gke: private nodes, DNS endpoint, Dataplane V2, cost allocation, monitoring, max_pods_per_node, nodepool labels/taints

GKE cluster hardening and observability options, plus per-nodepool max_pods_per_node, node_labels, and node_taints wired through the container cluster spec.

Author: geojaz

perfkitbenchmarker/configs/container_spec.py

@@ -243,6 +243,9 @@ def __init__(
     self.vm_spec: virtual_machine_spec.BaseVmSpec
     self.machine_families: list[str] | None
     self.sandbox_config: SandboxSpec | None
+    self.max_pods_per_node: int | None
+    self.node_labels: dict[str, str] | None
+    self.node_taints: list[str] | None
 
   @classmethod
   def _GetOptionDecoderConstructions(cls):
@@ -273,6 +276,18 @@ def _GetOptionDecoderConstructions(cls):
         ),
         'vm_spec': (spec.PerCloudConfigDecoder, {}),
         'sandbox_config': (_SandboxDecoder, {'default': None}),
+        'max_pods_per_node': (
+            option_decoders.IntDecoder,
+            {'default': None, 'none_ok': True, 'min': 1},
+        ),
+        'node_labels': (
+            option_decoders.TypeVerifier,
+            {'valid_types': (dict,), 'default': None, 'none_ok': True},
+        ),
+        'node_taints': (
+            option_decoders.TypeVerifier,
+            {'valid_types': (list,), 'default': None, 'none_ok': True},
+        ),
     })
     return result
 

perfkitbenchmarker/providers/gcp/flags.py

@@ -67,9 +67,7 @@
     'Whether or not we create a Confidential VM Instance',
 )
 GCE_CONFIDENTIAL_COMPUTE_TYPE = flags.DEFINE_string(
-    'gce_confidential_compute_type',
-    'sev',
-    'Type of Confidential VM Instance'
+    'gce_confidential_compute_type', 'sev', 'Type of Confidential VM Instance'
 )
 GCE_NETWORK_NAMES = flags.DEFINE_list(
     'gce_network_name',
@@ -561,6 +559,97 @@
     False,
     'Whether to enable shielded nodes.',
 )
+GKE_ENABLE_PRIVATE_NODES = flags.DEFINE_boolean(
+    'gke_enable_private_nodes',
+    False,
+    'Whether to create the cluster with private nodes (nodes have only internal'
+    ' IPs).',
+)
+GKE_ENABLE_DNS_ACCESS = flags.DEFINE_boolean(
+    'gke_enable_dns_access',
+    False,
+    'Whether to enable DNS-based control plane access (replaces the'
+    ' public/private IP endpoint model).',
+)
+GKE_ENABLE_IP_ACCESS = flags.DEFINE_boolean(
+    'gke_enable_ip_access',
+    True,
+    'Whether to enable IP-based control plane access. Disabling requires DNS'
+    ' access and is mutually exclusive with public clusters (nodes with public'
+    ' IPs).',
+)
+GKE_MASTER_IPV4_CIDR = flags.DEFINE_string(
+    'gke_master_ipv4_cidr',
+    None,
+    'CIDR range to use for the hosted master network. Required when private'
+    ' nodes are enabled without DNS access.',
+)
+GKE_ENABLE_DATAPLANE_V2 = flags.DEFINE_boolean(
+    'gke_enable_dataplane_v2',
+    False,
+    'Whether to enable GKE Dataplane V2 (eBPF-based datapath, Cilium under the'
+    ' hood). Requires cluster recreation; cannot be toggled on an existing'
+    ' cluster.',
+)
+GKE_ENABLE_MANAGED_PROMETHEUS = flags.DEFINE_boolean(
+    'gke_enable_managed_prometheus',
+    False,
+    'Whether to enable Google Cloud Managed Service for Prometheus on the'
+    ' cluster.',
+)
+GKE_ENABLE_COST_ALLOCATION = flags.DEFINE_boolean(
+    'gke_enable_cost_allocation',
+    False,
+    'Whether to enable GKE cost allocation tracking.',
+)
+GKE_MONITORING_COMPONENTS = flags.DEFINE_string(
+    'gke_monitoring_components',
+    'SYSTEM,API_SERVER,SCHEDULER,CONTROLLER_MANAGER',
+    'Comma-separated list of GKE monitoring components to enable '
+    '(e.g. SYSTEM,API_SERVER,SCHEDULER,CONTROLLER_MANAGER,POD,DEPLOYMENT,'
+    'STATEFULSET,DAEMONSET,HPA,STORAGE,CADVISOR,KUBELET).',
+)
+GKE_ENABLE_AGENT_SANDBOX = flags.DEFINE_boolean(
+    'gke_enable_agent_sandbox',
+    False,
+    'Whether to enable the GKE Agent Sandbox controller on the cluster. '
+    'Installs the managed agent-sandbox controller and CRDs, enabling '
+    'SandboxClaim/Sandbox/SandboxWarmPool reconciliation by GKE. This is '
+    'separate from the gvisor sandbox runtime (--sandbox=type=gvisor on a '
+    'node pool). Requires GKE 1.35.2-gke.1269000 or later. See '
+    'https://docs.cloud.google.com/kubernetes-engine/docs/how-to/agent-sandbox.',
+)
+
+
+def _ValidateGkePrivateNodeFlags(flags_dict):
+  if (
+      not flags_dict['gke_enable_ip_access']
+      and not flags_dict['gke_enable_dns_access']
+  ):
+    raise flags.ValidationError(
+        '--no-gke_enable_ip_access requires --gke_enable_dns_access.'
+    )
+  if (
+      flags_dict['gke_enable_private_nodes']
+      and not flags_dict['gke_enable_dns_access']
+      and not flags_dict['gke_master_ipv4_cidr']
+  ):
+    raise flags.ValidationError(
+        '--gke_enable_private_nodes without --gke_enable_dns_access requires'
+        ' --gke_master_ipv4_cidr.'
+    )
+  return True
+
+
+flags.register_multi_flags_validator(
+    [
+        'gke_enable_ip_access',
+        'gke_enable_dns_access',
+        'gke_enable_private_nodes',
+        'gke_master_ipv4_cidr',
+    ],
+    _ValidateGkePrivateNodeFlags,
+)
 GKE_ADDONS = flags.DEFINE_string(
     'gke_addons',
     '',

