feat: runtime manager

2026-04-28 20:39:18 +02:00
parent e0a99b346b
commit a7cee15115
289 changed files with 45660 additions and 2207 deletions
@@ -0,0 +1,612 @@
+// Package stopruntime implements the `stop` lifecycle operation owned by
+// Runtime Manager. The service is the single orchestrator behind both
+// the asynchronous `runtime:stop_jobs` consumer and the synchronous
+// `POST /api/v1/internal/runtimes/{game_id}/stop` REST handler. It is
+// also the inner stop step of the restart and patch services, which
+// call Run while holding the outer per-game lease.
+//
+// Lifecycle and failure-mode semantics follow `rtmanager/README.md
+// §Lifecycles → Stop`. Design rationale is captured in
+// `rtmanager/docs/services.md`.
+package stopruntime
+
+import (
+	"context"
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"strings"
+	"time"
+
+	"galaxy/rtmanager/internal/config"
+	"galaxy/rtmanager/internal/domain/health"
+	"galaxy/rtmanager/internal/domain/operation"
+	"galaxy/rtmanager/internal/domain/runtime"
+	"galaxy/rtmanager/internal/logging"
+	"galaxy/rtmanager/internal/ports"
+	"galaxy/rtmanager/internal/service/startruntime"
+	"galaxy/rtmanager/internal/telemetry"
+)
+
+// leaseReleaseTimeout bounds the deferred lease-release call. A fresh
+// background context is used so the release runs even when the request
+// context was already canceled.
+const leaseReleaseTimeout = 5 * time.Second
+
+// Input stores the per-call arguments for one stop operation.
+type Input struct {
+	// GameID identifies the platform game to stop.
+	GameID string
+
+	// Reason classifies the trigger of the stop. Required.
+	Reason StopReason
+
+	// OpSource classifies how the request entered Runtime Manager.
+	// Required: every operation_log entry carries an op_source.
+	OpSource operation.OpSource
+
+	// SourceRef stores the optional opaque per-source reference (Redis
+	// Stream entry id, REST request id, admin user id). Empty when the
+	// caller does not provide one. For inner calls invoked by the
+	// restart and patch orchestrators it carries the outer correlation
+	// id so the three operation_log entries share it.
+	SourceRef string
+}
+
+// Validate reports whether input carries the structural invariants the
+// service requires.
+func (input Input) Validate() error {
+	if strings.TrimSpace(input.GameID) == "" {
+		return fmt.Errorf("game id must not be empty")
+	}
+	if !input.OpSource.IsKnown() {
+		return fmt.Errorf("op source %q is unsupported", input.OpSource)
+	}
+	if err := input.Reason.Validate(); err != nil {
+		return err
+	}
+	return nil
+}
+
+// Result stores the deterministic outcome of one Handle / Run call.
+type Result struct {
+	// Record carries the runtime record installed by the operation.
+	// Populated on success and on idempotent replay; zero on failure.
+	Record runtime.RuntimeRecord
+
+	// Outcome reports whether the operation completed (success) or
+	// produced a stable failure code.
+	Outcome operation.Outcome
+
+	// ErrorCode stores the stable error code on failure, or
+	// `replay_no_op` on idempotent replay. Empty for fresh successes.
+	ErrorCode string
+
+	// ErrorMessage stores the operator-readable detail on failure.
+	// Empty for successes.
+	ErrorMessage string
+}
+
+// Dependencies groups the collaborators required by Service.
+type Dependencies struct {
+	// RuntimeRecords reads and updates the durable runtime record.
+	RuntimeRecords ports.RuntimeRecordStore
+
+	// OperationLogs records the success / failure audit entry.
+	OperationLogs ports.OperationLogStore
+
+	// Docker drives the Docker daemon (container stop).
+	Docker ports.DockerClient
+
+	// Leases serialises operations against the same game id.
+	Leases ports.GameLeaseStore
+
+	// HealthEvents publishes `runtime:health_events` and upserts the
+	// matching `health_snapshots` row. Used on the vanished-container
+	// path to emit `container_disappeared`.
+	HealthEvents ports.HealthEventPublisher
+
+	// Container groups the per-container settings consumed at stop time
+	// (the graceful stop timeout).
