[earakely-scale]

What changes

Fixes three flaky tests: two integration tests, one in each tier, whose failures came from image pulls, and a unit test that failed on the CodeBuild run.

run_bundle_test's killed-run test (fast tier). Every integration-local failure in the last 100 failed runs was this test (runs 43313173134, 43369378729, 43472071259 on main, 43538837017), each with public.ecr.aws/nginx/nginx:stable-alpine … toomanyrequests: Rate exceeded. The test's container sandbox ran that image. A local sandbox pulls a plain tag every time, so each run made an anonymous ECR Public pull. The test now runs the registry image the job already holds (mirror.gcr.io/library/registry:2, the :5000 service's), under its digest-pinned name (RepoDigests). A local sandbox doesn't pull an image it holds under a digest-pinned name. A machine without that image pulls it once, from mirror.gcr.io.

A compose's pulls of its base take turns (slow tier, product fix). Run 43505816495 (integration-local-slow (5/12), 2026-11-13) failed in browser_gateway_test's compose of the chromium env. The amd64 resolve failed with failed to copy: … could not fetch content descriptor … not found. The arm64 platform check was running at the same moment. clock_desktop_test on the same runner then failed 30 s later with failed to get reader from content store: content digest …: not found. composer.compose builds pairs in threads, and both of these ran a container of the same base at once:

• the resolve: docker create --platform linux/amd64 <base>
• builder._runs_here's emulation check: docker run --platform linux/arm64 <base>

Either one pulls the base when Docker lacks it, and neither held the image-pull: lock that host_images.pull holds. On the containerd image store, the two pulls left the base's amd64 manifest without all its layers, so every later container of it on that runner failed. Both now hold that lock (host_images.pull_lock), so pulls of one image take turns across threads and processes. The slow tier's MACHINE_LOCKS already include image-pull:.

user_chat_test's get_time test (unit). It set the gateway's clock with no rate, so it ran at 1, then expected get_time to answer the second it set. On #690's CodeBuild run it answered 15:00:01Z for 15:00:00Z. It now sets the clock still, as the file's other tests do.

Not covered: when a stored package lock lets a compose skip the resolve, BuildKit's own pull of the base in docker build doesn't take the lock. No CI failure has come from that path. An image already half-pulled by an earlier interrupted pull isn't repaired; this only stops our own concurrent pulls from making one.

Testing

• New unit test test_a_resolves_pull_of_its_base_takes_turns_with_a_platform_checks_and_a_deploys: a resolve, an arm64 platform check and a deploy's host_images.pull of one base at once, against a fake Docker. At most one pull is in flight. It fails without the fix, and passed 6 times out of 6 with it.
• tst/unit/env/envs, tst/unit/providers and tst/unit/bundle: 4398 passed. The one failure there, test_disk_recipe's Windows install, fails the same way on main on macOS (socat UNIX-CONNECT). ruff check is clean.
• On an x86 host with Docker's containerd image store (io.containerd.snapshotter.v1):
  • the killed-run test passed with no image pulled during it; its container ran from mirror.gcr.io/library/registry@sha256:a3d8aaa6…
  • browser_gateway_test.py and clock_desktop_test.py together with -n 2 --dist loadfile, as the slow shard runs them: 5 passed, twice (274 s and 116 s). The one skip is the Playwright/Puppeteer test's, as on main. Neither run's log shows a resolve, so these runs didn't exercise the race. The unit test above is what shows the lock holds.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; no new actionable issues were found.

Summary

The PR makes local pulls of the same base image take turns and removes two test timing dependencies.

• Local computer builds make pulls of the same base image take turns.
• The killed-run test starts its sandbox from the local registry image.
• The get_time test keeps its virtual clock fixed.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  R[Resolve creates container] --> L[Shared image-pull lock]
  C[Platform check runs container] --> L
  D[Deploy pulls image] --> L
  L --> P[One pull at a time]
  R --> T[Lock wait counts toward resolve timeout]

Loading

_{Reviews (7) · Last reviewed commit: "test(gateway): get_time's test sets the ..." · Reviewed by Greptile}