test(gateway): use deterministic CAR fixtures

[laurentsenta]

Contributes to #9654

• t0109-gateway-web-_redirects.sh
  • no change required
• t0111-gateway-writeable.sh
  • no change: we generate random data here. writable gateways will be covered later.
• t0112-gateway-cors.sh
  • no change required
• t0113-gateway-symlink.sh
• t0114-gateway-subdomains.sh
  • ipfs, no change for ipns yet
  • ipns (not doing ipnfs yet)
• t0115-gateway-dir-listing.sh
• t0116-gateway-cache.sh
• t0117-gateway-block.sh
• t0118-gateway-car.sh
• t0122-gateway-tar.sh
• t0123-gateway-json-cbor.sh
• t0124-gateway-ipns-record.sh (not doing ipns yet)
• t0400-api-no-gateway.sh

[hacdias]

Overall looks good to me. This is very nice 😃 I left some inline comments.

[laurentsenta]

Thanks for the review @hacdias, I patched your comments,
and I removed the IPNS change for now (it started breaking for no apparent reason). I'll add this in a second PR.

go test is failing at the moment but since we haven't touched code other than sharness shell scripts, I guess this is unrelated flakiness.

[hacdias]

I re-run the go test in GitHub and it's fine now. There are a few flaky tests and they seem to fail more often in GitHub Actions than Circle CI actually.

[lidel]

Thank you @laurentsenta! ❤️

This work is not only unblocking work on gateway conformance test suite that is not tied to Kubo, but ALSO reduces the surface work for switching ipfs add to produce CIDv1 by default.

[galargh]

I re-run the go test in GitHub and it's fine now. There are a few flaky tests and they seem to fail more often in GitHub Actions than Circle CI actually.

I checked success ratios of go-test in the last 24 hours. It's 29% for CircleCI and 29.4% for GitHub Actions. CircleCI can't do fractions so it seems to be the same thing. I didn't go further than 24 hours because we only started collecting GitHub Actions data 3 days ago and CircleCI only offers insights in 1, 7, 30, 60, 90 day intervals.

interop is the only test that seems to be doing worse on GHA. 92% vs 73.3%. Luckily, it's only on the surface. CircleCI cheats here a bit because it doesn't count cases where build failed and interop couldn't even start; like this one https://github.com/ipfs/kubo/pull/9677/checks?check_run_id=11653777666.

Overall, it seems we're pretty much on track to be able to disable CircleCI. I'll dig into the data more before the final cleanup PR but there's nothing worrying standing out.

Here's the WIP GHA monitoring dashboard if you want a peek: https://protocollabs.grafana.net/d/gieYOhbVk/github-actions-ipfs-kubo?orgId=1&from=now-3d&to=now (cc @BigLep since I think you might be interested ;)

[hacdias]

@galargh interesting. My "it seems" is just based on my observations on my PRs specifically, so I might just be unlucky :P Thanks for the link!