+	Container config.ContainerConfig
+
+	// Coordination supplies the per-game lease TTL.
+	Coordination config.CoordinationConfig
+
+	// Telemetry records stop outcomes and lease latency. Required.
+	Telemetry *telemetry.Runtime
+
+	// Logger records structured service-level events. Defaults to
+	// `slog.Default()` when nil.
+	Logger *slog.Logger
+
+	// Clock supplies the wall-clock used for operation timestamps.
+	// Defaults to `time.Now` when nil.
+	Clock func() time.Time
+
+	// NewToken supplies a unique opaque lease token. Defaults to a
+	// 32-byte random base64url string when nil. Tests may override.
+	NewToken func() string
+}
+
+// Service executes the stop lifecycle operation.
+type Service struct {
+	runtimeRecords ports.RuntimeRecordStore
+	operationLogs  ports.OperationLogStore
+	docker         ports.DockerClient
+	leases         ports.GameLeaseStore
+	healthEvents   ports.HealthEventPublisher
+
+	stopTimeout time.Duration
+	leaseTTL    time.Duration
+
+	telemetry *telemetry.Runtime
+	logger    *slog.Logger
+
+	clock    func() time.Time
+	newToken func() string
+}
+
+// NewService constructs one Service from deps.
+func NewService(deps Dependencies) (*Service, error) {
+	switch {
+	case deps.RuntimeRecords == nil:
+		return nil, errors.New("new stop runtime service: nil runtime records")
+	case deps.OperationLogs == nil:
+		return nil, errors.New("new stop runtime service: nil operation logs")
+	case deps.Docker == nil:
+		return nil, errors.New("new stop runtime service: nil docker client")
+	case deps.Leases == nil:
+		return nil, errors.New("new stop runtime service: nil lease store")
+	case deps.HealthEvents == nil:
+		return nil, errors.New("new stop runtime service: nil health events publisher")
+	case deps.Telemetry == nil:
+		return nil, errors.New("new stop runtime service: nil telemetry runtime")
+	}
+	if err := deps.Container.Validate(); err != nil {
+		return nil, fmt.Errorf("new stop runtime service: container config: %w", err)
+	}
+	if err := deps.Coordination.Validate(); err != nil {
+		return nil, fmt.Errorf("new stop runtime service: coordination config: %w", err)
+	}
+
+	clock := deps.Clock
+	if clock == nil {
+		clock = time.Now
+	}
+	logger := deps.Logger
+	if logger == nil {
+		logger = slog.Default()
+	}
+	logger = logger.With("service", "rtmanager.stopruntime")
+
+	newToken := deps.NewToken
+	if newToken == nil {
+		newToken = defaultTokenGenerator()
+	}
+
+	return &Service{
+		runtimeRecords: deps.RuntimeRecords,
+		operationLogs:  deps.OperationLogs,
+		docker:         deps.Docker,
+		leases:         deps.Leases,
+		healthEvents:   deps.HealthEvents,
+		stopTimeout:    deps.Container.StopTimeout,
+		leaseTTL:       deps.Coordination.GameLeaseTTL,
+		telemetry:      deps.Telemetry,
+		logger:         logger,
+		clock:          clock,
+		newToken:       newToken,
+	}, nil
+}
+
+// Handle executes one stop operation end-to-end. The Go-level error
+// return is reserved for non-business failures (nil context, nil
+// receiver). Every business outcome — success, idempotent replay, or
+// any of the stable failure modes — flows through Result.
+func (service *Service) Handle(ctx context.Context, input Input) (Result, error) {
+	if service == nil {
+		return Result{}, errors.New("stop runtime: nil service")
+	}
+	if ctx == nil {
+		return Result{}, errors.New("stop runtime: nil context")
+	}
+
+	opStartedAt := service.clock().UTC()
+
+	if err := input.Validate(); err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInvalidRequest,
+			errorMessage: err.Error(),
+		}), nil
+	}
+
+	token := service.newToken()
+	leaseStart := service.clock()
+	acquired, err := service.leases.TryAcquire(ctx, input.GameID, token, service.leaseTTL)
+	service.telemetry.RecordLeaseAcquireLatency(ctx, service.clock().Sub(leaseStart))
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeServiceUnavailable,
+			errorMessage: fmt.Sprintf("acquire game lease: %s", err.Error()),
+		}), nil
+	}
+	if !acquired {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeConflict,
+			errorMessage: "another lifecycle operation is in progress for this game",
+		}), nil
+	}
+	defer service.releaseLease(ctx, input.GameID, token)
+
+	return service.runUnderLease(ctx, input, opStartedAt)
+}
+
+// Run executes the stop lifecycle assuming the per-game lease is
+// already held by the caller. The method is reserved for orchestrator
+// services in `internal/service/` that compose stop with another
+// operation under a single outer lease (restart and patch). External
+// callers must use Handle.
+func (service *Service) Run(ctx context.Context, input Input) (Result, error) {
+	if service == nil {
+		return Result{}, errors.New("stop runtime: nil service")
+	}
+	if ctx == nil {
+		return Result{}, errors.New("stop runtime: nil context")
+	}
+
+	opStartedAt := service.clock().UTC()
+
+	if err := input.Validate(); err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInvalidRequest,
+			errorMessage: err.Error(),
+		}), nil
+	}
+
+	return service.runUnderLease(ctx, input, opStartedAt)
+}
+
+// runUnderLease executes the post-validation, lease-protected stop
+// steps shared by Handle and Run.
+func (service *Service) runUnderLease(ctx context.Context, input Input, opStartedAt time.Time) (Result, error) {
+	existing, err := service.runtimeRecords.Get(ctx, input.GameID)
+	if errors.Is(err, runtime.ErrNotFound) {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeNotFound,
+			errorMessage: fmt.Sprintf("runtime record for game %q does not exist", input.GameID),
+		}), nil
+	}
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("load runtime record: %s", err.Error()),
+		}), nil
+	}
+
+	switch existing.Status {
+	case runtime.StatusStopped, runtime.StatusRemoved:
+		return service.recordReplayNoOp(ctx, opStartedAt, input, existing), nil
+	case runtime.StatusRunning:
+		// proceed
+	default:
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("runtime record has unsupported status %q", existing.Status),
+		}), nil
+	}
+
+	if err := service.docker.Stop(ctx, existing.CurrentContainerID, service.stopTimeout); err != nil {
+		if errors.Is(err, ports.ErrContainerNotFound) {
+			return service.handleVanished(ctx, input, opStartedAt, existing), nil
+		}
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeServiceUnavailable,
+			errorMessage: fmt.Sprintf("docker stop: %s", err.Error()),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	}
+
+	updateNow := service.clock().UTC()
+	err = service.runtimeRecords.UpdateStatus(ctx, ports.UpdateStatusInput{
+		GameID:              input.GameID,
+		ExpectedFrom:        runtime.StatusRunning,
+		ExpectedContainerID: existing.CurrentContainerID,
+		To:                  runtime.StatusStopped,
+		Now:                 updateNow,
+	})
+	if errors.Is(err, runtime.ErrConflict) {
+		// CAS race: a concurrent reconciler / restart already moved the
+		// record. The desired terminal state was reached by another path.
+		return service.recordReplayNoOp(ctx, opStartedAt, input, existing), nil
+	}
+	if errors.Is(err, runtime.ErrNotFound) {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeNotFound,
+			errorMessage: fmt.Sprintf("runtime record for game %q vanished mid-stop", input.GameID),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	}
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("update runtime status: %s", err.Error()),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	}
+
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:      input.GameID,
+		OpKind:      operation.OpKindStop,
+		OpSource:    input.OpSource,
+		SourceRef:   input.SourceRef,
+		ImageRef:    existing.CurrentImageRef,
+		ContainerID: existing.CurrentContainerID,
+		Outcome:     operation.OutcomeSuccess,
+		StartedAt:   opStartedAt,
+		FinishedAt:  &finishedAt,
+	})
+	service.telemetry.RecordStopOutcome(ctx, string(operation.OutcomeSuccess), string(input.Reason), string(input.OpSource))
+
+	record := existing
+	record.Status = runtime.StatusStopped
+	stoppedAt := updateNow
+	record.StoppedAt = &stoppedAt
+	record.LastOpAt = updateNow
+
+	logArgs := []any{
+		"game_id", input.GameID,
+		"container_id", existing.CurrentContainerID,
+		"reason", string(input.Reason),
+		"op_source", string(input.OpSource),
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.InfoContext(ctx, "runtime stopped", logArgs...)
+
+	return Result{
+		Record:  record,
+		Outcome: operation.OutcomeSuccess,
+	}, nil
+}
+
+// handleVanished records the success outcome for the case where docker
+// stop reports the container as already gone. It updates the record to
+// removed, publishes container_disappeared, and returns success.
+func (service *Service) handleVanished(ctx context.Context, input Input, opStartedAt time.Time, existing runtime.RuntimeRecord) Result {
+	updateNow := service.clock().UTC()
+	err := service.runtimeRecords.UpdateStatus(ctx, ports.UpdateStatusInput{
+		GameID:              input.GameID,
+		ExpectedFrom:        runtime.StatusRunning,
+		ExpectedContainerID: existing.CurrentContainerID,
+		To:                  runtime.StatusRemoved,
+		Now:                 updateNow,
+	})
+	if errors.Is(err, runtime.ErrConflict) {
+		return service.recordReplayNoOp(ctx, opStartedAt, input, existing)
+	}
+	if err != nil && !errors.Is(err, runtime.ErrNotFound) {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("update runtime status to removed: %s", err.Error()),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		})
+	}
+
+	service.bestEffortPublishHealth(ctx, ports.HealthEventEnvelope{
+		GameID:      input.GameID,
+		ContainerID: existing.CurrentContainerID,
+		EventType:   health.EventTypeContainerDisappeared,
+		OccurredAt:  updateNow,
+		Details:     emptyHealthDetails(),
+	})
+
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:      input.GameID,
+		OpKind:      operation.OpKindStop,
+		OpSource:    input.OpSource,
+		SourceRef:   input.SourceRef,
+		ImageRef:    existing.CurrentImageRef,
+		ContainerID: existing.CurrentContainerID,
+		Outcome:     operation.OutcomeSuccess,
+		StartedAt:   opStartedAt,
+		FinishedAt:  &finishedAt,
+	})
+	service.telemetry.RecordStopOutcome(ctx, string(operation.OutcomeSuccess), string(input.Reason), string(input.OpSource))
+	service.telemetry.RecordHealthEvent(ctx, string(health.EventTypeContainerDisappeared))
+
+	record := existing
+	record.Status = runtime.StatusRemoved
+	record.CurrentContainerID = ""
+	removedAt := updateNow
+	record.RemovedAt = &removedAt
+	record.LastOpAt = updateNow
+
+	logArgs := []any{
+		"game_id", input.GameID,
+		"container_id", existing.CurrentContainerID,
+		"reason", string(input.Reason),
+		"op_source", string(input.OpSource),
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.InfoContext(ctx, "runtime stop on vanished container", logArgs...)
+
+	return Result{
+		Record:  record,
+		Outcome: operation.OutcomeSuccess,
+	}
+}
+
+// recordReplayNoOp records the idempotent replay outcome and returns the
+// existing record unchanged.
+func (service *Service) recordReplayNoOp(ctx context.Context, opStartedAt time.Time, input Input, existing runtime.RuntimeRecord) Result {
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:      input.GameID,
+		OpKind:      operation.OpKindStop,
+		OpSource:    input.OpSource,
+		SourceRef:   input.SourceRef,
+		ImageRef:    existing.CurrentImageRef,
+		ContainerID: existing.CurrentContainerID,
+		Outcome:     operation.OutcomeSuccess,
+		ErrorCode:   startruntime.ErrorCodeReplayNoOp,
+		StartedAt:   opStartedAt,
+		FinishedAt:  &finishedAt,
+	})
+	service.telemetry.RecordStopOutcome(ctx, string(operation.OutcomeSuccess), string(input.Reason), string(input.OpSource))
+
+	logArgs := []any{
+		"game_id", input.GameID,
+		"container_id", existing.CurrentContainerID,
+		"reason", string(input.Reason),
+		"op_source", string(input.OpSource),
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.InfoContext(ctx, "runtime stop replay no-op", logArgs...)
+
+	return Result{
+		Record:    existing,
+		Outcome:   operation.OutcomeSuccess,
+		ErrorCode: startruntime.ErrorCodeReplayNoOp,
+	}
+}
+
+// failureCtx groups the inputs to recordFailure so the runUnderLease
+// method stays readable.
+type failureCtx struct {
+	opStartedAt  time.Time
+	input        Input
+	errorCode    string
+	errorMessage string
+	containerID  string
+	imageRef     string
+}
+
+// recordFailure records the failure operation_log entry and emits
+// telemetry. The runtime record stays untouched.
+func (service *Service) recordFailure(ctx context.Context, fc failureCtx) Result {
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:       fc.input.GameID,
+		OpKind:       operation.OpKindStop,
+		OpSource:     fc.input.OpSource,
+		SourceRef:    fc.input.SourceRef,
+		ImageRef:     fc.imageRef,
+		ContainerID:  fc.containerID,
+		Outcome:      operation.OutcomeFailure,
+		ErrorCode:    fc.errorCode,
+		ErrorMessage: fc.errorMessage,
+		StartedAt:    fc.opStartedAt,
+		FinishedAt:   &finishedAt,
+	})
+	service.telemetry.RecordStopOutcome(ctx, string(operation.OutcomeFailure), string(fc.input.Reason), string(fc.input.OpSource))
+
+	logArgs := []any{
+		"game_id", fc.input.GameID,
+		"reason", string(fc.input.Reason),
+		"op_source", string(fc.input.OpSource),
+		"error_code", fc.errorCode,
+		"error_message", fc.errorMessage,
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.WarnContext(ctx, "runtime stop failed", logArgs...)
+
+	return Result{
+		Outcome:      operation.OutcomeFailure,
+		ErrorCode:    fc.errorCode,
+		ErrorMessage: fc.errorMessage,
+	}
+}
+
+// releaseLease releases the per-game lease in a fresh background context
+// so a canceled request context does not leave the lease pinned for its
+// TTL.
+func (service *Service) releaseLease(ctx context.Context, gameID, token string) {
+	cleanupCtx, cancel := context.WithTimeout(context.Background(), leaseReleaseTimeout)
+	defer cancel()
+	if err := service.leases.Release(cleanupCtx, gameID, token); err != nil {
+		service.logger.WarnContext(ctx, "release game lease",
+			"game_id", gameID,
+			"err", err.Error(),
+		)
+	}
+}
+
+// bestEffortAppend writes one operation_log entry. A failure is logged
+// and discarded; the durable runtime record (or its absence) remains
+// the source of truth.
+func (service *Service) bestEffortAppend(ctx context.Context, entry operation.OperationEntry) {
+	if _, err := service.operationLogs.Append(ctx, entry); err != nil {
+		service.logger.ErrorContext(ctx, "append operation log",
+			"game_id", entry.GameID,
+			"op_kind", string(entry.OpKind),
+			"outcome", string(entry.Outcome),
+			"error_code", entry.ErrorCode,
+			"err", err.Error(),
+		)
+	}
+}
+
+// bestEffortPublishHealth emits one health event + snapshot upsert.
+// Failures degrade silently per `rtmanager/README.md §Notification
+// Contracts`; the runtime record remains the source of truth.
+func (service *Service) bestEffortPublishHealth(ctx context.Context, envelope ports.HealthEventEnvelope) {
+	if err := service.healthEvents.Publish(ctx, envelope); err != nil {
+		service.logger.ErrorContext(ctx, "publish health event",
+			"game_id", envelope.GameID,
+			"container_id", envelope.ContainerID,
+			"event_type", string(envelope.EventType),
+			"err", err.Error(),
+		)
+	}
+}
+
+// defaultTokenGenerator returns a function that produces 32-byte
+// base64url-encoded tokens. Mirrors the start service: a degraded
+// entropy source falls back to a sentinel token so the next TryAcquire
+// observes a collision rather than a panic.
+func defaultTokenGenerator() func() string {
+	return func() string {
+		var buf [32]byte
+		if _, err := rand.Read(buf[:]); err != nil {
+			return "rtmanager-fallback-token"
+		}
+		return base64.RawURLEncoding.EncodeToString(buf[:])
+	}
+}
+
+// emptyHealthDetails returns the canonical empty-object payload required
+// by the `container_disappeared` AsyncAPI variant.
+func emptyHealthDetails() json.RawMessage {
+	return json.RawMessage("{}")
+}