perfkitbenchmarker/providers/gcp/gce_network.py

@@ -914,6 +914,7 @@ class GceNetwork(network.BaseNetwork):
   def __init__(self, network_spec: GceNetworkSpec):
     super().__init__(network_spec)
     self.project: str | None = network_spec.project
+    self._zone: str = network_spec.zone
     self.vpn_gateway: Dict[str, GceVpnGateway] = {}
 
     #  Figuring out the type of network here.
@@ -1231,6 +1232,60 @@ def _GetNumberVms(self) -> int:
         for group_spec in benchmark_spec.config.vm_groups.values()
     )
 
+  def _CreateCloudNat(self):
+    """Provision a Cloud Router + NAT so private resources can egress.
+
+    Called during network provisioning so NAT has time to fully propagate
+    before any cluster lifecycle code starts. Shared across all resources
+    in the network.
+    """
+    region = util.GetRegionFromZone(self._zone)
+    router_name = f'{self.primary_subnet_name}-router'
+    nat_name = f'{self.primary_subnet_name}-nat'
+
+    router_cmd = util.GcloudCommand(
+        self, 'compute', 'routers', 'create', router_name
+    )
+    router_cmd.flags['network'] = self.primary_subnet_name
+    router_cmd.flags['region'] = region
+    router_cmd.flags.pop('zone', None)
+    _, stderr, retcode = router_cmd.Issue(raise_on_failure=False)
+    if retcode and 'already exists' not in stderr:
+      logging.warning('Cloud Router create failed: %s', stderr)
+
+    nat_cmd = util.GcloudCommand(
+        self, 'compute', 'routers', 'nats', 'create', nat_name
+    )
+    nat_cmd.flags['router'] = router_name
+    nat_cmd.flags['region'] = region
+    nat_cmd.flags.pop('zone', None)
+    nat_cmd.args.append('--auto-allocate-nat-external-ips')
+    nat_cmd.args.append('--nat-all-subnet-ip-ranges')
+    _, stderr, retcode = nat_cmd.Issue(raise_on_failure=False)
+    if retcode and 'already exists' not in stderr:
+      logging.warning('Cloud NAT create failed: %s', stderr)
+
+  def _DeleteCloudNat(self):
+    """Best-effort teardown of the NAT and router this network created."""
+    region = util.GetRegionFromZone(self._zone)
+    router_name = f'{self.primary_subnet_name}-router'
+    nat_name = f'{self.primary_subnet_name}-nat'
+
+    nat_cmd = util.GcloudCommand(
+        self, 'compute', 'routers', 'nats', 'delete', nat_name
+    )
+    nat_cmd.flags['router'] = router_name
+    nat_cmd.flags['region'] = region
+    nat_cmd.flags.pop('zone', None)
+    nat_cmd.Issue(raise_on_failure=False)
+
+    router_cmd = util.GcloudCommand(
+        self, 'compute', 'routers', 'delete', router_name
+    )
+    router_cmd.flags['region'] = region
+    router_cmd.flags.pop('zone', None)
+    router_cmd.Issue(raise_on_failure=False)
+
   def Create(self):
     """Creates the actual network."""
     if not self.is_existing_network:
@@ -1244,6 +1299,8 @@ def Create(self):
             lambda rule: self.external_nets_rules[rule].Create(),
             list(self.external_nets_rules.keys()),
         )
+      if gcp_flags.GKE_ENABLE_PRIVATE_NODES.value:
+        self._CreateCloudNat()
       if getattr(self, 'vpn_gateway', False):
         background_tasks.RunThreaded(
             lambda gateway: self.vpn_gateway[gateway].Create(),
@@ -1257,6 +1314,12 @@ def Delete(self):
     if self.placement_group:
       self.placement_group.Delete()
     if not self.is_existing_network:
+      # Always attempt NAT+router cleanup: both gcloud calls use
+      # raise_on_failure=False so this is a no-op when none was created.
+      # Checking the live flag here is wrong because teardown runs from a
+      # restored pickle and the flag may default to False even when a NAT
+      # was created at provision time.
+      self._DeleteCloudNat()
       if getattr(self, 'vpn_gateway', False):
         background_tasks.RunThreaded(
             lambda gateway: self.vpn_gateway[gateway].Delete(),

perfkitbenchmarker/providers/gcp/google_kubernetes_engine.py

@@ -168,10 +168,37 @@ def _RunClusterCreateCommand(self, cmd: util.GcloudCommand):
         )
       cmd.flags['release-channel'] = self.release_channel
 
+    if gcp_flags.GKE_ENABLE_PRIVATE_NODES.value:
+      cmd.args.append('--enable-private-nodes')
+      # GKE requires VPC-native (alias IPs) when private nodes are enabled.
+      # Without this gcloud rejects the create with:
+      #   Cannot specify --enable-private-nodes without --enable-ip-alias.
+      cmd.args.append('--enable-ip-alias')
+    else:
+      cmd.args.append('--no-enable-private-nodes')
+    if gcp_flags.GKE_ENABLE_DNS_ACCESS.value:
+      cmd.args.append('--enable-dns-access')
+    else:
+      cmd.args.append('--no-enable-dns-access')
+    if gcp_flags.GKE_ENABLE_IP_ACCESS.value:
+      cmd.args.append('--enable-ip-access')
+    else:
+      cmd.args.append('--no-enable-ip-access')
+    if gcp_flags.GKE_ENABLE_DATAPLANE_V2.value:
+      cmd.args.append('--enable-dataplane-v2')
+    if gcp_flags.GKE_ENABLE_AGENT_SANDBOX.value:
+      cmd.args.append('--enable-agent-sandbox')
+    if gcp_flags.GKE_MASTER_IPV4_CIDR.value:
+      cmd.flags['master-ipv4-cidr'] = gcp_flags.GKE_MASTER_IPV4_CIDR.value
+
     if FLAGS.gke_enable_alpha:
       cmd.args.append('--enable-kubernetes-alpha')
       cmd.args.append('--no-enable-autorepair')
-    cmd.flags['monitoring'] = 'SYSTEM,API_SERVER,SCHEDULER,CONTROLLER_MANAGER'
+    cmd.flags['monitoring'] = gcp_flags.GKE_MONITORING_COMPONENTS.value
+    if gcp_flags.GKE_ENABLE_MANAGED_PROMETHEUS.value:
+      cmd.args.append('--enable-managed-prometheus')
+    if gcp_flags.GKE_ENABLE_COST_ALLOCATION.value:
+      cmd.args.append('--enable-cost-allocation')
 
