feat: implement clock expiry scanning and handling for game records (#53)

Reviewed-on: #53

modules/coordinator/src/main/scala/de/nowchess/coordinator/service/AutoScaler.scala

@@ -10,9 +10,11 @@ import io.fabric8.kubernetes.client.KubernetesClient
 import io.micrometer.core.instrument.{Gauge, MeterRegistry}
 import io.quarkus.scheduler.Scheduled
 import org.jboss.logging.Logger
+import io.fabric8.kubernetes.client.KubernetesClientException
 import scala.jdk.CollectionConverters.*
 
 import java.util.concurrent.atomic.AtomicReference
+import java.util.concurrent.ConcurrentHashMap
 import scala.compiletime.uninitialized
 
 @ApplicationScoped
@@ -37,9 +39,19 @@ class AutoScaler:
   private var meterRegistry: MeterRegistry = uninitialized
   // scalafix:on DisableSyntax.var
 
-  private val log           = Logger.getLogger(classOf[AutoScaler])
-  private val lastScaleTime = new java.util.concurrent.atomic.AtomicLong(0L)
-  private val avgLoadRef    = new AtomicReference[Double](0.0)
+  private val log                  = Logger.getLogger(classOf[AutoScaler])
+  private val lastScaleTime        = new java.util.concurrent.atomic.AtomicLong(0L)
+  private val avgLoadRef           = new AtomicReference[Double](0.0)
+  private val drainingForScaleDown = ConcurrentHashMap.newKeySet[String]()
+
+  def isDrainingForScaleDown(instanceId: String): Boolean =
+    drainingForScaleDown.contains(instanceId)
+
+  def clearDraining(instanceId: String): Unit =
+    drainingForScaleDown.remove(instanceId)
+
+  def clearDrainingByPodName(podName: String): Unit =
+    drainingForScaleDown.asScala.find(podName.contains).foreach(drainingForScaleDown.remove)
 
   private def kubeClientOpt: Option[KubernetesClient] =
     if kubeClientInstance.isUnsatisfied then None
@@ -125,49 +137,69 @@ class AutoScaler:
           else if avgLoad < config.scaleDownThreshold * config.maxGamesPerCore && instances.size > config.scaleMinReplicas
           then scaleDown()
 
