kubectx assets from shared release
• node — 1 extra: .exe bare binary naming difference
• ollama — 2 extras: Ollama-darwin.zip (case difference?)
• uuidv7 — 16 extras: exotic arches (thumbeb, armeb, riscv32)
• yq — 1 extra: naming difference
• ffmpeg — 21 extras: many platform/format combinations

Source/Naming Diffs

• [-] aliasman — source tarball naming differences (GitHub archive format)
• [-] duckdns.sh — source tarball naming differences
• [-] serviceman — source naming + version differences

Stale Data (rate-limited, need re-fetch with token)

• [-] go — live has 98 extra versions (Go didn't fetch golang.org)
• [-] lf — live has 30 extra versions
• [-] postgres, psql — Go has v17, live has v18
• [-] ffmpeg — Go has older, live has newer

Cross-Package Issues

• kubectx/kubens — shared GitHub release, assets for both packages appear in each. Need to split by asset name prefix.

Remaining Action Items

1. hugo-extended exclude: Deferred — keep matching Node.js behavior for now
2. kubectx/kubens split: Filter assets by name prefix in shared release
3. bun baseline in legacy: Resolved — baseline is legacy amd64, non-baseline tagged as v3 variant
4. Re-fetch with GITHUB_TOKEN: Fix rate-limited/stale packages
5. Unknown asset notifications: Log new/unrecognized assets to _notices/

Deferred Decisions

1. Consolidate cmd/classify and cmd/webicached duplication: Both have their own classifyPackage switch, isMetaAsset, detectFormat, and GitHub API types (ghRelease, ghAsset, etc.). cmd/classify is a diagnostic tool (CSV output), cmd/webicached is the production pipeline ([]storage.Asset). Shared pieces could move to internal/ packages. Keep separate dispatchers since they return different types.