实现持久化 assignment handoff

This commit is contained in:
2026-09-20 18:20:22 +00:00
parent 22dc3fef63
commit 0d8b36380d
10 changed files with 464 additions and 27 deletions
+96
View File
@@ -0,0 +1,96 @@
// Package assignmentqueue implements the durable assignment handoff with JetStream.
package assignmentqueue
import (
"context"
"errors"
"fmt"
"strings"
"time"
"github.com/nats-io/nats.go"
"github.com/nats-io/nats.go/jetstream"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskassignment"
)
type publishAPI interface {
PublishMsg(context.Context, *nats.Msg, ...jetstream.PublishOpt) (*jetstream.PubAck, error)
}
// Publisher implements the scheduler dispatcher with one subject per backend.
type Publisher struct {
JetStream publishAPI
SubjectBase string
}
func (p Publisher) Dispatch(ctx context.Context, assignment taskassignment.Assignment) error {
if p.JetStream == nil {
return errors.New("JetStream publisher is required")
}
body, err := taskassignment.Marshal(assignment)
if err != nil {
return err
}
base := strings.TrimSuffix(p.SubjectBase, ".")
if base == "" {
return errors.New("assignment subject base is required")
}
message := &nats.Msg{
Subject: base + "." + string(assignment.Backend),
Header: nats.Header{jetstream.MsgIDHeader: []string{assignment.ID}},
Data: body,
}
if _, err := p.JetStream.PublishMsg(ctx, message); err != nil {
return fmt.Errorf("publish assignment %s: %w", assignment.ID, err)
}
return nil
}
type Handler interface {
Handle(context.Context, taskassignment.Assignment) (bool, error)
}
// Message is the subset of jetstream.Msg needed by one reconciliation.
type Message interface {
Data() []byte
DoubleAck(context.Context) error
NakWithDelay(time.Duration) error
InProgress() error
TermWithReason(string) error
}
// Processor maps one delivery to one idempotent worker reconciliation.
type Processor struct {
TrustDomain string
Handler Handler
RetryDelay time.Duration
}
func (p Processor) Process(ctx context.Context, message Message) error {
if p.Handler == nil {
return errors.New("assignment handler is required")
}
assignment, err := taskassignment.Unmarshal(message.Data(), p.TrustDomain)
if err != nil {
return errors.Join(err, message.TermWithReason("invalid assignment"))
}
done, err := p.Handler.Handle(ctx, assignment)
if err != nil {
delay := p.RetryDelay
if delay <= 0 {
delay = 15 * time.Second
}
return errors.Join(err, message.NakWithDelay(delay))
}
if done {
if err := message.DoubleAck(ctx); err != nil {
return fmt.Errorf("ack assignment %s: %w", assignment.ID, err)
}
return nil
}
if err := message.InProgress(); err != nil {
return fmt.Errorf("extend assignment %s acknowledgement: %w", assignment.ID, err)
}
return nil
}
+130
View File
@@ -0,0 +1,130 @@
package assignmentqueue
import (
"context"
"errors"
"testing"
"time"
runnerv1 "gitea.dev/actionslib/runner/v1"
"github.com/nats-io/nats.go"
"github.com/nats-io/nats.go/jetstream"
"google.golang.org/protobuf/types/known/structpb"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskassignment"
)
func testAssignment(t *testing.T) taskassignment.Assignment {
t.Helper()
fields, err := structpb.NewStruct(map[string]any{"repository": "owner/repo"})
if err != nil {
t.Fatal(err)
}
assignment, err := taskassignment.New(&runnerv1.Task{
Id: 42,
Context: fields,
WorkflowPayload: []byte("jobs:\n publish:\n runs-on: [self-hosted, pod]\n steps: []\n"),
}, "ddupan.top")
if err != nil {
t.Fatal(err)
}
return assignment
}
type fakePublisher struct{ message *nats.Msg }
func (p *fakePublisher) PublishMsg(_ context.Context, message *nats.Msg, _ ...jetstream.PublishOpt) (*jetstream.PubAck, error) {
p.message = message
return &jetstream.PubAck{}, nil
}
func TestPublisherUsesBackendSubjectAndAssignmentDeduplication(t *testing.T) {
api := &fakePublisher{}
publisher := Publisher{JetStream: api, SubjectBase: "ci.assignment"}
if err := publisher.Dispatch(context.Background(), testAssignment(t)); err != nil {
t.Fatal(err)
}
if api.message.Subject != "ci.assignment.pod" {
t.Fatalf("subject = %q", api.message.Subject)
}
if api.message.Header.Get(jetstream.MsgIDHeader) != "gitea-task-42" {
t.Fatalf("message ID = %q", api.message.Header.Get(jetstream.MsgIDHeader))
}
}
type fakeHandler struct {
done bool
err error
}
func (h fakeHandler) Handle(context.Context, taskassignment.Assignment) (bool, error) {
return h.done, h.err
}
type fakeMessage struct {
data []byte
acked int
nacked time.Duration
inProgress int
terminated int
}
func (m *fakeMessage) Data() []byte { return m.data }
func (m *fakeMessage) DoubleAck(context.Context) error { m.acked++; return nil }
func (m *fakeMessage) NakWithDelay(delay time.Duration) error { m.nacked = delay; return nil }
func (m *fakeMessage) InProgress() error { m.inProgress++; return nil }
func (m *fakeMessage) TermWithReason(string) error { m.terminated++; return nil }
func encodedAssignment(t *testing.T) []byte {
t.Helper()
data, err := taskassignment.Marshal(testAssignment(t))
if err != nil {
t.Fatal(err)
}
return data
}
func TestProcessorAcknowledgesOnlyCompletedAssignment(t *testing.T) {
message := &fakeMessage{data: encodedAssignment(t)}
processor := Processor{TrustDomain: "ddupan.top", Handler: fakeHandler{done: true}}
if err := processor.Process(context.Background(), message); err != nil {
t.Fatal(err)
}
if message.acked != 1 || message.inProgress != 0 || message.nacked != 0 {
t.Fatalf("message = %#v", message)
}
}
func TestProcessorKeepsRunningAssignmentPending(t *testing.T) {
message := &fakeMessage{data: encodedAssignment(t)}
processor := Processor{TrustDomain: "ddupan.top", Handler: fakeHandler{}}
if err := processor.Process(context.Background(), message); err != nil {
t.Fatal(err)
}
if message.acked != 0 || message.inProgress != 1 {
t.Fatalf("message = %#v", message)
}
}
func TestProcessorRetriesBackendFailureAndTerminatesPoisonMessage(t *testing.T) {
retry := &fakeMessage{data: encodedAssignment(t)}
processor := Processor{
TrustDomain: "ddupan.top",
Handler: fakeHandler{err: errors.New("backend unavailable")},
RetryDelay: time.Minute,
}
if err := processor.Process(context.Background(), retry); err == nil {
t.Fatal("expected backend error")
}
if retry.nacked != time.Minute {
t.Fatalf("retry delay = %s", retry.nacked)
}
poison := &fakeMessage{data: []byte("not-json")}
if err := processor.Process(context.Background(), poison); err == nil {
t.Fatal("expected decode error")
}
if poison.terminated != 1 || poison.nacked != 0 {
t.Fatalf("poison message = %#v", poison)
}
}
+125
View File
@@ -0,0 +1,125 @@
// Package taskassignment defines the durable handoff between the scheduler and workers.
package taskassignment
import (
"bytes"
"encoding/json"
"errors"
"fmt"
"slices"
"gitea.dev/actionslib/pkg/model"
runnerv1 "gitea.dev/actionslib/runner/v1"
"google.golang.org/protobuf/proto"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskidentity"
)
const wireVersion = 1
type Backend string
const (
BackendPod Backend = "pod"
BackendVM Backend = "vm"
)
// Assignment is the only document persisted in the handoff queue.
type Assignment struct {
ID string
Backend Backend
Task *runnerv1.Task
Identity taskidentity.Identity
}
type envelope struct {
Version int `json:"version"`
ID string `json:"id"`
Backend Backend `json:"backend"`
Task []byte `json:"task"`
Identity taskidentity.Identity `json:"identity"`
}
// New derives all trusted assignment fields from the task fetched from Gitea.
func New(task *runnerv1.Task, trustDomain string) (Assignment, error) {
if task == nil || task.GetId() <= 0 {
return Assignment{}, errors.New("positive Gitea task ID is required")
}
identity, err := taskidentity.FromTask(task, trustDomain)
if err != nil {
return Assignment{}, err
}
backend, err := backendFromTask(task)
if err != nil {
return Assignment{}, err
}
return Assignment{
ID: fmt.Sprintf("gitea-task-%d", task.GetId()),
Backend: backend,
Task: task,
Identity: identity,
}, nil
}
func backendFromTask(task *runnerv1.Task) (Backend, error) {
workflow, err := model.ReadWorkflow(bytes.NewReader(task.GetWorkflowPayload()))
if err != nil {
return "", fmt.Errorf("parse task workflow for backend: %w", err)
}
jobIDs := workflow.GetJobIDs()
if len(jobIDs) != 1 || workflow.GetJob(jobIDs[0]) == nil {
return "", fmt.Errorf("task workflow must contain exactly one non-empty job")
}
labels := workflow.GetJob(jobIDs[0]).RunsOnLabels()
if !slices.Contains(labels, "self-hosted") {
return "", fmt.Errorf("task runs-on labels must include self-hosted: %v", labels)
}
hasPod := slices.Contains(labels, string(BackendPod))
hasVM := slices.Contains(labels, string(BackendVM))
if hasPod == hasVM {
return "", fmt.Errorf("task runs-on labels must select exactly one of pod or vm: %v", labels)
}
if hasPod {
return BackendPod, nil
}
return BackendVM, nil
}
// Marshal encodes a versioned assignment. Protobuf preserves the exact Gitea task.
func Marshal(assignment Assignment) ([]byte, error) {
if assignment.Task == nil {
return nil, errors.New("assignment task is required")
}
task, err := proto.Marshal(assignment.Task)
if err != nil {
return nil, fmt.Errorf("marshal Gitea task: %w", err)
}
return json.Marshal(envelope{
Version: wireVersion,
ID: assignment.ID, Backend: assignment.Backend,
Task: task, Identity: assignment.Identity,
})
}
// Unmarshal re-derives trusted fields instead of trusting duplicated queue metadata.
func Unmarshal(data []byte, trustDomain string) (Assignment, error) {
var wire envelope
if err := json.Unmarshal(data, &wire); err != nil {
return Assignment{}, fmt.Errorf("decode assignment: %w", err)
}
if wire.Version != wireVersion {
return Assignment{}, fmt.Errorf("unsupported assignment version %d", wire.Version)
}
task := new(runnerv1.Task)
if err := proto.Unmarshal(wire.Task, task); err != nil {
return Assignment{}, fmt.Errorf("unmarshal Gitea task: %w", err)
}
canonical, err := New(task, trustDomain)
if err != nil {
return Assignment{}, err
}
if wire.ID != canonical.ID || wire.Backend != canonical.Backend || wire.Identity != canonical.Identity {
return Assignment{}, errors.New("assignment metadata does not match its Gitea task")
}
return canonical, nil
}
@@ -0,0 +1,75 @@
package taskassignment
import (
"bytes"
"testing"
runnerv1 "gitea.dev/actionslib/runner/v1"
"google.golang.org/protobuf/types/known/structpb"
)
func task(t *testing.T, labels string) *runnerv1.Task {
t.Helper()
context, err := structpb.NewStruct(map[string]any{"repository": "owner/repo"})
if err != nil {
t.Fatal(err)
}
return &runnerv1.Task{
Id: 42,
Context: context,
WorkflowPayload: []byte("jobs:\n publish:\n runs-on: " + labels + "\n steps: []\n"),
}
}
func TestNewSelectsBackendFromRunsOn(t *testing.T) {
for _, test := range []struct {
labels string
backend Backend
}{
{"[self-hosted, pod]", BackendPod},
{"[self-hosted, vm]", BackendVM},
} {
assignment, err := New(task(t, test.labels), "ddupan.top")
if err != nil {
t.Fatal(err)
}
if assignment.Backend != test.backend || assignment.ID != "gitea-task-42" {
t.Fatalf("assignment = %#v", assignment)
}
}
}
func TestNewRejectsAmbiguousBackend(t *testing.T) {
for _, labels := range []string{
"[self-hosted]",
"[self-hosted, pod, vm]",
"[pod]",
} {
if _, err := New(task(t, labels), "ddupan.top"); err == nil {
t.Fatalf("expected labels %s to fail", labels)
}
}
}
func TestAssignmentWireRoundTripAndValidation(t *testing.T) {
want, err := New(task(t, "[self-hosted, pod]"), "ddupan.top")
if err != nil {
t.Fatal(err)
}
data, err := Marshal(want)
if err != nil {
t.Fatal(err)
}
got, err := Unmarshal(data, "ddupan.top")
if err != nil {
t.Fatal(err)
}
if got.ID != want.ID || got.Backend != want.Backend || got.Identity != want.Identity || !bytes.Equal(got.Task.WorkflowPayload, want.Task.WorkflowPayload) {
t.Fatalf("round trip = %#v, want %#v", got, want)
}
tampered := bytes.Replace(data, []byte(`"backend":"pod"`), []byte(`"backend":"vm"`), 1)
if _, err := Unmarshal(tampered, "ddupan.top"); err == nil {
t.Fatal("expected tampered backend to fail")
}
}
+4 -14
View File
@@ -5,19 +5,13 @@ package taskscheduler
import (
"context"
"errors"
"fmt"
runnerv1 "gitea.dev/actionslib/runner/v1"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskidentity"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskassignment"
)
// Assignment is the immutable input handed to a Pod or VM provisioner.
type Assignment struct {
ID string
Task *runnerv1.Task
Identity taskidentity.Identity
}
type Assignment = taskassignment.Assignment
// Dispatcher creates exactly one executor for an already assigned Gitea task.
// It must not register another runner or ask Gitea for a task.
@@ -36,13 +30,9 @@ func (s *Scheduler) Run(ctx context.Context, task *runnerv1.Task) error {
if s.Dispatcher == nil {
return errors.New("executor dispatcher is required")
}
identity, err := taskidentity.FromTask(task, s.TrustDomain)
assignment, err := taskassignment.New(task, s.TrustDomain)
if err != nil {
return err
}
return s.Dispatcher.Dispatch(ctx, Assignment{
ID: fmt.Sprintf("gitea-task-%d", task.GetId()),
Task: task,
Identity: identity,
})
return s.Dispatcher.Dispatch(ctx, assignment)
}
+5 -4
View File
@@ -6,8 +6,8 @@ import (
"errors"
"strconv"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskassignment"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskidentity"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskscheduler"
)
// Phase is observed from Kubernetes or OpenSandbox, never stored by the worker.
@@ -38,7 +38,7 @@ type Metadata struct {
// Every method must be idempotent.
type Backend interface {
Find(context.Context, string) (*Executor, error)
Create(context.Context, taskscheduler.Assignment, Metadata) (*Executor, error)
Create(context.Context, taskassignment.Assignment, Metadata) (*Executor, error)
BindIdentity(context.Context, *Executor, taskidentity.Identity) error
Delete(context.Context, *Executor) error
}
@@ -59,7 +59,7 @@ type Worker struct {
// Handle performs one reconciliation. Done means the queue message may be
// acknowledged. A false result should remain pending and be reconciled again.
func (w Worker) Handle(ctx context.Context, assignment taskscheduler.Assignment) (done bool, err error) {
func (w Worker) Handle(ctx context.Context, assignment taskassignment.Assignment) (done bool, err error) {
if w.Backend == nil || w.Tasks == nil {
return false, errors.New("backend and Gitea task state are required")
}
@@ -114,11 +114,12 @@ func (w Worker) Handle(ctx context.Context, assignment taskscheduler.Assignment)
}
// BackendMetadata is the shared metadata contract for Pods and OpenSandbox.
func BackendMetadata(assignment taskscheduler.Assignment) Metadata {
func BackendMetadata(assignment taskassignment.Assignment) Metadata {
return Metadata{
Labels: map[string]string{
"ci.ddupan.top/assignment-id": assignment.ID,
"ci.ddupan.top/task-id": strconv.FormatInt(assignment.Task.GetId(), 10),
"ci.ddupan.top/backend": string(assignment.Backend),
},
Annotations: map[string]string{
"ci.ddupan.top/repository": assignment.Identity.Repository,
+7 -6
View File
@@ -6,8 +6,8 @@ import (
runnerv1 "gitea.dev/actionslib/runner/v1"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskassignment"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskidentity"
"git.ddupan.top/panxiao81/gitea-dynamic-runner/internal/taskscheduler"
)
type fakeBackend struct {
@@ -18,7 +18,7 @@ type fakeBackend struct {
}
func (b *fakeBackend) Find(context.Context, string) (*Executor, error) { return b.executor, nil }
func (b *fakeBackend) Create(_ context.Context, _ taskscheduler.Assignment, _ Metadata) (*Executor, error) {
func (b *fakeBackend) Create(_ context.Context, _ taskassignment.Assignment, _ Metadata) (*Executor, error) {
b.created++
b.executor = &Executor{Name: "executor", IdentityTarget: "pod-uid", Phase: PhaseRunning}
return b.executor, nil
@@ -45,10 +45,11 @@ func (t *fakeTasks) Report(_ context.Context, _ int64, phase Phase) error {
return nil
}
func assignment() taskscheduler.Assignment {
return taskscheduler.Assignment{
ID: "gitea-task-42",
Task: &runnerv1.Task{Id: 42},
func assignment() taskassignment.Assignment {
return taskassignment.Assignment{
ID: "gitea-task-42",
Backend: taskassignment.BackendPod,
Task: &runnerv1.Task{Id: 42},
Identity: taskidentity.Identity{
Repository: "owner/repo",
Task: "publish",