+  private def patchRolloutReplicas(
+      kube: KubernetesClient,
+      direction: String,
+      delta: Int,
+      canScale: Int => Boolean,
+      atLimit: Int => Unit,
+      onSuccess: (Int, Int) => Unit,
+      maxRetries: Int = 3,
+  ): Unit =
+    def attempt(retries: Int): Unit =
+      try
+        Option(
+          kube
+            .genericKubernetesResources(argoApiVersion, argoKind)
+            .inNamespace(config.k8sNamespace)
+            .withName(config.k8sRolloutName)
+            .get(),
+        ).foreach { rollout =>
+          rolloutSpec(rollout).foreach { spec =>
+            spec.get("replicas") match
+              case current: Integer =>
+                val n = current.intValue()
+                if !canScale(n) then atLimit(n)
+                else
+                  spec.put("replicas", Integer.valueOf(n + delta))
+                  kube
+                    .genericKubernetesResources(argoApiVersion, argoKind)
+                    .inNamespace(config.k8sNamespace)
+                    .resource(rollout)
+                    .update()
+                  meterRegistry.counter("nowchess.coordinator.scale.events", "direction", direction).increment()
+                  onSuccess(n, n + delta)
+              case _ => ()
+          }
+        }
+      catch
+        case ex: KubernetesClientException if ex.getCode == 409 =>
+          if retries > 0 then
+            log.debugf("Conflict scaling %s %s, retrying (%d left)", direction, config.k8sRolloutName, retries - 1)
+            attempt(retries - 1)
+          else
+            meterRegistry.counter("nowchess.coordinator.scale.failures", "direction", direction).increment()
+            log.errorf(ex, "Failed to scale %s %s after conflict retries", direction, config.k8sRolloutName)
+        case ex: Exception =>
+          meterRegistry.counter("nowchess.coordinator.scale.failures", "direction", direction).increment()
+          log.errorf(ex, "Failed to scale %s %s", direction, config.k8sRolloutName)
+    attempt(maxRetries)
+
   def scaleUp(): Unit =
     log.info("Scaling up Argo Rollout")
     kubeClientOpt match
-      case None =>
-        log.warn("Kubernetes client not available, cannot scale")
+      case None => log.warn("Kubernetes client not available, cannot scale")
       case Some(kube) =>
-        try
-          Option(
-            kube
-              .genericKubernetesResources(argoApiVersion, argoKind)
-              .inNamespace(config.k8sNamespace)
-              .withName(config.k8sRolloutName)
-              .get(),
-          ).foreach { rollout =>
-            rolloutSpec(rollout).foreach { spec =>
-              spec.get("replicas") match
-                case replicas: Integer =>
-                  val currentReplicas = replicas.intValue()
-                  val maxReplicas     = config.scaleMaxReplicas
-
-                  if currentReplicas < maxReplicas then
-                    spec.put("replicas", Integer.valueOf(currentReplicas + 1))
-                    kube
-                      .genericKubernetesResources(argoApiVersion, argoKind)
-                      .inNamespace(config.k8sNamespace)
-                      .resource(rollout)
-                      .update()
-                    meterRegistry.counter("nowchess.coordinator.scale.events", "direction", "up").increment()
-                    log.infof(
-                      "Scaled up %s from %d to %d replicas",
-                      config.k8sRolloutName,
-                      currentReplicas,
-                      currentReplicas + 1,
-                    )
-                    loadBalancer.rebalance
-                  else log.infof("Already at max replicas %d for %s", maxReplicas, config.k8sRolloutName)
-                case _ => ()
-            }
-          }
-        catch
-          case ex: Exception =>
-            meterRegistry.counter("nowchess.coordinator.scale.failures", "direction", "up").increment()
-            log.errorf(ex, "Failed to scale up %s", config.k8sRolloutName)
+        patchRolloutReplicas(
+          kube,
+          direction = "up",
+          delta = 1,
+          canScale = _ < config.scaleMaxReplicas,
+          atLimit = n => log.infof("Already at max replicas %d for %s", n, config.k8sRolloutName),
+          onSuccess = (from, to) =>
+            log.infof("Scaled up %s from %d to %d replicas", config.k8sRolloutName, from, to)
+            loadBalancer.rebalance,
+        )
 
   def scaleDown(): Unit =
     log.info("Scaling down Argo Rollout")
@@ -177,6 +209,7 @@ class AutoScaler:
 
     underloadedInstance.foreach { inst =>
       log.infof("Marking instance %s for drain before scale-down", inst.instanceId)
+      drainingForScaleDown.add(inst.instanceId)
       failoverService
         .onInstanceStreamDropped(inst.instanceId)
         .subscribe()
@@ -187,42 +220,34 @@ class AutoScaler:
     }
 
     kubeClientOpt match
-      case None =>
-        log.warn("Kubernetes client not available, cannot scale")
+      case None => log.warn("Kubernetes client not available, cannot scale")
       case Some(kube) =>
-        try
-          Option(
-            kube
-              .genericKubernetesResources(argoApiVersion, argoKind)
-              .inNamespace(config.k8sNamespace)
-              .withName(config.k8sRolloutName)
-              .get(),
-          ).foreach { rollout =>
-            rolloutSpec(rollout).foreach { spec =>
-              spec.get("replicas") match
-                case replicas: Integer =>
-                  val currentReplicas = replicas.intValue()
-                  val minReplicas     = config.scaleMinReplicas
+        patchRolloutReplicas(
+          kube,
+          direction = "down",
+          delta = -1,
+          canScale = _ > config.scaleMinReplicas,
+          atLimit = n => log.infof("Already at min replicas %d for %s", n, config.k8sRolloutName),
+          onSuccess = (from, to) =>
+            log.infof("Scaled down %s from %d to %d replicas", config.k8sRolloutName, from, to)
+            underloadedInstance.foreach(inst => forceDeletePod(inst.instanceId, kube)),
+        )
 
