test(bdd): retry transient LLM eligibility

Signed-off-by: Mike Camp <mcamp@nvidia.com>

Author: mikeyrcamp

tests/bdd/steps/nvcf_cli_steps.go

@@ -21,12 +21,15 @@ import (
 	"context"
 	"fmt"
 	"strings"
+	"time"
 
 	"github.com/cucumber/godog"
 
 	"nvcf-bdd/dsl"
 )
 
+var modelInvocationRetryInterval = time.Second
+
 func registerNVCFCLISteps(ctx *godog.ScenarioContext, sc *ScenarioContext) {
 	ctx.Step(`^I use NVCF CLI config "([^"]*)"$`, sc.iUseNVCFCLIConfig)
 	ctx.Step(`^I successfully create function "([^"]*)" from image "([^"]*)" with CLI options:$`, sc.iSuccessfullyCreateFunction)
@@ -101,13 +104,39 @@ func (sc *ScenarioContext) iSuccessfullyInvokeModel(
 	timeout string,
 	doc *godog.DocString,
 ) error {
-	return sc.runNVCFCLI(ctx,
+	args := []string{
 		"function", "invoke",
 		"--inference-url", inferenceURL,
 		"--model-name", model,
 		"--request-body", doc.Content,
 		"--timeout", timeout,
-	)
+	}
+	retryFor, retryTimeoutErr := time.ParseDuration(timeout + "s")
+	deadline := time.Now().Add(retryFor)
+
+	for {
+		err := sc.runNVCFCLI(ctx, args...)
+		if err == nil {
+			return nil
+		}
+		if retryTimeoutErr != nil || retryFor <= 0 ||
+			!strings.Contains(combinedOutput(sc.LastResult), "no_eligible_candidates") {
+			return err
+		}
+
+		remaining := time.Until(deadline)
+		if remaining <= 0 {
+			return err
+		}
+		waitFor := min(modelInvocationRetryInterval, remaining)
+		timer := time.NewTimer(waitFor)
+		select {
+		case <-ctx.Done():
+			timer.Stop()
+			return ctx.Err()
+		case <-timer.C:
+		}
+	}
 }
 
 func (sc *ScenarioContext) iSuccessfullyUndeploySelectedFunction(ctx context.Context) error {

tests/bdd/steps/nvcf_cli_steps_test.go

@@ -22,6 +22,7 @@ import (
 	"errors"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/cucumber/godog"
 
@@ -119,6 +120,55 @@ func TestNVCFCLIInvocationAdaptersExposeAllArguments(t *testing.T) {
 	}
 }
 
+func TestNVCFCLIModelInvocationRetriesNoEligibleCandidates(t *testing.T) {
+	sc, fake := newScenarioContext(t)
+	t.Setenv("NVCF_CLI", "nvcf-cli")
+	sc.NVCFCLIConfig = "config.yaml"
+	fake.runResults = []harness.Result{
+		{ExitCode: 1, Stderr: `API error 404: {"code":"no_eligible_candidates"}`},
+		{ExitCode: 0, Stdout: `{"object":"chat.completion"}`},
+	}
+
+	previousInterval := modelInvocationRetryInterval
+	modelInvocationRetryInterval = time.Nanosecond
+	t.Cleanup(func() { modelInvocationRetryInterval = previousInterval })
+
+	err := sc.iSuccessfullyInvokeModel(
+		context.Background(),
+		"model/name",
+		"/v1/chat/completions",
+		"1",
+		&godog.DocString{Content: `{"messages":[]}`},
+	)
+	if err != nil {
+		t.Fatalf("invoke model: %v", err)
+	}
+	if len(fake.runs) != 2 {
+		t.Fatalf("runs = %d, want 2", len(fake.runs))
+	}
+}
+
+func TestNVCFCLIModelInvocationDoesNotRetryOtherErrors(t *testing.T) {
+	sc, fake := newScenarioContext(t)
+	t.Setenv("NVCF_CLI", "nvcf-cli")
+	sc.NVCFCLIConfig = "config.yaml"
+	fake.result = harness.Result{ExitCode: 1, Stderr: "API error 401: unauthorized"}
+
+	err := sc.iSuccessfullyInvokeModel(
+		context.Background(),
+		"model/name",
+		"/v1/chat/completions",
+		"1",
+		&godog.DocString{Content: `{"messages":[]}`},
+	)
+	if err == nil || !strings.Contains(err.Error(), "exit code = 1, want 0") {
+		t.Fatalf("error = %v, want exit-zero assertion failure", err)
+	}
+	if len(fake.runs) != 1 {
+		t.Fatalf("runs = %d, want 1", len(fake.runs))
+	}
+}
+
 func TestNVCFCLISuccessStepRequiresExitZero(t *testing.T) {
 	sc, fake := newScenarioContext(t)
 	t.Setenv("NVCF_CLI", "nvcf-cli")