test: verify workflow version isolation and pause recovery races
This commit is contained in:
202
sdk/go/workflow_process_integration_test.go
Normal file
202
sdk/go/workflow_process_integration_test.go
Normal file
@@ -0,0 +1,202 @@
|
||||
package orbit
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"testing"
|
||||
"time"
|
||||
)
|
||||
|
||||
// 此测试由脚本在原服务退出后单独运行；所有流程操作只走公开HTTP/WSS SDK。
|
||||
// 两旧节点均Wait确认SIGKILL后才等待原时钟到期并启动两新节点。
|
||||
func TestIntegrationWorkflowProcessPauseRecovery(t *testing.T) {
|
||||
a := startOrbitProcess(t)
|
||||
b := startOrbitProcess(t)
|
||||
_, queue, token := integrationWorkerClient(t)
|
||||
api := processClient(t, a, token)
|
||||
ctx, cancel := context.WithTimeout(context.Background(), 50*time.Second)
|
||||
defer cancel()
|
||||
handler := "pause.process." + workerTestID(t)
|
||||
spec := workflowSDKSpec(queue, handler)
|
||||
businessKey := WorkflowRef("workflow", "business_key")
|
||||
spec.Body.Steps[0].Task.BusinessKey = &businessKey
|
||||
initial := *spec.Body.Steps[0].Task
|
||||
query := initial
|
||||
query.Handler = handler + ".query"
|
||||
query.ResultType = WorkflowType{Kind: "STRING", Enum: []string{"PENDING", "DONE"}}
|
||||
query.TimeoutMS = 20000
|
||||
parallel := WorkflowParallel{}
|
||||
for _, branch := range []struct {
|
||||
step WorkflowStep
|
||||
typ WorkflowType
|
||||
}{
|
||||
{WorkflowStep{Key: "initial", Kind: "TASK", Task: &initial}, spec.InputType},
|
||||
{WorkflowStep{Key: "signal", Kind: "SIGNAL", Signal: &WorkflowSignal{Name: "external.done", PayloadVersion: "v1", Correlation: WorkflowRef("workflow", "business_key"), PayloadType: spec.InputType, TimeoutMS: 30000}}, spec.InputType},
|
||||
{WorkflowStep{Key: "timer", Kind: "TIMER", Timer: &WorkflowTimer{DelayMS: 5000}}, WorkflowType{Kind: "NULL"}},
|
||||
{WorkflowStep{Key: "query", Kind: "POLL", Poll: &WorkflowPoll{Task: query, MaxRounds: 2, TimeoutMS: 30000, IntervalMS: 5000, Pending: []string{"PENDING"}, Succeeded: []string{"DONE"}}}, query.ResultType},
|
||||
} {
|
||||
parallel.Branches = append(parallel.Branches, WorkflowParallelBranch{Name: branch.step.Key, Body: WorkflowBlock{Steps: []WorkflowStep{branch.step}, OutputType: branch.typ, Output: WorkflowRef(branch.step.Key)}})
|
||||
}
|
||||
spec.Body.Steps = append([]WorkflowStep{{Key: "join", Kind: "PARALLEL", Parallel: ¶llel}}, spec.Body.Steps...)