-                  if currentReplicas > minReplicas then
-                    spec.put("replicas", Integer.valueOf(currentReplicas - 1))
-                    kube
-                      .genericKubernetesResources(argoApiVersion, argoKind)
-                      .inNamespace(config.k8sNamespace)
-                      .resource(rollout)
-                      .update()
-                    meterRegistry.counter("nowchess.coordinator.scale.events", "direction", "down").increment()
-                    log.infof(
-                      "Scaled down %s from %d to %d replicas",
-                      config.k8sRolloutName,
-                      currentReplicas,
-                      currentReplicas - 1,
-                    )
-                  else log.infof("Already at min replicas %d for %s", minReplicas, config.k8sRolloutName)
-                case _ => ()
-            }
-          }
-        catch
-          case ex: Exception =>
-            meterRegistry.counter("nowchess.coordinator.scale.failures", "direction", "down").increment()
-            log.errorf(ex, "Failed to scale down %s", config.k8sRolloutName)
+  private def forceDeletePod(instanceId: String, kube: KubernetesClient): Unit =
+    try
+      val pods = kube
+        .pods()
+        .inNamespace(config.k8sNamespace)
+        .withLabel(config.k8sRolloutLabelSelector)
+        .list()
+        .getItems
+        .asScala
+      pods.find(_.getMetadata.getName.contains(instanceId)) match
+        case Some(pod) =>
+          kube.pods().inNamespace(config.k8sNamespace).withName(pod.getMetadata.getName).withGracePeriod(0L).delete()
+          log.infof("Force-deleted pod for drained instance %s", instanceId)
+        case None =>
+          log.debugf("No pod found for drained instance %s, skipping deletion", instanceId)
+    catch
+      case ex: Exception =>
+        log.warnf(ex, "Failed to force-delete pod for drained instance %s", instanceId)

modules/coordinator/src/main/scala/de/nowchess/coordinator/service/HealthMonitor.scala

@@ -88,7 +88,8 @@ class HealthMonitor:
     if evicted.nonEmpty then
       log.warnf("Evicted %d stale instances: %s", evicted.size, evicted.mkString(", "))
       evicted.foreach(deleteK8sPod)
-      autoScaler.scaleUp()
+      val unexpectedEvictions = evicted.filterNot(autoScaler.isDrainingForScaleDown)
+      if unexpectedEvictions.nonEmpty then autoScaler.scaleUp()
     val instances = instanceRegistry.getAllInstances
     val failed = instances.collect { inst =>
       val isHealthy = checkHealth(inst.instanceId)
@@ -99,7 +100,8 @@ class HealthMonitor:
         Some(inst.instanceId)
       else None
     }.flatten
-    if failed.nonEmpty then autoScaler.scaleUp()
+    val unexpectedFailures = failed.filterNot(autoScaler.isDrainingForScaleDown)
+    if unexpectedFailures.nonEmpty then autoScaler.scaleUp()
 
   private def checkHealth(instanceId: String): Boolean =
     val redisHealthy = checkRedisHeartbeat(instanceId)
@@ -227,12 +229,10 @@ class HealthMonitor:
     }
 
   private def handlePodGone(pod: Pod): Unit =
+    val podName = pod.getMetadata.getName
+    autoScaler.clearDrainingByPodName(podName)
     findRegisteredInstance(pod).foreach { inst =>
-      log.warnf(
-        "Pod %s deleted — triggering failover for %s",
-        pod.getMetadata.getName,
-        inst.instanceId,
-      )
+      log.warnf("Pod %s deleted — triggering failover for %s", podName, inst.instanceId)
       failoverService
         .onInstanceStreamDropped(inst.instanceId)
         .subscribe()