deployment.go

// Copyright 2022 Namespace Labs Inc; All rights reserved.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.

package kubernetes

import (
	"context"
	"crypto/sha256"
	"embed"
	"encoding/json"
	"fmt"
	"io"
	"io/fs"
	"path/filepath"
	"strings"
	"time"

	"github.com/dustin/go-humanize"
	specs "github.com/opencontainers/image-spec/specs-go/v1"
	"golang.org/x/exp/slices"
	corev1 "k8s.io/api/core/v1"
	v1 "k8s.io/api/core/v1"
	"k8s.io/apimachinery/pkg/api/resource"
	"k8s.io/apimachinery/pkg/util/intstr"
	appsv1 "k8s.io/client-go/applyconfigurations/apps/v1"
	applycorev1 "k8s.io/client-go/applyconfigurations/core/v1"
	applymetav1 "k8s.io/client-go/applyconfigurations/meta/v1"
	rbacv1 "k8s.io/client-go/applyconfigurations/rbac/v1"
	"namespacelabs.dev/foundation/framework/kubernetes/kubedef"
	"namespacelabs.dev/foundation/framework/kubernetes/kubenaming"
	"namespacelabs.dev/foundation/framework/secrets"
	"namespacelabs.dev/foundation/internal/console"
	"namespacelabs.dev/foundation/internal/fnerrors"
	"namespacelabs.dev/foundation/internal/protos"
	"namespacelabs.dev/foundation/internal/runtime"
	xsecrets "namespacelabs.dev/foundation/internal/secrets"
	"namespacelabs.dev/foundation/internal/support"
	"namespacelabs.dev/foundation/library/kubernetes/rbac"
	"namespacelabs.dev/foundation/schema"
	"namespacelabs.dev/foundation/std/execution/defs"
	"namespacelabs.dev/foundation/std/resources"
	"namespacelabs.dev/foundation/std/runtime/constants"
	"namespacelabs.dev/go-ids"
	"sigs.k8s.io/yaml"
)

//go:embed defaults/*.yaml
var defaults embed.FS

var DeployAsPodsInTests = true

const (
	kubeNode schema.PackageName = "namespacelabs.dev/foundation/std/runtime/kubernetes"

	revisionHistoryLimit int32 = 10

	defaultOneShotDeadline = time.Hour
)

type definitions []definition

type definition interface {
	defs.MakeDefinition
	AppliedResource() any
}

type serverRunState struct {
	operations definitions
}

func getArg(c *applycorev1.ContainerApplyConfiguration, name string) (string, bool) {
	for _, arg := range c.Args {
		if !strings.HasPrefix(arg, "-") {
			continue
		}
		// Remove up to two dashes.
		cleaned := strings.TrimPrefix(strings.TrimPrefix(arg, "-"), "-")
		parts := strings.SplitN(cleaned, "=", 2)
		if len(parts) != 2 {
			continue
		}

		if parts[0] == name {
			return parts[1], true
		}
	}

	return "", false
}

type deployOpts struct {
	secrets   secrets.GroundedSecrets
	platforms []specs.Platform
}

func prepareDeployment(ctx context.Context, target BoundNamespace, deployable runtime.DeployableSpec, opts deployOpts, s *serverRunState) error {
	if deployable.MainContainer.Image.Repository == "" {
		return fnerrors.InternalError("kubernetes: no repository defined in image: %v", deployable.MainContainer.Image)
	}

	secCtx, err := makeSecurityContext(deployable.MainContainer, deployable.Name, console.Debug(ctx))
	if err != nil {
		return err
	}

	reses, err := containerResources(deployable.MainContainer)
	if err != nil {
		return err
	}

	ctrName := kubedef.ServerCtrName(deployable)
	containers := []string{ctrName}
	mainContainer := applycorev1.Container().
		WithName(ctrName).
		WithTerminationMessagePolicy(corev1.TerminationMessageFallbackToLogsOnError).
		WithImage(deployable.MainContainer.Image.RepoAndDigest()).
		WithArgs(deployable.MainContainer.Args...).
		WithCommand(deployable.MainContainer.Command...).
		WithSecurityContext(secCtx).
		WithResources(reses)

	for _, port := range deployable.MainContainer.ContainerPorts {
		p := applycorev1.ContainerPort().WithContainerPort(port.ContainerPort)
		if port.Name != "" {
			p = p.WithName(port.Name)
		}
		if port.HostPort != 0 {
			p = p.WithHostPort(port.HostPort)
		}
		if port.Protocol == schema.Endpoint_Port_UDP {
			p = p.WithProtocol(v1.ProtocolUDP)
		}
		mainContainer.WithPorts(p)
	}

	switch deployable.Attachable {
	case runtime.AttachableKind_WITH_STDIN_ONLY:
		mainContainer = mainContainer.WithStdin(true).WithStdinOnce(true)

	case runtime.AttachableKind_WITH_TTY:
		mainContainer = mainContainer.WithStdin(true).WithStdinOnce(true).WithTTY(true)
	}

	mainEnv := slices.Clone(deployable.MainContainer.Env)

	if deployable.MainContainer.WorkingDir != "" {
		mainContainer = mainContainer.WithWorkingDir(deployable.MainContainer.WorkingDir)
	}

	spec := applycorev1.PodSpec().
		WithEnableServiceLinks(false) // Disable service injection via environment variables.

