perf(npm): optimize loading npm resolution snapshot from lockfile by only loading necessary version info

[nathanwhit]

This is sort of a POC, the actual code is sus but works well enough to see the perf impact on real world projects.

Basically when we're loading the npm resolution snapshot from a lockfile, we read (and deserialize!) a bunch of registry.json files from the cache dir.

The thing is, those registry.json files can be pretty massive (e.g. nextJS's registry.json is ~6MB), and reading and deserializing them is very expensive.

We don't really need the entire NpmPackageInfo though, we just want the information for a specific version (the one in the lockfile).

This PR optimizes this by creating an index of the registry.json files that gives the file offsets containing the version info for each version. Then, when we load the snapshot we read the index to get the byte offsets for the version we need, and only load (and deserialize) that part of the registry.json file.

---

Results (this is in a project from create-next-app)

Running an empty file with --node-modules-dir=auto (which eagerly loads the resolution snapshot):

❯ hyperfine --warmup 20 'deno run --node-modules-dir=auto -A empty.js' '/Users/nathanwhit/Documents/Code/deno/target/release/deno run --node-modules-dir=auto -A empty.js'
Benchmark 1: deno run --node-modules-dir=auto -A empty.js
  Time (mean ± σ):     213.8 ms ±   2.7 ms    [User: 197.2 ms, System: 17.0 ms]
  Range (min … max):   210.0 ms … 218.4 ms    13 runs

Benchmark 2: /Users/nathanwhit/Documents/Code/deno/target/release/deno run --node-modules-dir=auto -A empty.js
  Time (mean ± σ):      19.8 ms ±   0.6 ms    [User: 13.2 ms, System: 7.8 ms]
  Range (min … max):    18.8 ms …  21.9 ms    138 runs

Summary
  /Users/nathanwhit/Documents/Code/deno/target/release/deno run --node-modules-dir=auto -A empty.js ran
   10.82 ± 0.33 times faster than deno run --node-modules-dir=auto -A empty.js

Deno install with a populated cache dir, but setting up node modules from scratch:

❯ hyperfine --warmup 2 --shell=none --prepare "rm -rf node_modules" "deno i" "/Users/nathanwhit/Documents/Code/deno/target/release/deno i"
Benchmark 1: deno i
  Time (mean ± σ):     395.5 ms ±   6.6 ms    [User: 194.7 ms, System: 200.8 ms]
  Range (min … max):   388.0 ms … 409.8 ms    10 runs

Benchmark 2: /Users/nathanwhit/Documents/Code/deno/target/release/deno i
  Time (mean ± σ):     187.1 ms ±  13.5 ms    [User: 10.0 ms, System: 196.6 ms]
  Range (min … max):   167.9 ms … 211.6 ms    10 runs

Summary
  /Users/nathanwhit/Documents/Code/deno/target/release/deno i ran
    2.11 ± 0.16 times faster than deno i

A no-op deno install (node_modules already set up):

❯ hyperfine --warmup 20 --shell=none "deno i" "/Users/nathanwhit/Documents/Code/deno/target/release/deno i"
Benchmark 1: deno i
  Time (mean ± σ):     212.2 ms ±   4.8 ms    [User: 194.8 ms, System: 15.9 ms]
  Range (min … max):   206.5 ms … 225.2 ms    14 runs

Benchmark 2: /Users/nathanwhit/Documents/Code/deno/target/release/deno i
  Time (mean ± σ):      11.5 ms ±   0.3 ms    [User: 6.2 ms, System: 5.0 ms]
  Range (min … max):    10.8 ms …  12.5 ms    255 runs

Summary
  /Users/nathanwhit/Documents/Code/deno/target/release/deno i ran
   18.53 ± 0.67 times faster than deno i

[lucacasonato]

Wow, this is super great