|
||||
in := publishWorkflowSDK(t, ctx, api, spec)
|
||||
in.BusinessKey = handler
|
||||
startKey := workerTestID(t)
|
||||
started, err := api.StartWorkflow(ctx, startKey, in)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
v := started.Instance
|
||||
work := workflowWorkerSDK(t, ctx, api, queue, handler)
|
||||
queries := workflowWorkerSDK(t, ctx, api, queue, query.Handler)
|
||||
first, err := work.NextAssignment(ctx)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if _, err = work.Start(ctx, workerTestID(t), first.Attempt()); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
firstQuery := finishWorkflowSDK(t, ctx, queries, WorkerSucceeded, json.RawMessage(`"PENDING"`))
|
||||
pollBefore := waitWorkflowStepSDK(t, ctx, api, v.ID, "query", func(d WorkflowStepDetail) bool {
|
||||
return d.Poll != nil && d.Poll.UsedRounds == 1 && d.Poll.NextPollAt != nil
|
||||
})
|
||||
timerBefore, err := api.GetWorkflowStep(ctx, v.ID, "timer")
|
||||
if err != nil || timerBefore.Wait == nil || timerBefore.ActivatedAt == nil || timerBefore.Wait.State != "WAITING" {
|
||||
t.Fatal("timer not waiting before pause", err)
|
||||
}
|
||||
signalBefore, err := api.GetWorkflowStep(ctx, v.ID, "signal")
|
||||
if err != nil || signalBefore.Wait == nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
pauseKey := workerTestID(t)
|
||||
pauseInput := WorkflowControlRequest{Operation: "PAUSE", ExpectedControlRevision: v.ControlRevision, Reason: "persist pause across all controller exits"}
|
||||
paused, err := api.ControlWorkflow(ctx, v.ID, pauseKey, pauseInput)
|
||||
if err != nil || !paused.Instance.Paused {
|
||||
t.Fatal(err)
|
||||
}
|
||||
signalKey := workerTestID(t)
|
||||
signalInput := SendWorkflowSignalRequest{Name: "external.done", PayloadVersion: "v1", TargetStepKey: "signal", TargetActivationSequence: signalBefore.Step.ActivationSequence, CorrelationKey: v.BusinessKey, Payload: workflowExact, Reason: "verified result while paused", EvidenceRef: "proof:paused-process"}
|
||||
accepted, err := api.SendWorkflowSignal(ctx, v.ID, signalKey, signalInput)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if _, err = work.Finish(ctx, workerTestID(t), WorkerFinish{Ref: first.Attempt(), Outcome: WorkerSucceeded, Result: workflowExact}); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
initialTask, err := api.GetTask(ctx, first.Task.ID)
|
||||
if err != nil || initialTask.State != TaskSucceeded || initialTask.AttemptCount != 1 {
|
||||
t.Fatal("paused result not committed", err)
|
||||
}
|
||||
work.Close()
|
||||
queries.Close()
|
||||
a.kill(t)
|
||||
b.kill(t)
|
||||
// 以首次激活时间计算原TIMER唤醒，不用进程启动时间重新计时。
|
||||
wake := timerBefore.ActivatedAt.Add(5 * time.Second)
|
||||
if !time.Now().Before(wake) {
|
||||
t.Fatal("all controllers must stop before the original timer wake")
|
||||
}
|
||||
if pollBefore.Poll.NextPollAt.After(wake) {
|
||||
wake = *pollBefore.Poll.NextPollAt
|
||||
}
|
||||
if wait := time.Until(wake.Add(100 * time.Millisecond)); wait > 0 {
|
||||
select {
|
||||
case <-time.After(wait):
|
||||
case <-ctx.Done():
|
||||
t.Fatal(ctx.Err())
|
||||
}
|
||||
}
|
||||
c := startOrbitProcess(t)
|
||||
d := startOrbitProcess(t)
|
||||
if a.nodeID == b.nodeID || c.nodeID == d.nodeID || c.nodeID == a.nodeID || c.nodeID == b.nodeID || d.nodeID == a.nodeID || d.nodeID == b.nodeID {
|
||||
t.Fatal("node identity reused")
|
||||
}
|
||||