     user = util.GetDefaultUser()
     if FLAGS.gcp_service_account:
@@ -209,6 +236,10 @@ def _GetKubeconfig(self):
     cmd = self._GcloudCommand(
         'container', 'clusters', 'get-credentials', self.name
     )
+    if gcp_flags.GKE_ENABLE_DNS_ACCESS.value:
+      # Private-node clusters are unreachable via the IP endpoint; use the
+      # DNS-based control plane endpoint instead.
+      cmd.args.append('--dns-endpoint')
     env = os.environ.copy()
     env['KUBECONFIG'] = FLAGS.kubeconfig
     cmd.IssueRetryable(env=env)
@@ -377,6 +408,10 @@ def _Create(self):
     cmd = self._GcloudCommand('container', 'clusters', 'create', self.name)
     if self.default_nodepool.network:
       cmd.flags['network'] = self.default_nodepool.network.network_resource.name
+      if self.default_nodepool.network.subnet_resource:
+        cmd.flags['subnetwork'] = (
+            self.default_nodepool.network.subnet_resource.name
+        )
 
     if gcp_flags.GKE_ENABLE_SHIELDED_NODES.value:
       cmd.args.append('--enable-shielded-nodes')
@@ -576,10 +611,19 @@ def _AddNodeParamsToCmd(
     if nodepool_config.sandbox_config is not None:
       cmd.flags['sandbox'] = nodepool_config.sandbox_config.ToSandboxFlag()
 
+    if nodepool_config.max_pods_per_node is not None:
+      cmd.flags['max-pods-per-node'] = nodepool_config.max_pods_per_node
+
     if self.image_type:
       cmd.flags['image-type'] = self.image_type
 
-    cmd.flags['node-labels'] = f'pkb_nodepool={nodepool_config.name}'
+    labels = {}
+    if nodepool_config.node_labels:
+      labels.update(nodepool_config.node_labels)
+    labels['pkb_nodepool'] = nodepool_config.name
+    cmd.flags['node-labels'] = ','.join(f'{k}={v}' for k, v in labels.items())
+    if nodepool_config.node_taints:
+      cmd.flags['node-taints'] = ','.join(nodepool_config.node_taints)
     if nodepool_config.min_nodes != nodepool_config.max_nodes:
       cmd.args.append('--enable-autoscaling')
       cmd.flags['min-nodes'] = nodepool_config.min_nodes
@@ -673,6 +717,10 @@ def _Create(self):
     )
     if self.default_nodepool.network:
       cmd.flags['network'] = self.default_nodepool.network.network_resource.name
+      if self.default_nodepool.network.subnet_resource:
+        cmd.flags['subnetwork'] = (
+            self.default_nodepool.network.subnet_resource.name
+        )
     cmd.flags['labels'] = util.MakeFormattedDefaultTags()
 
     if self.enable_aam:

perfkitbenchmarker/resources/container_service/container.py

@@ -187,6 +187,9 @@ def __init__(
     # Defined by GceVirtualMachineConfig. Used by google_kubernetes_engine
     # pylint: disable=g-missing-from-attributes
     self.sandbox_config: container_spec_lib.SandboxSpec | None = None
+    self.node_labels: dict[str, str] | None = None
+    self.node_taints: list[str] | None = None
+    self.max_pods_per_node: int | None = None
     self.max_local_disks: int | None
     self.ssd_interface: str | None
     self.threads_per_core: int

perfkitbenchmarker/resources/container_service/container_cluster.py

@@ -116,6 +116,9 @@ def _InitializeNodePool(
         nodepool_spec.machine_families,
     )
     nodepool_config.sandbox_config = nodepool_spec.sandbox_config
+    nodepool_config.node_labels = nodepool_spec.node_labels
+    nodepool_config.node_taints = nodepool_spec.node_taints
+    nodepool_config.max_pods_per_node = nodepool_spec.max_pods_per_node
     nodepool_config.zone = zone
     nodepool_config.num_nodes = nodepool_spec.vm_count
     if nodepool_spec.min_vm_count is None: