Keep local readiness responsive when optional models stall #273

Merged
m merged 2 commits from codex/chat-semantic-inventory into main 2026-08-24 16:39:42 +00:00
3 changed files with 72 additions and 31 deletions
+43 -28
View File
@@ -6,6 +6,7 @@ import (
"fmt"
"net/http"
"strings"
"sync"
"time"
"git.cin.su/m/Assistant/contracts/capability"
@@ -31,35 +32,16 @@ type fleetStatusResponse struct {
func fleetReadinessHandler(client *http.Client, preferred capability.FleetTier, endpoints []fleetEndpoint) http.Handler {
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
members := make([]capability.FleetMember, 0, len(endpoints))
for _, endpoint := range endpoints {
member := capability.FleetMember{ProviderInstanceID: capability.ProviderInstanceID(endpoint.ProviderID), ModelDescriptorID: capability.ModelDescriptorID(endpoint.ModelID), Tier: endpoint.Tier, Reason: "NOT_CONFIGURED"}
if strings.TrimSpace(endpoint.URL) == "" {
members = append(members, member)
continue
}
request, err := http.NewRequestWithContext(r.Context(), http.MethodGet, endpoint.URL, nil)
if err == nil {
for name, value := range endpoint.Headers {
request.Header.Set(name, value)
}
probeClient := endpoint.Client
if probeClient == nil {
probeClient = client
}
response, requestErr := probeClient.Do(request)
if requestErr == nil {
response.Body.Close()
if response.StatusCode >= 200 && response.StatusCode < 300 {
member.Ready, member.Reason = true, "AVAILABLE"
} else {
member.Reason = fmt.Sprintf("status=%s", response.Status)
}
} else {
member.Reason = requestErr.Error()
}
}
members = append(members, member)
members = make([]capability.FleetMember, len(endpoints))
var waitGroup sync.WaitGroup
for index, endpoint := range endpoints {
waitGroup.Add(1)
go func(index int, endpoint fleetEndpoint) {
defer waitGroup.Done()
members[index] = probeFleetEndpoint(r.Context(), client, endpoint)
}(index, endpoint)
}
waitGroup.Wait()
readiness, err := fleetdomain.BuildFleetReadiness(time.Now().UnixMilli(), members)
response := fleetStatusResponse{Readiness: readiness, Preferred: preferred}
if err == nil {
@@ -76,4 +58,37 @@ func fleetReadinessHandler(client *http.Client, preferred capability.FleetTier,
})
}
func probeFleetEndpoint(parent context.Context, fallbackClient *http.Client, endpoint fleetEndpoint) capability.FleetMember {
member := capability.FleetMember{ProviderInstanceID: capability.ProviderInstanceID(endpoint.ProviderID), ModelDescriptorID: capability.ModelDescriptorID(endpoint.ModelID), Tier: endpoint.Tier, Reason: "NOT_CONFIGURED"}
if strings.TrimSpace(endpoint.URL) == "" {
return member
}
ctx, cancel := context.WithTimeout(parent, 2*time.Second)
defer cancel()
request, err := http.NewRequestWithContext(ctx, http.MethodGet, endpoint.URL, nil)
if err != nil {
member.Reason = err.Error()
return member
}
for name, value := range endpoint.Headers {
request.Header.Set(name, value)
}
probeClient := endpoint.Client
if probeClient == nil {
probeClient = fallbackClient
}
response, requestErr := probeClient.Do(request)
if requestErr != nil {
member.Reason = requestErr.Error()
return member
}
defer response.Body.Close()
if response.StatusCode >= 200 && response.StatusCode < 300 {
member.Ready, member.Reason = true, "AVAILABLE"
} else {
member.Reason = fmt.Sprintf("status=%s", response.Status)
}
return member
}
func fleetContext(ctx context.Context) context.Context { return ctx }
+24
View File
@@ -5,6 +5,7 @@ import (
"net/http/httptest"
"strings"
"testing"
"time"
"git.cin.su/m/Assistant/contracts/capability"
)
@@ -26,6 +27,29 @@ func TestFleetReadinessSelectsBestAvailablePrimaryTierBeforeLifeboat(t *testing.
}
}
func TestFleetReadinessDoesNotWaitForSlowMembersSerially(t *testing.T) {
slow := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
time.Sleep(150 * time.Millisecond)
w.WriteHeader(http.StatusServiceUnavailable)
}))
defer slow.Close()
fast := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(http.StatusOK) }))
defer fast.Close()
handler := fleetReadinessHandler(fast.Client(), capability.FleetTierFast, []fleetEndpoint{
{ProviderID: "slow", ModelID: "slow-model", Tier: capability.FleetTierMedium, URL: slow.URL},
{ProviderID: "lifeboat", ModelID: "lifeboat-model", Tier: capability.FleetTierLifeboat, URL: fast.URL},
})
started := time.Now()
recorder := httptest.NewRecorder()
handler.ServeHTTP(recorder, httptest.NewRequest(http.MethodGet, "/v1/fleet/readiness", nil))
if recorder.Code != http.StatusOK || !strings.Contains(recorder.Body.String(), `"selected_tier":"LIFEBOAT"`) {
t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
}
if elapsed := time.Since(started); elapsed >= 300*time.Millisecond {
t.Fatalf("fleet readiness was serialized: elapsed=%s", elapsed)
}
}
func TestFleetReadinessSendsConfiguredProviderHeaders(t *testing.T) {
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.Header.Get("Authorization") != "Bearer fleet-key" {
+5 -3
View File
@@ -126,7 +126,9 @@ func main() {
}
cpuEndpoint := strings.TrimRight(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), "/") + "/models"
cpuHealthProbe := health.OptionalHTTPProbeWithHeaders(cpuReadinessClient, cpuEndpoint, cpuFleetHeaders)
cpuHealthProbe.Timeout = 10 * time.Second
// Opportunistic providers must not hide a ready local lifeboat. Keep their
// readiness budget short; their detailed fleet state is reported separately.
cpuHealthProbe.Timeout = 2 * time.Second
mediumEndpoint := ""
mediumHeaders := map[string]string{}
mediumHealthProbe := health.Probe{}
@@ -136,7 +138,7 @@ func main() {
mediumHeaders["Authorization"] = "Bearer " + mediumAPIKey
}
mediumHealthProbe = health.OptionalHTTPProbeWithHeaders(mediumReadinessClient, mediumEndpoint, mediumHeaders)
mediumHealthProbe.Timeout = 10 * time.Second
mediumHealthProbe.Timeout = 2 * time.Second
}
readinessProbes := []health.Probe{
health.HTTPProbe(readinessClient, strings.TrimRight(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), "/")+"/healthz"),
@@ -149,7 +151,7 @@ func main() {
if gpuConfigured {
gpuHeaders := map[string]string{"Authorization": "Bearer " + gpuKey}
gpuProbe := health.OptionalHTTPProbeWithHeaders(readinessClient, strings.TrimRight(gpuURL, "/")+"/models", gpuHeaders)
gpuProbe.Timeout = 10 * time.Second
gpuProbe.Timeout = 2 * time.Second
readinessProbes = append(readinessProbes, gpuProbe)
}