api = processClient(t, c, token)
|
||||
other := processClient(t, d, token)
|
||||
for _, entry := range []struct {
|
||||
key string
|
||||
before WorkflowStepDetail
|
||||
}{{"signal", signalBefore}, {"timer", timerBefore}} {
|
||||
after := waitWorkflowStepSDK(t, ctx, api, v.ID, entry.key, func(d WorkflowStepDetail) bool { return d.Step.State == "SUCCEEDED" })
|
||||
if after.Wait == nil || after.Wait.ID != entry.before.Wait.ID || !after.Wait.DeadlineAt.Equal(entry.before.Wait.DeadlineAt) || after.Step.ActivationSequence != entry.before.Step.ActivationSequence || *after.Step.ActivationID != *entry.before.Step.ActivationID {
|
||||
t.Fatal("wait identity or original deadline changed across pause/restart")
|
||||
}
|
||||
}
|
||||
initialAfter := waitWorkflowStepSDK(t, ctx, api, v.ID, "initial", func(d WorkflowStepDetail) bool { return d.Step.State == "SUCCEEDED" })
|
||||
if initialAfter.Step.TaskID == nil || *initialAfter.Step.TaskID != first.Task.ID {
|
||||
t.Fatal("old result replaced after restart")
|
||||
}
|
||||
current, err := other.GetWorkflow(ctx, v.ID)
|
||||
if err != nil || !current.Paused || current.ControlRevision != paused.Instance.ControlRevision || current.DefinitionVersionID != v.DefinitionVersionID || !current.DeadlineAt.Equal(v.DeadlineAt) {
|
||||
t.Fatal("paused workflow changed on new nodes", err)
|
||||
}
|
||||
tail, err := api.GetWorkflowStep(ctx, v.ID, "work")
|
||||
if err != nil || tail.Step.State != "BLOCKED" || tail.Step.TaskID != nil {
|
||||
t.Fatal("paused successor created", err)
|
||||
}
|
||||
pollAfter, err := api.GetWorkflowStep(ctx, v.ID, "query")
|
||||
if err != nil || pollAfter.Poll == nil || pollAfter.Poll.UsedRounds != 1 || pollAfter.Poll.LastTaskID != firstQuery.Task.ID || pollAfter.Poll.CurrentTaskID != nil || pollAfter.Poll.NextPollAt == nil || !pollAfter.Poll.NextPollAt.Equal(*pollBefore.Poll.NextPollAt) || !pollAfter.Poll.DeadlineAt.Equal(pollBefore.Poll.DeadlineAt) || pollAfter.Poll.MaxRounds != 2 || pollAfter.Poll.MaxAttemptsPerRound != 2 {
|
||||
t.Fatal("paused query used another round or reset original budget", err)
|
||||
}
|
||||
received, err := api.GetWorkflowSignal(ctx, v.ID, accepted.Signal.ID)
|
||||
if err != nil || received.State != "CONSUMED" || !received.AcceptedAt.Equal(accepted.Signal.AcceptedAt) || received.TargetActivationSequence != accepted.Signal.TargetActivationSequence {
|
||||
t.Fatal("accepted signal lost", err)
|
||||
}
|
||||
// 在另一个新节点恢复原PAUSE、SIGNAL和START回执，不能修改当前状态或预算。
|
||||
duplicate, err := other.ControlWorkflow(ctx, v.ID, pauseKey, pauseInput)
|
||||
if err != nil || !duplicate.Deduplicated || duplicate.Instance.ControlRevision != paused.Instance.ControlRevision {
|
||||
t.Fatal("pause receipt lost", err)
|
||||
}
|
||||
signalDuplicate, err := other.SendWorkflowSignal(ctx, v.ID, signalKey, signalInput)
|
||||
if err != nil || !signalDuplicate.Deduplicated || signalDuplicate.Signal.ID != accepted.Signal.ID || signalDuplicate.Signal.State != accepted.Signal.State {
|
||||
t.Fatal("signal receipt changed", err)
|
||||
}
|
||||
startDuplicate, err := other.StartWorkflow(ctx, startKey, in)
|
||||
if err != nil || !startDuplicate.Deduplicated || startDuplicate.Instance.ID != v.ID || !startDuplicate.Instance.DeadlineAt.Equal(v.DeadlineAt) {
|
||||
t.Fatal("start receipt changed", err)
|
||||
}
|
||||
resumeKey := workerTestID(t)
|
||||
resumeInput := WorkflowControlRequest{Operation: "RESUME", ExpectedControlRevision: paused.Instance.ControlRevision, Reason: "resume persisted pause through another node"}
|
||||
resumed, err := other.ControlWorkflow(ctx, v.ID, resumeKey, resumeInput)
|
||||
if err != nil || resumed.Instance.Paused || resumed.Instance.ControlRevision != 3 || !resumed.Instance.DeadlineAt.Equal(v.DeadlineAt) {
|
||||
t.Fatal("resume changed fixed state", err)
|
||||
}
|
||||
queries = workflowWorkerSDK(t, ctx, api, queue, query.Handler)
|
||||
lastQuery := finishWorkflowSDK(t, ctx, queries, WorkerSucceeded, json.RawMessage(`"DONE"`))
|
||||
work = workflowWorkerSDK(t, ctx, other, queue, handler)
|
||||
last := finishWorkflowSDK(t, ctx, work, WorkerSucceeded, workflowExact)
|
||||
finished := waitWorkflowSDK(t, ctx, api, v.ID, "SUCCEEDED")
|
||||
if finished.Paused || finished.ControlRevision != 3 || !finished.DeadlineAt.Equal(v.DeadlineAt) || finished.DefinitionVersionID != v.DefinitionVersionID || string(finished.Output) != string(workflowExact) {
|
||||
t.Fatal("final output or original version/deadline changed")
|
||||
}
|
||||
finalPoll, err := api.GetWorkflowStep(ctx, v.ID, "query")
|
||||
if err != nil || finalPoll.Poll == nil || finalPoll.Poll.UsedRounds != 2 || !finalPoll.Poll.DeadlineAt.Equal(pollBefore.Poll.DeadlineAt) || finalPoll.Poll.CurrentTaskID != nil {
|
||||
t.Fatal("query did not preserve the two-round budget", err)
|
||||
}
|
||||
rounds, err := api.ListWorkflowPollRounds(ctx, v.ID, "query", *finalPoll.Step.ActivationID, HistoryPageRequest{})
|
||||
if err != nil || len(rounds.Items) != 2 || rounds.NextCursor != "" || rounds.Items[0].Round != 2 || rounds.Items[0].TaskID != lastQuery.Task.ID || rounds.Items[1].Round != 1 || rounds.Items[1].TaskID != firstQuery.Task.ID {
|
||||
t.Fatal("query history duplicated or reset after paused downtime", err)
|
||||
}
|
||||
tasks, err := api.ListTasks(ctx, ListTasksRequest{QueueID: queue, BusinessKey: handler, Limit: 100})
|
||||
if err != nil || len(tasks.Items) != 4 || tasks.NextCursor != "" {
|
||||
t.Fatal("unexpected tasks in paused workflow", err)
|
||||
}
|
||||
for _, assignment := range []WorkerAssignment{first, firstQuery, lastQuery, last} {
|
||||
task, err := api.GetTask(ctx, assignment.Task.ID)
|
||||
if err != nil || task.State != TaskSucceeded || task.AttemptCount != 1 {
|
||||
t.Fatal("duplicated execution after full stop", err)
|
||||
}
|
||||
history, err := api.ListTaskAttempts(ctx, task.ID, HistoryPageRequest{})
|
||||
if err != nil || len(history.Items) != 1 || history.NextCursor != "" {
|
||||
t.Fatal("historical attempts lost or duplicated", err)
|
||||
}
|
||||
}
|
||||
unchanged, err := api.GetTask(ctx, initialTask.ID)
|
||||
if err != nil || !unchanged.DeadlineAt.Equal(initialTask.DeadlineAt) || unchanged.Retry != initialTask.Retry {
|
||||
t.Fatal("original task policy changed", err)
|
||||
}
|
||||
resumeDuplicate, err := api.ControlWorkflow(ctx, v.ID, resumeKey, resumeInput)
|
||||
if err != nil || !resumeDuplicate.Deduplicated || resumeDuplicate.Instance.ControlRevision != 3 {
|
||||
t.Fatal("resume receipt changed", err)
|
||||
}
|
||||
t.Log("both original nodes SIGKILL confirmed; paused task/signal/timer collected by two fresh nodes; public SDK resume preserves version, receipts, deadlines, two query rounds and four single attempts")
|
||||
}
|
||||
Reference in New Issue
Block a user