Feature detection caches features as not available if a builder is polled before it is fully initialized

Contributing guidelines

• I've read the contributing guidelines and wholeheartedly agree

I've found a bug and checked that ...

• ... the documentation does not mention anything about my problem
• ... there are no open or closed issues that are related to my problem

Description

When running docker buildx build --platform linux/amd64 --platform linux/arm64 --type=docker . against a docker:dind container, I sometimes get:

ERROR: failed to build: docker exporter does not currently support exporting manifest lists

But this happens intermittently, and only under load, which seems to imply that the error message is incorrect. Maybe the builder isn't fully initialized or something, but it's returning an error message implying that the driver is incompatible. This can't be given that most of the time the same test passes just fine and I am using the default builder driver every time.

Per @thaJeztah:

Looks like it does feature detection here;
https://github.com/docker/buildx/blob/1d7cda1232b38146345c8d5d73d6ee88de03eda1/build/opt.go#L224-L233
Looking at that code, it uses a sync.Once, and discards errors (if ... err == nil), so if it has a failure when trying to detect features the first time, it will discard the error, and considers it “not supported” for the duration of the buildx instance ;
https://github.com/docker/buildx/blob/461bd9e5d134057c42a099977f27af9bd47a1080/util/dockerutil/client.go#L76-L93

If the only way to check whether a builder is ready for use is a retry-loop then this will result in the builder's feature getting stripped away for the lifetime of the builder which is bad.

Expected behaviour

The result of features detection should only be cached if the builder is fully initialized

Actual behaviour

If an error occurs while communicating with a builder, a feature is marked as "not available" forever

Buildx version

github.com/docker/buildx v0.25.0 faaea65da4ba0e58a13cd9cadcb950c51cf3b3c9

Docker info

Builders list

NAME/NODE           DRIVER/ENDPOINT     STATUS    BUILDKIT   PLATFORMS
default             docker
 \_ default          \_ default         running   v0.22.0    linux/amd64 (+3), linux/arm64, linux/arm (+2), linux/ppc64le, (2 more)
desktop-linux*      docker
 \_ desktop-linux    \_ desktop-linux   running   v0.22.0    linux/amd64 (+3), linux/arm64, linux/arm (+2), linux/ppc64le, (2 more)

Configuration

N/A

Build logs

Additional info

No response

Start by reading the feature-detection logic in build/opt.go around the linked lines and the builder client handling in util/dockerutil/client.go. Reproduce the docker buildx build command under load, then verify that a transient initialization error no longer permanently marks the feature unavailable and that normal feature detection still works.