	if deployable.MainContainer.TerminationGracePeriodSeconds > 0 {
		spec = spec.WithTerminationGracePeriodSeconds(deployable.MainContainer.TerminationGracePeriodSeconds)
	}

	if target.planning != nil {
		nodeSelector := target.planning.DefaultNodeSelector

		for _, c := range target.planning.OverrideNodeSelector {
			if c.DeployablePackageRef != nil && c.DeployablePackageRef.Equals(deployable.PackageRef) {
				nodeSelector = c.NodeSelector
			}
		}

		labelList := map[string]string{}
		for _, v := range nodeSelector {
			labelList[v.Name] = v.Value
		}
		spec = spec.WithNodeSelector(labelList)
	}

	for _, toleration := range deployable.Tolerations {
		t := applycorev1.Toleration().
			WithEffect(v1.TaintEffect(toleration.Effect)).
			WithKey(toleration.Key)

		if toleration.Operator != "" {
			t = t.WithOperator(v1.TolerationOperator(toleration.Operator))
		}

		if toleration.Value != "" {
			t = t.WithValue(toleration.Value)
		}

		spec = spec.WithTolerations(t)
	}

	// nodeSelector := spec.NodeSelector
	nodeSelector := map[string]string{}
	for _, nodeSelectorItem := range deployable.NodeSelector {
		nodeSelector[nodeSelectorItem.Key] = nodeSelectorItem.Value
	}
	spec = spec.WithNodeSelector(nodeSelector)

	// Explicitly allow all pods on all available platforms.
	// On GKE, workloads are not allowed on ARM nodes by default, even if all nodes are ARM.
	// https://cloud.google.com/kubernetes-engine/docs/how-to/prepare-arm-workloads-for-deployment#overview
	// TODO make node affinity configurable.
	var archs []string
	for _, plat := range opts.platforms {
		archs = append(archs, plat.Architecture)
	}

	spec = spec.WithAffinity(applycorev1.Affinity().WithNodeAffinity(
		applycorev1.NodeAffinity().WithRequiredDuringSchedulingIgnoredDuringExecution(
			applycorev1.NodeSelector().WithNodeSelectorTerms(
				applycorev1.NodeSelectorTerm().WithMatchExpressions(
					applycorev1.NodeSelectorRequirement().
						WithKey(kubedef.KubernetesIoArch).
						WithOperator("In").
						WithValues(archs...))))))

	labels := kubedef.MakeLabels(target.env, deployable)
	annotations := kubedef.MakeAnnotations(target.env)
	deploymentId := kubedef.MakeDeploymentId(deployable)

	tmpl := applycorev1.PodTemplateSpec().
		WithAnnotations(annotations).
		WithLabels(labels)

	var initVolumeMounts []*applycorev1.VolumeMountApplyConfiguration
	// Key: PackageRef.CanonicalString().
	initArgs := map[string][]string{}

	var serviceAccount string // May be specified by a SpecExtension.
	var createServiceAccount bool
	var serviceAccountAnnotations []*kubedef.SpecExtension_Annotation

	if len(deployable.Permissions.GetClusterRole()) > 0 {
		createServiceAccount = true
		serviceAccount = kubedef.MakeDeploymentId(deployable)

		for _, role := range deployable.Permissions.ClusterRole {
			var clusterRole *runtime.ComputedResource
			for _, res := range deployable.ComputedResources {
				if res.ResourceInstanceID == role.ResourceId {
					clusterRole = &res
					// XXX check dups?
				}
			}

			if clusterRole == nil {
				return fnerrors.New("permissions refer to %q which is not present", role.Label)
			}

			const roleInstance = "library.kubernetes.rbac.ClusterRoleInstance"
			if clusterRole.InstanceType.GetProtoType() != roleInstance {
				return fnerrors.New("expected resource class to be %q, got %q", roleInstance, clusterRole.InstanceType.GetProtoType())
			}

			instance := &rbac.ClusterRoleInstance{}
			if err := json.Unmarshal(clusterRole.InstanceSerializedJSON, instance); err != nil {
				return fnerrors.New("failed to unmarshal instance %q: %w", roleInstance, err)
			}

			s.operations = append(s.operations, kubedef.Apply{
				Description: fmt.Sprintf("%s: bind Cluster Role", role.Label),
				Resource: rbacv1.ClusterRoleBinding(fmt.Sprintf("binding:%s:%s", kubedef.MakeDeploymentId(deployable), instance.Name)).
					WithLabels(labels).
					WithAnnotations(kubedef.BaseAnnotations()).
					WithRoleRef(rbacv1.RoleRef().
						WithAPIGroup("rbac.authorization.k8s.io").
						WithKind("ClusterRole").
						WithName(instance.Name)).
					WithSubjects(rbacv1.Subject().
						WithKind("ServiceAccount").
						WithNamespace(target.namespace).
						WithName(serviceAccount)),
			})
		}
	}

	var specifiedSec *kubedef.SpecExtension_SecurityContext

	probes := deployable.Probes
	for _, input := range deployable.Extensions {
		specExt := &kubedef.SpecExtension{}
		containerExt := &kubedef.ContainerExtension{}
		initContainerExt := &kubedef.InitContainerExtension{}

		switch {
		case input.Impl.MessageIs(specExt):
			if err := input.Impl.UnmarshalTo(specExt); err != nil {
				return fnerrors.InternalError("failed to unmarshal SpecExtension: %w", err)
			}

			for _, vol := range specExt.Volume {
				k8svol, err := toK8sVol(vol)
				if err != nil {
					return err
				}
				spec = spec.WithVolumes(k8svol)
			}

			for _, toleration := range specExt.Toleration {
				t := applycorev1.Toleration().WithEffect(v1.TaintEffect(toleration.Effect)).WithKey(toleration.Key)

				if toleration.Operator != "" {
					t = t.WithOperator(v1.TolerationOperator(toleration.Operator))
				}

				if toleration.Value != "" {
					t = t.WithValue(toleration.Value)
				}

				spec = spec.WithTolerations(t)
			}

			if len(specExt.Annotation) > 0 {
				m := map[string]string{}
				for _, annotation := range specExt.Annotation {
					m[annotation.Key] = annotation.Value
				}
				tmpl = tmpl.WithAnnotations(m)
			}

			if specExt.EnsureServiceAccount {
				createServiceAccount = true
				if specExt.ServiceAccount == "" {
					return fnerrors.NewWithLocation(deployable.ErrorLocation, "ensure_service_account requires service_account to be set")
				}
			}

			serviceAccountAnnotations = append(serviceAccountAnnotations, specExt.ServiceAccountAnnotation...)

			if specExt.ServiceAccount != "" {
				if serviceAccount != "" && serviceAccount != specExt.ServiceAccount {
					return fnerrors.NewWithLocation(deployable.ErrorLocation, "incompatible service accounts defined, %q vs %q",
						serviceAccount, specExt.ServiceAccount)
				}
				serviceAccount = specExt.ServiceAccount
			}

			if specExt.SecurityContext != nil {
				if !protos.CheckConsolidate(specExt.SecurityContext, &specifiedSec) {
					return fnerrors.NewWithLocation(deployable.ErrorLocation, "incompatible securitycontext defined, %v vs %v",
						specifiedSec, specExt.SecurityContext)
				}
			}

		case input.Impl.MessageIs(containerExt):
			if err := input.Impl.UnmarshalTo(containerExt); err != nil {
				return fnerrors.InternalError("failed to unmarshal ContainerExtension: %w", err)
			}

			for _, mount := range containerExt.VolumeMount {
				volumeMount := applycorev1.VolumeMount().
					WithName(mount.Name).
					WithReadOnly(mount.ReadOnly).
					WithMountPath(mount.MountPath)
				mainContainer = mainContainer.WithVolumeMounts(volumeMount)
				if mount.MountOnInit {
					// Volume mounts may declare to be available also during server initialization.
					// E.g. Initializing the schema of a data store requires early access to server secrets.
					// The volume mount provider has full control over whether the volume is available.
					initVolumeMounts = append(initVolumeMounts, volumeMount)
				}
			}

			var err error
			mainEnv, err = support.MergeEnvs(mainEnv, containerExt.Env)
			if err != nil {
				return err
			}

			if containerExt.Args != nil {
				mainContainer = mainContainer.WithArgs(containerExt.Args...)
			} else {
				// Deprecated path.
				for _, arg := range containerExt.ArgTuple {
					if currentValue, found := getArg(mainContainer, arg.Name); found && currentValue != arg.Value {
						return fnerrors.NewWithLocation(deployable.ErrorLocation, "argument '%s' is already set to '%s' but would be overwritten to '%s' by container extension", arg.Name, currentValue, arg.Value)
					}
					mainContainer = mainContainer.WithArgs(fmt.Sprintf("--%s=%s", arg.Name, arg.Value))
				}
			}

			probes = append(probes, containerExt.Probe...)

			// Deprecated path.
			for _, initContainer := range containerExt.InitContainer {
				key := initContainer.PackageName
				initArgs[key] = append(initArgs[key], initContainer.Arg...)
			}

		case input.Impl.MessageIs(initContainerExt):
			if err := input.Impl.UnmarshalTo(initContainerExt); err != nil {
				return fnerrors.InternalError("failed to unmarshal InitContainerExtension: %w", err)
			}

			var key string
			if initContainerExt.PackageRef != nil {
				key = initContainerExt.PackageRef.Canonical()
			} else {
				key = initContainerExt.PackageName
			}
			initArgs[key] = append(initArgs[key], initContainerExt.Args...)

		default:
			return fnerrors.InternalError("unused startup input: %s", input.Impl.GetTypeUrl())
		}
	}

	if len(deployable.Annotations) > 0 {
		x := runtime.ResolvableSinkMap{}

		if err := runtime.ResolveResolvables(ctx, deployable.RuntimeConfig, nil, deployable.Annotations, &x); err != nil {
			return err
		}

		tmpl = tmpl.WithAnnotations(x)
	}

	var readinessProbe, livenessProbe *schema.Probe
	for _, probe := range probes {
		switch probe.Kind {
		case runtime.FnServiceLivez:
			if !protos.CheckConsolidate(probe, &livenessProbe) {
				return fnerrors.BadInputError("inconsistent live probe definition")
			}
		case runtime.FnServiceReadyz:
			if !protos.CheckConsolidate(probe, &readinessProbe) {
				return fnerrors.BadInputError("inconsistent ready probe definition")
			}
		default:
			return fnerrors.BadInputError("%s: unknown probe kind", probe.Kind)
		}
	}

	probevalues := perEnvConfMapping[target.env.GetPurpose()]
	if readinessProbe != nil || livenessProbe != nil {
		if probevalues == nil {
			return fnerrors.InternalError("%s: no constants configured", target.env.GetPurpose())
		}
	}

	if readinessProbe != nil {
		probe, err := toK8sProbe(applycorev1.Probe().WithInitialDelaySeconds(probevalues.readinessInitialDelay),
			probevalues, readinessProbe)
		if err != nil {
			return err
		}
		mainContainer = mainContainer.WithReadinessProbe(probe)
	}

	if livenessProbe != nil {
		probe, err := toK8sProbe(applycorev1.Probe().WithInitialDelaySeconds(probevalues.livenessInitialDelay),
			probevalues, livenessProbe)
		if err != nil {
			return err
		}
		mainContainer = mainContainer.WithLivenessProbe(probe)
	}

	// XXX Think through this, should it also be hashed + immutable?
	secretId := fmt.Sprintf("ns-managed-%s-%s", deployable.Name, deployable.GetId())
	seccol := newSecretCollector(opts.secrets, secretId)

	ensure := kubedef.EnsureDeployment{
		Deployable:    deployable,
		InhibitEvents: deployable.Class == schema.DeployableClass_MANUAL || (target.namespace == kubedef.AdminNamespace && !deployable.Focused),
		SchedCategory: []string{
			runtime.DeployableCategory(deployable),
			runtime.OwnedByDeployable(deployable),
		},
		SetContainerFields: slices.Clone(deployable.SetContainerField),
	}

	if _, err := fillEnv(ctx, deployable.RuntimeConfig, mainContainer, mainEnv, seccol, &ensure); err != nil {
		return err
	}

	volumes := deployable.Volumes
	mounts := deployable.MainContainer.Mounts

	var volumeTemplates []*applycorev1.PersistentVolumeClaimApplyConfiguration

	volumeDefs := map[string]*volumeDef{}
	for k, volume := range volumes {
		if volume.Name == "" {
			return fnerrors.InternalError("volume #%d is missing a name", k)
		}

		name := kubedef.MakeVolumeName(volume)
		volumeDef := &volumeDef{name: name}
		volumeDefs[volume.Name] = volumeDef

		switch volume.Kind {
		case constants.VolumeKindEphemeral:
			ev := &schema.EphemeralVolume{}
			if err := volume.Definition.UnmarshalTo(ev); err != nil {
				return fnerrors.InternalError("%s: failed to unmarshal hostDir volume definition: %w", volume.Name, err)
			}

			emptydir := applycorev1.EmptyDirVolumeSource()
			if ev.SizeBytes > 0 {
				quantity := resource.NewScaledQuantity(int64(ev.SizeBytes), 0)
				emptydir = emptydir.WithSizeLimit(*quantity)
			}

			spec = spec.WithVolumes(applycorev1.Volume().WithName(name).WithEmptyDir(emptydir))

		case constants.VolumeKindHostPath:
			pv := &schema.HostPathVolume{}
			if err := volume.Definition.UnmarshalTo(pv); err != nil {
				return fnerrors.InternalError("%s: failed to unmarshal hostDir volume definition: %w", volume.Name, err)
			}

			switch {
			case pv.Directory != "":
				spec = spec.WithVolumes(applycorev1.Volume().WithName(name).
					WithHostPath(applycorev1.HostPathVolumeSource().
						WithType(v1.HostPathDirectoryOrCreate).WithPath(pv.Directory)))
			}

		case constants.VolumeKindPersistent:
			pv := &schema.PersistentVolume{}
			if err := volume.Definition.UnmarshalTo(pv); err != nil {
				return fnerrors.InternalError("%s: failed to unmarshal persistent volume definition: %w", volume.Name, err)
			}

			if pv.Template {
				if deployable.Class != schema.DeployableClass_STATEFUL {
					return fnerrors.InternalError("volume %q is a template, but the server is not stateful", volume.Name)
				}
			}

			if pv.Id == "" {
				return fnerrors.BadInputError("%s: persistent ID is missing", volume.Name)
			}

			v, pvc, err := makePersistentVolume(target.namespace, target.env, deployable.ErrorLocation, volume.Owner, name, pv.Id, pv.SizeBytes, pv.Template, annotations)
			if err != nil {
				return err
			}

			if v != nil {
				spec = spec.WithVolumes(v)
			}

			if pvc != nil {
				if pv.Template {
					volumeTemplates = append(volumeTemplates, pvc)
				} else {
					// spec = spec.WithVolumes(v)
					s.operations = append(s.operations, kubedef.Apply{
						Description: fmt.Sprintf("Persistent storage for %s (%s)", volume.Owner, humanize.Bytes(pv.SizeBytes)),
						Resource:    pvc,
					})
				}
			}

		case constants.VolumeKindWorkspaceSync:
			volumeDef.isWorkspaceSync = true

		case constants.VolumeKindConfigurable:
			cv := &schema.ConfigurableVolume{}
			if err := volume.Definition.UnmarshalTo(cv); err != nil {
				return fnerrors.InternalError("%s: failed to unmarshal configurable volume definition: %w", volume.Name, err)
			}

			configs := newDataItemCollector()

			configHash := sha256.New()

			projected := applycorev1.ProjectedVolumeSource()

			var configmapItems []*applycorev1.KeyToPathApplyConfiguration
			var secretItems []*applycorev1.KeyToPathApplyConfiguration
			for _, entry := range cv.Entries {
				switch {
				case entry.Inline != nil:
					configmapItems = append(configmapItems, makeConfigEntry(configHash, entry, entry.Inline, configs).WithPath(entry.Path))

				case entry.InlineSet != nil:
					for _, rsc := range entry.InlineSet.Resource {
						configmapItems = append(configmapItems, makeConfigEntry(configHash, entry, rsc, configs).WithPath(filepath.Join(entry.Path, rsc.Path)))
					}

				case entry.SecretRef != nil:
					resource, err := opts.secrets.Get(ctx, entry.SecretRef)
					if err != nil {
						return err
					}

					if resource.FileContents != nil {
						secretItems = append(secretItems, makeConfigEntry(configHash, entry, resource.FileContents, seccol.items).WithPath(entry.Path))
					} else if resource.Spec.Generate != nil {
						alloc := seccol.allocateGenerated(resource.Ref, resource.Spec)
						projected = projected.WithSources(applycorev1.VolumeProjection().WithSecret(
							applycorev1.SecretProjection().WithName(alloc.Name).WithItems(
								applycorev1.KeyToPath().WithKey(alloc.Key).WithPath(entry.Path)),
						))
					}

				case entry.KubernetesSecretRef != nil:
					projected = projected.WithSources(applycorev1.VolumeProjection().WithSecret(
						applycorev1.SecretProjection().WithName(entry.KubernetesSecretRef.SecretName).
							WithItems(applycorev1.KeyToPath().WithKey(entry.KubernetesSecretRef.SecretKey).WithPath(entry.Path)),
					))
				}
			}

			if len(configmapItems) > 0 {
				configId := "ns-static-" + ids.EncodeToBase32String(configHash.Sum(nil))

				projected = projected.WithSources(applycorev1.VolumeProjection().WithConfigMap(
					applycorev1.ConfigMapProjection().WithName(configId).WithItems(configmapItems...)))

				// Needs to be declared before it's used.
				s.operations = append(s.operations, kubedef.Apply{
					Description: "Static configuration",
					Resource: applycorev1.ConfigMap(configId, target.namespace).
						WithAnnotations(annotations).
						WithLabels(labels).
						WithLabels(map[string]string{
							kubedef.K8sKind: kubedef.K8sStaticConfigKind,
						}).
						WithData(configs.data).
						WithBinaryData(configs.binaryData),
				})
			}

			if len(secretItems) > 0 {
				projected = projected.WithSources(applycorev1.VolumeProjection().WithSecret(
					applycorev1.SecretProjection().WithName(secretId).WithItems(secretItems...)))
			}

			spec = spec.WithVolumes(applycorev1.Volume().WithName(name).WithProjected(projected))

		default:
			return fnerrors.InternalError("%s: unsupported volume type", volume.Kind)
		}
	}

	for k, mount := range mounts {
		if mount.Path == "" {
			return fnerrors.InternalError("mount #%d is missing a path", k)
		}

		if mount.VolumeRef == nil {
			return fnerrors.InternalError("mount %q is missing a target volume", mount.Path)
		}

		volumeDef, ok := volumeDefs[mount.VolumeRef.Name]
		if !ok {
			return fnerrors.InternalError("unknown target volume %q for mount %q", mount.VolumeRef.Name, mount.Path)
		}

		if !volumeDef.isWorkspaceSync {
			mainContainer = mainContainer.WithVolumeMounts(applycorev1.VolumeMount().
				WithMountPath(mount.Path).
				WithName(volumeDef.name).
				WithReadOnly(mount.Readonly))
		}
	}

	regularResources := slices.Clone(deployable.ResourceDeps)

	const projectedSecretsVolName = "ns-projected-secrets"

	var secretProjections []*applycorev1.SecretProjectionApplyConfiguration
	var injected []*kubedef.OpEnsureRuntimeConfig_InjectedResource
	for _, res := range deployable.SecretResources {
		alloc, err := seccol.Allocate(ctx, res.SecretRef)
		if err != nil {
			return err
		}

		serialized, err := xsecrets.Serialize(res.ResourceRef)
		if err != nil {
			return err
		}

		secretProjections = append(secretProjections,
			applycorev1.SecretProjection().WithName(alloc.Name).WithItems(
				applycorev1.KeyToPath().WithKey(alloc.Key).WithPath(serialized.RelPath)))

		injected = append(injected, &kubedef.OpEnsureRuntimeConfig_InjectedResource{
			ResourceRef:    res.ResourceRef,
			SerializedJson: serialized.JSON,
		})
	}

	for _, dep := range deployable.PlannedResourceDeps {
		handled := false
		for _, x := range deployable.ComputedResources {
			if dep.ResourceInstanceId == x.ResourceInstanceID {
				injected = append(injected, &kubedef.OpEnsureRuntimeConfig_InjectedResource{
					ResourceRef:    dep.ResourceRef,
					SerializedJson: x.InstanceSerializedJSON,
				})
				handled = true
				break
			}
		}

		if !handled {
			return fnerrors.InternalError("%s: resource value is missing", dep.ResourceRef.Canonical())
		}
	}

	if len(secretProjections) > 0 {
		src := applycorev1.ProjectedVolumeSource()
		for _, p := range secretProjections {
			src = src.WithSources(applycorev1.VolumeProjection().WithSecret(p))
		}

		spec = spec.WithVolumes(applycorev1.Volume().WithName(projectedSecretsVolName).WithProjected(src))
		mainContainer = mainContainer.WithVolumeMounts(applycorev1.VolumeMount().
			WithName(projectedSecretsVolName).
			WithMountPath(xsecrets.SecretBaseMountPath).
			WithReadOnly(true))
	}

	// Before sidecars so they have access to the "runtime config" volume.
	if deployable.RuntimeConfig != nil || len(regularResources) > 0 || len(injected) > 0 {
		slices.SortFunc(regularResources, func(a, b *resources.ResourceDependency) int {
			return strings.Compare(a.ResourceInstanceId, b.ResourceInstanceId)
		})

		ensureConfig := kubedef.EnsureRuntimeConfig{
			Description:          "Runtime configuration",
			RuntimeConfig:        deployable.RuntimeConfig,
			Deployable:           deployable,
			ResourceDependencies: regularResources,
			InjectedResources:    injected,
			BuildVCS:             deployable.BuildVCS,
			PersistConfiguration: deployable.MountRuntimeConfigPath != "",
		}

		s.operations = append(s.operations, ensureConfig)

		// Make sure we wait for the runtime configuration to be created before
		// deploying a new deployment or statefulset.
		ensure.RuntimeConfigDependency = kubedef.RuntimeConfigOutput(deployable)

		if deployable.MountRuntimeConfigPath != "" {
			ensure.ConfigurationVolumeName = "namespace-rtconfig"
			mainContainer = mainContainer.WithVolumeMounts(
				applycorev1.VolumeMount().
					WithMountPath(deployable.MountRuntimeConfigPath).
					WithName(ensure.ConfigurationVolumeName).
					WithReadOnly(true))
		}
	}

	for _, sidecar := range deployable.Sidecars {
		if sidecar.Name == "" {
			return fnerrors.InternalError("sidecar name is missing")
		}

		name := sidecarName(sidecar, "sidecar")
		for _, c := range containers {
			if name == c {
				return fnerrors.NewWithLocation(deployable.ErrorLocation, "duplicate sidecar container name: %s", name)
			}
		}

		containers = append(containers, name)

		sidecarSecCtx, err := makeSecurityContext(sidecar.ContainerRunOpts, sidecar.Name, console.Debug(ctx))
		if err != nil {
			return fnerrors.AttachLocation(deployable.ErrorLocation, err)
		}

		reses, err := containerResources(deployable.MainContainer)
		if err != nil {
			return fnerrors.AttachLocation(deployable.ErrorLocation, err)
		}

		scntr := applycorev1.Container().
			WithName(name).
			WithTerminationMessagePolicy(corev1.TerminationMessageFallbackToLogsOnError).
			WithImage(sidecar.Image.RepoAndDigest()).
			WithArgs(sidecar.Args...).
			WithCommand(sidecar.Command...).
			WithSecurityContext(sidecarSecCtx).
			WithResources(reses)

		// XXX remove this
		scntr = scntr.WithEnv(
			applycorev1.EnvVar().WithName("FN_KUBERNETES_NAMESPACE").WithValue(target.namespace),
			applycorev1.EnvVar().WithName("FN_SERVER_ID").WithValue(deployable.GetId()),
			applycorev1.EnvVar().WithName("FN_SERVER_NAME").WithValue(deployable.Name),
		)

		if _, err := fillEnv(ctx, deployable.RuntimeConfig, scntr, sidecar.Env, seccol, &ensure); err != nil {
			return err
		}

		// Share all mounts with sidecards for now.
		// XXX security review this.
		scntr.VolumeMounts = mainContainer.VolumeMounts
		spec.WithContainers(scntr)
	}

	for _, init := range deployable.Inits {
		if init.Name == "" {
			return fnerrors.InternalError("sidecar name is missing")
		}

		name := sidecarName(init, "init")
		for _, c := range containers {
			if name == c {
				return fnerrors.NewWithLocation(deployable.ErrorLocation, "duplicate init container name: %s", name)
			}
		}

		containers = append(containers, name)

		reses, err := containerResources(deployable.MainContainer)
		if err != nil {
			return fnerrors.AttachLocation(deployable.ErrorLocation, err)
		}

		scntr := applycorev1.Container().
			WithName(name).
			WithTerminationMessagePolicy(corev1.TerminationMessageFallbackToLogsOnError).
			WithImage(init.Image.RepoAndDigest()).
			WithArgs(append(init.Args, initArgs[init.BinaryRef.Canonical()]...)...).
			WithCommand(init.Command...).
			WithVolumeMounts(initVolumeMounts...).
			WithResources(reses)

		if _, err := fillEnv(ctx, deployable.RuntimeConfig, scntr, init.Env, seccol, &ensure); err != nil {
			return err
		}

		spec.WithInitContainers(scntr)
	}

	s.operations = append(s.operations, seccol.planDeployment(target.namespace, annotations, labels)...)

	podSecCtx := applycorev1.PodSecurityContext()
	if specifiedSec == nil {
		toparse := map[string]any{
			"defaults/pod.podsecuritycontext.yaml": podSecCtx,
		}

		for path, obj := range toparse {
			contents, err := fs.ReadFile(defaults, path)
			if err != nil {
				return fnerrors.InternalError("internal kubernetes data failed to read: %w", err)
			}

			if err := yaml.Unmarshal(contents, obj); err != nil {
				return fnerrors.InternalError("%s: failed to parse defaults: %w", path, err)
			}
		}
	} else {
		if specifiedSec.RunAsUser != 0 {
			podSecCtx = podSecCtx.WithRunAsUser(specifiedSec.RunAsUser)
		}
		if specifiedSec.RunAsGroup != 0 {
			podSecCtx = podSecCtx.WithRunAsGroup(specifiedSec.RunAsGroup)
		}
		if specifiedSec.FsGroup != 0 {
			podSecCtx = podSecCtx.WithFSGroup(specifiedSec.FsGroup)
		}
	}

	if _, err := runAsToPodSecCtx(podSecCtx, deployable.MainContainer.RunAs); err != nil {
		return fnerrors.AttachLocation(deployable.ErrorLocation, err)
	}

	if deployable.MainContainer.HostNetwork {
		spec = spec.WithHostNetwork(true).WithDNSPolicy(corev1.DNSClusterFirstWithHostNet)
	}

	spec = spec.
		WithSecurityContext(podSecCtx).
		WithContainers(mainContainer).
		WithAutomountServiceAccountToken(serviceAccount != "")

	if serviceAccount != "" {
		spec = spec.WithServiceAccountName(serviceAccount)
	}

	tmpl = tmpl.WithSpec(spec)

	if createServiceAccount {
		annotations := map[string]string{}
		for _, ann := range serviceAccountAnnotations {
			annotations[ann.Key] = ann.Value
		}

		s.operations = append(s.operations, kubedef.Apply{
			Description: "Service Account",
			Resource: applycorev1.ServiceAccount(serviceAccount, target.namespace).
				WithLabels(labels).
				WithAnnotations(annotations),
		})
	} else {
		if len(serviceAccountAnnotations) > 0 {
			return fnerrors.NewWithLocation(deployable.ErrorLocation, "can't set service account annotations without ensure_service_account")
		}
	}

	// We don't deploy managed deployments or statefulsets in tests, as these are one-shot
	// servers which we want to control a bit more carefully. For example, we want to deploy
	// them with restart_policy=never, which we would otherwise not be able to do with
	// deployments.
	if deployAsPods(target.env) || isOneShotLike(deployable.Class) {
		desc := fmt.Sprintf("Server %s", deployable.Name)
		if isOneShotLike(deployable.Class) {
			desc = fmt.Sprintf("One-shot %s", deployable.Name)

			// TODO make this configurable
			tmpl.Spec = tmpl.Spec.WithActiveDeadlineSeconds(int64(defaultOneShotDeadline.Seconds()))
		}

		ensure.Description = firstStr(deployable.Description, desc)
		ensure.Resource = applycorev1.Pod(deploymentId, target.namespace).
			WithAnnotations(annotations).
			WithAnnotations(tmpl.Annotations).
			WithLabels(labels).
			WithLabels(tmpl.Labels).
			WithSpec(tmpl.Spec.WithRestartPolicy(corev1.RestartPolicyNever))
	} else {
		replicas := int32(1)
		if deployable.Replicas > 0 {
			replicas = deployable.Replicas
		}

		switch deployable.Class {
		case schema.DeployableClass_STATELESS:
			ensure.Description = firstStr(deployable.Description, fmt.Sprintf("Server Deployment %s", deployable.Name))
			deployment := appsv1.
				Deployment(deploymentId, target.namespace).
				WithAnnotations(annotations).
				WithLabels(labels).
				WithSpec(appsv1.DeploymentSpec().
					WithReplicas(replicas).
					WithRevisionHistoryLimit(revisionHistoryLimit).
					WithTemplate(tmpl).
					WithSelector(applymetav1.LabelSelector().WithMatchLabels(kubedef.SelectById(deployable))))
			if deployable.ConfigImage != nil {
				deployment.WithAnnotations(map[string]string{
					kubedef.K8sConfigImage: deployable.ConfigImage.RepoAndDigest(),
				})
			}

			ensure.Resource = deployment

		case schema.DeployableClass_STATEFUL:
			ensure.Description = firstStr(deployable.Description, fmt.Sprintf("Server StatefulSet %s", deployable.Name))
			statefulSet := appsv1.
				StatefulSet(deploymentId, target.namespace).
				WithAnnotations(annotations).
				WithLabels(labels).
				WithSpec(appsv1.StatefulSetSpec().
					WithReplicas(replicas).
					WithRevisionHistoryLimit(revisionHistoryLimit).
					WithTemplate(tmpl).
					WithVolumeClaimTemplates(volumeTemplates...).
					WithSelector(applymetav1.LabelSelector().WithMatchLabels(kubedef.SelectById(deployable))))
			if deployable.ConfigImage != nil {
				statefulSet.WithAnnotations(map[string]string{
					kubedef.K8sConfigImage: deployable.ConfigImage.RepoAndDigest(),
				})
			}

			ensure.Resource = statefulSet

		case schema.DeployableClass_DAEMONSET:
			ensure.Description = firstStr(deployable.Description, fmt.Sprintf("Server DaemonSet %s", deployable.Name))
			deployment := appsv1.
				DaemonSet(deploymentId, target.namespace).
				WithAnnotations(annotations).
				WithLabels(labels).
				WithSpec(appsv1.DaemonSetSpec().
					WithRevisionHistoryLimit(revisionHistoryLimit).
					WithTemplate(tmpl).
					WithSelector(applymetav1.LabelSelector().WithMatchLabels(kubedef.SelectById(deployable))))
			if deployable.ConfigImage != nil {
				deployment.WithAnnotations(map[string]string{
					kubedef.K8sConfigImage: deployable.ConfigImage.RepoAndDigest(),
				})
			}

			ensure.Resource = deployment

		default:
			return fnerrors.InternalError("%s: unsupported deployable class", deployable.Class)
		}
	}

	s.operations = append(s.operations, ensure)
	return nil
}

func containerResources(ctr runtime.ContainerRunOpts) (*applycorev1.ResourceRequirementsApplyConfiguration, error) {
	if ctr.ResourceLimits == nil && ctr.ResourceRequests == nil {
		return nil, nil
	}

	r := applycorev1.ResourceRequirements()

	rl, err := parseResourceList(ctr.ResourceLimits)
	if err != nil {
		return nil, err
	} else if rl != nil {
		r = r.WithLimits(rl)
	}

	reqs, err := parseResourceList(ctr.ResourceRequests)
	if err != nil {
		return nil, err
	} else if reqs != nil {
		r = r.WithRequests(reqs)
	}

	return r, nil
}

func parseResourceList(limits *schema.Container_ResourceLimits) (corev1.ResourceList, error) {
	if limits == nil || (limits.Cpu == "" && limits.Memory == "") {
		return nil, nil
	}

	rl := corev1.ResourceList{}
	if limits.Cpu != "" {
		x, err := resource.ParseQuantity(limits.Cpu)
		if err != nil {
			return nil, err
		}
		rl[corev1.ResourceCPU] = x
	}

	if limits.Memory != "" {
		x, err := resource.ParseQuantity(limits.Memory)
		if err != nil {
			return nil, err
		}
		rl[corev1.ResourceMemory] = x
	}

	return rl, nil
}

func firstStr(strs ...string) string {
	for _, str := range strs {
		if str != "" {
			return str
		}
	}
	return ""
}

func isOneShotLike(class schema.DeployableClass) bool {
	return class == schema.DeployableClass_ONESHOT || class == schema.DeployableClass_MANUAL
}

type collector struct {
	data       map[string]string
	binaryData map[string][]byte
}

func newDataItemCollector() *collector {
	return &collector{data: map[string]string{}, binaryData: map[string][]byte{}}
}

func (cm *collector) set(key string, rsc []byte) {
	cm.binaryData[key] = rsc
}

func makeConfigEntry(hash io.Writer, entry *schema.ConfigurableVolume_Entry, rsc *schema.FileContents, cm *collector) *applycorev1.KeyToPathApplyConfiguration {
	key := kubenaming.DomainFragLike(entry.Path, rsc.Path)
	fmt.Fprintf(hash, "%s:", key)
	_, _ = hash.Write(rsc.Contents)
	fmt.Fprintln(hash)
	cm.set(key, rsc.Contents)

	return applycorev1.KeyToPath().WithKey(key)
}

func sidecarName(o runtime.SidecarRunOpts, prefix string) string {
	return fmt.Sprintf("%s-%s", prefix, o.Name)
}

func deployEndpoint(ctx context.Context, r BoundNamespace, deployable runtime.DeployableSpec, endpoint *schema.Endpoint, s *serverRunState) error {
	serviceSpec := applycorev1.ServiceSpec().WithSelector(kubedef.SelectById(deployable))

	var applied []*applycorev1.ServicePortApplyConfiguration
	for _, port := range endpoint.Ports {
		if port.Port == nil {
			continue
		}

		p := applycorev1.ServicePort().
			WithName(port.Port.Name).
			WithPort(port.ExportedPort).
			WithTargetPort(intstr.FromInt(int(port.Port.ContainerPort)))

		if port.Port.Protocol == schema.Endpoint_Port_UDP {
			p = p.WithProtocol(corev1.ProtocolUDP)
		} else {
			p = p.WithProtocol(corev1.ProtocolTCP)
		}

		applied = append(applied, p)
	}

	if len(applied) == 0 {
		return nil
	}

	serviceSpec = serviceSpec.WithPorts(applied...)

	if endpoint.Type == schema.Endpoint_LOAD_BALANCER {
		serviceSpec = serviceSpec.WithType(corev1.ServiceTypeLoadBalancer)
	}

	serviceAnnotations, err := kubedef.MakeServiceAnnotations(endpoint)
	if err != nil {
		return err
	}

	for _, md := range endpoint.ServiceMetadata {
		x := &schema.ServiceAnnotations{}
		if md.Details.MessageIs(x) {
			if err := md.Details.UnmarshalTo(x); err != nil {
				return fnerrors.InternalError("failed to unmarshal ServiceAnnotations: %w", err)
			}
			for _, kv := range x.KeyValue {
				serviceAnnotations[kv.Key] = kv.Value
			}
		}
	}

	if err := validateServiceName(endpoint.AllocatedName); err != nil {
		return fnerrors.AttachLocation(deployable.ErrorLocation, err)
	}

	s.operations = append(s.operations, kubedef.Apply{
		Description: fmt.Sprintf("Service %s:%s", deployable.Name, endpoint.ServiceName),
		Resource: applycorev1.
			Service(endpoint.AllocatedName, r.namespace).
			WithLabels(kubedef.MakeServiceLabels(r.env, deployable, endpoint)).
			WithAnnotations(serviceAnnotations).
			WithSpec(serviceSpec),
		SchedCategory: []string{
			runtime.OwnedByDeployable(deployable),
			kubedef.MakeServicesCat(deployable),
		},
		SchedAfterCategory: []string{
			runtime.DeployableCategory(deployable),
		},
	})

	return nil
}