open-nomad/client/allocrunner/alloc_runner_unix_test.go
Mahmood Ali 4afd7835e3 Fail alloc if alloc runner prestart hooks fail
When an alloc runner prestart hook fails, the task runners aren't invoked
and they remain in a pending state.

This leads to terrible results, some of which are:
* Lockup in GC process as reported in https://github.com/hashicorp/nomad/pull/5861
* Lockup in shutdown process as TR.Shutdown() waits for WaitCh to be closed
* Alloc not being restarted/rescheduled to another node (as it's still in
  pending state)
* Unexpected restart of alloc on a client restart, potentially days/weeks after
  alloc expected start time!

Here, we treat all tasks to have failed if alloc runner prestart hook fails.
This fixes the lockups, and permits the alloc to be rescheduled on another node.

While it's desirable to retry alloc runner in such failures, I opted to treat it
out of scope.  I'm afraid of some subtles about alloc and task runners and their
idempotency that's better handled in a follow up PR.

This might be one of the root causes for
https://github.com/hashicorp/nomad/issues/5840 .
2019-07-02 18:35:47 +08:00

294 lines
8.1 KiB
Go

// +build !windows
package allocrunner
import (
"encoding/json"
"fmt"
"os"
"syscall"
"testing"
"time"
"github.com/hashicorp/nomad/client/consul"
"github.com/hashicorp/nomad/client/state"
"github.com/hashicorp/nomad/nomad/mock"
"github.com/hashicorp/nomad/nomad/structs"
"github.com/hashicorp/nomad/testutil"
"github.com/stretchr/testify/require"
)
// TestAllocRunner_Restore_RunningTerminal asserts that restoring a terminal
// alloc with a running task properly kills the running the task. This is meant
// to simulate a Nomad agent crash after receiving an updated alloc with
// DesiredStatus=Stop, persisting the update, but crashing before terminating
// the task.
func TestAllocRunner_Restore_RunningTerminal(t *testing.T) {
t.Parallel()
// 1. Run task
// 2. Shutdown alloc runner
// 3. Set alloc.desiredstatus=false
// 4. Start new alloc runner
// 5. Assert task and logmon are cleaned up
alloc := mock.Alloc()
task := alloc.Job.TaskGroups[0].Tasks[0]
task.Driver = "mock_driver"
task.Config = map[string]interface{}{
"run_for": "1h",
}
conf, cleanup := testAllocRunnerConfig(t, alloc.Copy())
defer cleanup()
// Maintain state for subsequent run
conf.StateDB = state.NewMemDB(conf.Logger)
// Start and wait for task to be running
ar, err := NewAllocRunner(conf)
require.NoError(t, err)
go ar.Run()
defer destroy(ar)
testutil.WaitForResult(func() (bool, error) {
s := ar.AllocState()
return s.ClientStatus == structs.AllocClientStatusRunning, fmt.Errorf("expected running, got %s", s.ClientStatus)
}, func(err error) {
require.NoError(t, err)
})
// Shutdown the AR and manually change the state to mimic a crash where
// a stopped alloc update is received, but Nomad crashes before
// stopping the alloc.
ar.Shutdown()
select {
case <-ar.ShutdownCh():
case <-time.After(30 * time.Second):
require.Fail(t, "AR took too long to exit")
}
// Assert logmon is still running. This is a super ugly hack that pulls
// logmon's PID out of its reattach config, but it does properly ensure
// logmon gets cleaned up.
ls, _, err := conf.StateDB.GetTaskRunnerState(alloc.ID, task.Name)
require.NoError(t, err)
require.NotNil(t, ls)
logmonReattach := struct {
Pid int
}{}
err = json.Unmarshal([]byte(ls.Hooks["logmon"].Data["reattach_config"]), &logmonReattach)
require.NoError(t, err)
logmonProc, _ := os.FindProcess(logmonReattach.Pid)
require.NoError(t, logmonProc.Signal(syscall.Signal(0)))
// Fake alloc terminal during Restore()
alloc.DesiredStatus = structs.AllocDesiredStatusStop
alloc.ModifyIndex++
alloc.AllocModifyIndex++
// Start a new alloc runner and assert it gets stopped
conf2, cleanup2 := testAllocRunnerConfig(t, alloc)
defer cleanup2()
// Use original statedb to maintain hook state
conf2.StateDB = conf.StateDB
// Restore, start, and wait for task to be killed
