Adding the CodeQL Bundle to the toolcache seems to take a surprisingly long amount of time. We haven't collected large scale telemetry yet, but here's some anecdotal numbers from looking at the logs on PR checks:

Linux: ~10–15 secs
macOS: ~15–20 secs
Windows: ~2 minutes (!)

Therefore in this PR, we add a feature flag that extracts the CodeQL bundle directly to the toolcache, eliminating the copy to toolcache step completely.

I've tested the change by adding a PR check, temporarily modifying this to running 10 runs on each platform, and running this a bunch of times. Next steps are dogfooding this change internally using the feature flag.

Merge / deployment checklist

Confirm this change is backwards compatible with existing workflows.
Confirm thereadme has been updated if necessary.
Confirm thechangelog has been updated if necessary.

henrymercer added12 commits

December 2, 2024 17:58

Enable specifying extraction destination

be26fe6

Add feature flag for extracting directly to toolcache

6695503

Extract directly to toolcache when feature enabled

198fb4b

Add PR check

028cece

Write toolcache marker file

0b20485

PR check: Ensure bundle added to toolcache

8c0517f

PR check: Check toolcache empty at start

f8ccf38

Run PR check on Windows too

d57328a

Add more error handling to zstd extraction

d5bcf48

Ensure destination directory exists when streaming extraction

b2a4442

Tweak log wording

93f3303

Record telemetry for cache operation duration

7bcf845

henrymercer requested a review froma team as acode owner

December 3, 2024 18:29

dbartol previously approved these changes

Dec 3, 2024

View reviewed changes

Copy link

dbartol left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

LGTM. Looks like a relatively simple change, plus some plumbing to pass the destination directory through to all the places the write the files.

On a GitHub-hosted runner, it looks like we'll still add the bundle to the toolcache, even though the cache will disappear after the job. Is this just because we have to put itsomewhere, and we might as well put it in the toolcache for consistency with the less-common self-hosted runner scenario?

Copy link

ContributorAuthor

henrymercer commentedDec 4, 2024

On a GitHub-hosted runner, it looks like we'll still add the bundle to the toolcache, even though the cache will disappear after the job. Is this just because we have to put itsomewhere, and we might as well put it in the toolcache for consistency with the less-common self-hosted runner scenario?

Yes, pretty much. It's also consistent with the situation where the GitHub-hosted runner images have caught up to the latest version of CodeQL and the tools we want are already installed. Plus if anyone's still using a hardcoded toolcache path to find CodeQL, that'll still work.