Optionally use download_queue for brew install

[MikeMcQuaid]

Allowing using HOMEBREW_DOWNLOAD_CONCURRENCY to use the DownloadQueue for brew install by downloading and extracting bottles in parallel.

This requires some fixes in e.g. Dependency and FormulaInstaller to be able to front-load all downloads and handle parallelisation of bottle pouring.

Behaviour without HOMEBREW_DOWNLOAD_CONCURRENCY set should be unchanged.

Attestations are not handled for now and the UI should be improved before we roll this out to users.

Post-install upgrades are not yet parallelised.

Performance testing on my M1 Max MacBook Pro (that we do make -j10 on):

Benchmark 1: HOMEBREW_DOWNLOAD_CONCURRENCY=40 brew install ack wget ffmpeg
  Time (mean ± σ):     107.526 s ±  3.498 s    [User: 31.709 s, System: 65.326 s]
  Range (min … max):   105.053 s … 110.000 s    2 runs
 
Benchmark 2: HOMEBREW_DOWNLOAD_CONCURRENCY=20 brew install ack wget ffmpeg
  Time (mean ± σ):     100.833 s ±  3.007 s    [User: 31.519 s, System: 62.292 s]
  Range (min … max):   98.706 s … 102.959 s    2 runs
 
Benchmark 3: HOMEBREW_DOWNLOAD_CONCURRENCY=10 brew install ack wget ffmpeg
  Time (mean ± σ):     103.043 s ±  1.297 s    [User: 31.360 s, System: 63.587 s]
  Range (min … max):   102.126 s … 103.960 s    2 runs
 
Benchmark 4: HOMEBREW_DOWNLOAD_CONCURRENCY=2 brew install ack wget ffmpeg
  Time (mean ± σ):     139.044 s ±  3.662 s    [User: 31.711 s, System: 64.411 s]
  Range (min … max):   136.454 s … 141.633 s    2 runs
 
Benchmark 5: brew install ack wget ffmpeg
  Time (mean ± σ):     162.942 s ±  0.163 s    [User: 32.008 s, System: 57.277 s]
  Range (min … max):   162.827 s … 163.057 s    2 runs
 
Summary
  HOMEBREW_DOWNLOAD_CONCURRENCY=20 brew install ack wget ffmpeg ran
    1.02 ± 0.03 times faster than HOMEBREW_DOWNLOAD_CONCURRENCY=10 brew install ack wget ffmpeg
    1.07 ± 0.05 times faster than HOMEBREW_DOWNLOAD_CONCURRENCY=40 brew install ack wget ffmpeg
    1.38 ± 0.05 times faster than HOMEBREW_DOWNLOAD_CONCURRENCY=2 brew install ack wget ffmpeg
    1.62 ± 0.05 times faster than brew install ack wget ffmpeg

[carlocab]

Can't say I understand all of this, but it looks reasonable to me. Minor nits below.

[Bo98]

Need to incorporate the retry logic in

brew/Library/Homebrew/bottle.rb

Lines 123 to 132 in e804860

	rescue DownloadError
	raise unless fallback_on_error?
	retry
	rescue Resource::BottleManifest::Error
	raise if @fetch_tab_retried
	@fetch_tab_retried = true
	resource.clear_cache
	retry

since we're not calling fetch_bottle_tab anymore. Perhaps by some custom callbacks from the queue or something

The final # Do nothing rescue in the installer is also important for CI when we enable it there.

[MikeMcQuaid]

@Bo98 I think this should do it? #20285

[Bo98]

I don't see where that PR changes this

[MikeMcQuaid]

@Bo98 This is triggering a manual retry on bottle manifest download errors. That PR is automatically triggering a manual retry on any bottle manifest errors under the download queue.

Do you mean this logic needs reincorporated for the non-download queue case and that's missing/changed and still needs resolved beyond #20282?

[Bo98]

Oh right I see what you mean - I misunderstood there.

We do however also need a retry (with a clear_cache) if it gets a Resource::BottleManifest::Error triggered by verify_download_integrity of the bottle manifest:

brew/Library/Homebrew/resource.rb

Lines 316 to 319 in 6261551

	def verify_download_integrity(_filename)
	# We don't have a checksum, but we can at least try parsing it.
	tab
	end

Try modify a locally-cached manifest to e.g. remove all arm64 bottles and see if that works as a reproducer. (I've not retested this after the changes yet.)

[MikeMcQuaid]

@Bo98 gotcha, thanks, on it (and holding release on it)

[MikeMcQuaid]

@Bo98 Not blocking release but working on a fix for this as it only affects download queue case.

[MikeMcQuaid]

#20288