acoustic-carpenter-78188
05/21/2026, 10:41 PMRemoteFileOutputReader.ReadError already populates the Go-level io.ExecutionError.IsRecoverable bool from ContainerError.Kind, but it does not populate the proto-level core.ExecutionError.Recoverability field on the wrapped *core.ExecutionError. Since the Go bool is dropped during proto encoding, any downstream service that consumes the ExecutionError over the wire sees the proto3 zero (NON_RECOVERABLE) for every failure regardless of the container's intent.
Set Recoverability from errorDoc.Error.Kind so the bit travels intact. The Go bool continues to be set in parallel for in-process callers; both agree.
Motivation
In our setup, the leasor reads ExecutionError.Recoverability off the wire to honor NonRecoverableError from the SDK without re-reading error.pb. Without this fix, NonRecoverableError is indistinguishable from a recoverable failure on the wire, defeating the SDK-side contract that NonRecoverableError consumes no retries.
This is a strict superset behavior change — Recoverability was previously zero (NON_RECOVERABLE) for both recoverable and non-recoverable error.pb contents, so the field was useless. Now both cases set it correctly.
Tests
Existing TestReadOrigin extended:
• user sub-test (NON_RECOVERABLE) now also asserts ee.GetRecoverability() == ContainerError_NON_RECOVERABLE
• system sub-test (RECOVERABLE) now also asserts ee.GetRecoverability() == ContainerError_RECOVERABLE
Test plan
• go test ./flyteplugins/go/tasks/pluginmachinery/ioutils/...
flyteorg/flyteacoustic-carpenter-78188
05/21/2026, 11:24 PM