0.32.7

src/balance/lease_runtime.rs

@@ -40,9 +40,26 @@ impl Lease {
             "ok"
         };
         let before = self.member.state.healthy();
-        self.member
-            .state
-            .observe_passive(&self.health, &self.passive_health, reason, elapsed);
+        let failure_detail = if is_5xx {
+            format!("status={status}")
+        } else {
+            reason.into()
+        };
+        let ejected = self.member.state.observe_passive_with_detail(
+            &self.health,
+            &self.passive_health,
+            reason,
+            reason,
+            &failure_detail,
+            elapsed,
+        );
+        if let Some((first, failures)) = ejected {
+            tracing::warn!(backend = %self.member.definition.id, origin = %self.member.backend.origin, source = "passive",
+                reason, first_error_class = first.class,
+                first_error_at_ms = first.at_ms, first_error_detail = %first.detail,
+                trigger_error_detail = %failure_detail, failures,
+                "backend excluded from traffic");
+        }
         let after = self.member.state.healthy();
         if before != after {
             tracing::info!(
@@ -62,6 +79,11 @@ impl Lease {
             self.member.state.errors.fetch_add(1, Ordering::Relaxed);
         }
         self.member.state.record_failure(class, detail);
+        // GOAWAY(NO_ERROR) is a graceful H2 connection shutdown, not an
+        // origin failure. It can still be surfaced in access diagnostics.
+        if matches!(class, "h2_graceful_goaway" | "h2_graceful_close") {
+            return;
+        }
         let passive_reason = match class {
             "connect" | "dns" | "tls" => "connect_error",
             "timeout"
@@ -72,12 +94,21 @@ impl Lease {
             _ => "io_error",
         };
         let before = self.member.state.healthy();
-        self.member.state.observe_passive(
+        let ejected = self.member.state.observe_passive_with_detail(
             &self.health,
             &self.passive_health,
             passive_reason,
+            class,
+            detail,
             elapsed,
         );
+        if let Some((first, failures)) = ejected {
+            tracing::warn!(backend = %self.member.definition.id, origin = %self.member.backend.origin, source = "passive",
+                reason = passive_reason, first_error_class = first.class,
+                first_error_at_ms = first.at_ms, first_error_detail = %first.detail,
+                trigger_error_detail = %detail, failures,
+                "backend excluded from traffic");
+        }
         let after = self.member.state.healthy();
         if before != after {
             tracing::info!(
@@ -326,8 +357,18 @@ async fn probe(member: Arc<Member>, settings: HealthCheck) {
     };
     let before = member.state.healthy();
     member.state.observe_active(&settings, result, elapsed);
-    if before != member.state.healthy() {
-        tracing::info!(backend = %member.definition.id, healthy = member.state.healthy(), reason = result, "backend health changed");
+    let after = member.state.healthy();
+    if before && !after {
+        let first = member.state.first_active_failure();
+        tracing::warn!(backend = %member.definition.id, origin = %member.backend.origin, source = "active",
+            reason = result,
+            first_error_class = first.as_ref().map_or("active_health", |f| f.class),
+            first_error_at_ms = first.as_ref().map_or(0, |f| f.at_ms),
+            first_error_detail = %first.as_ref().map_or(result, |f| f.detail.as_str()),
+            "backend excluded from traffic");
+    }
+    if before != after {
+        tracing::info!(backend = %member.definition.id, healthy = after, reason = result, "backend health changed");
     }
 }
 /// One supervisor, at most 16 probes, at most one outstanding probe per member.