ar2, err := NewAllocRunner(conf2)
require.NoError(t, err)
require.NoError(t, ar2.Restore())
go ar2.Run()
defer destroy(ar2)
select {
case <-ar2.WaitCh():
case <-time.After(30 * time.Second):
}
// Assert logmon was cleaned up
require.Error(t, logmonProc.Signal(syscall.Signal(0)))
// Assert consul was cleaned up:
// 2 removals (canary+noncanary) during prekill
// 2 removals (canary+noncanary) during exited
// 2 removals (canary+noncanary) during stop
consulOps := conf2.Consul.(*consul.MockConsulServiceClient).GetOps()
require.Len(t, consulOps, 6)
for _, op := range consulOps {
require.Equal(t, "remove", op.Op)
}
// Assert terminated task event was emitted
events := ar2.AllocState().TaskStates[task.Name].Events
require.Len(t, events, 4)
require.Equal(t, events[0].Type, structs.TaskReceived)
require.Equal(t, events[1].Type, structs.TaskSetup)
require.Equal(t, events[2].Type, structs.TaskStarted)
require.Equal(t, events[3].Type, structs.TaskTerminated)
}
// TestAllocRunner_Restore_CompletedBatch asserts that restoring a completed
// batch alloc doesn't run it again
func TestAllocRunner_Restore_CompletedBatch(t *testing.T) {
t.Parallel()
// 1. Run task and wait for it to complete
// 2. Start new alloc runner
// 3. Assert task didn't run again
alloc := mock.Alloc()
alloc.Job.Type = structs.JobTypeBatch
task := alloc.Job.TaskGroups[0].Tasks[0]
task.Driver = "mock_driver"
task.Config = map[string]interface{}{
"run_for": "2ms",
}
conf, cleanup := testAllocRunnerConfig(t, alloc.Copy())
defer cleanup()
// Maintain state for subsequent run
conf.StateDB = state.NewMemDB(conf.Logger)
// Start and wait for task to be running
ar, err := NewAllocRunner(conf)
require.NoError(t, err)
go ar.Run()
defer destroy(ar)
testutil.WaitForResult(func() (bool, error) {
s := ar.AllocState()
if s.ClientStatus != structs.AllocClientStatusComplete {
return false, fmt.Errorf("expected complete, got %s", s.ClientStatus)
}
return true, nil
}, func(err error) {
require.NoError(t, err)
})
// once job finishes, it shouldn't run again
require.False(t, ar.shouldRun())
initialRunEvents := ar.AllocState().TaskStates[task.Name].Events
require.Len(t, initialRunEvents, 4)
ls, ts, err := conf.StateDB.GetTaskRunnerState(alloc.ID, task.Name)
require.NoError(t, err)
require.NotNil(t, ls)
require.Equal(t, structs.TaskStateDead, ts.State)
// Start a new alloc runner and assert it gets stopped
conf2, cleanup2 := testAllocRunnerConfig(t, alloc)
defer cleanup2()
// Use original statedb to maintain hook state
conf2.StateDB = conf.StateDB
// Restore, start, and wait for task to be killed
ar2, err := NewAllocRunner(conf2)
require.NoError(t, err)
require.NoError(t, ar2.Restore())
go ar2.Run()
defer destroy(ar2)
// AR waitCh must be closed even when task doesn't run again
select {
case <-ar2.WaitCh():
case <-time.After(10 * time.Second):
require.Fail(t, "alloc.waitCh wasn't closed")
}
// TR waitCh must be closed too!
select {
case <-ar2.tasks[task.Name].WaitCh():
case <-time.After(10 * time.Second):
require.Fail(t, "tr.waitCh wasn't closed")
}
// Assert that events are unmodified, which they would if task re-run
events := ar2.AllocState().TaskStates[task.Name].Events
require.Equal(t, initialRunEvents, events)
}
// TestAllocRunner_PreStartFailuresLeadToFailed asserts that if an alloc
// prestart hooks failed, then the alloc and subsequent tasks transition
// to failed state
func TestAllocRunner_PreStartFailuresLeadToFailed(t *testing.T) {
t.Parallel()
alloc := mock.Alloc()
alloc.Job.Type = structs.JobTypeBatch
task := alloc.Job.TaskGroups[0].Tasks[0]
task.Driver = "mock_driver"
task.Config = map[string]interface{}{
"run_for": "2ms",
}
alloc.Job.TaskGroups[0].RestartPolicy = &structs.RestartPolicy{
Attempts: 0,
}
conf, cleanup := testAllocRunnerConfig(t, alloc.Copy())
defer cleanup()
// Maintain state for subsequent run
conf.StateDB = state.NewMemDB(conf.Logger)
// Start and wait for task to be running
ar, err := NewAllocRunner(conf)
require.NoError(t, err)
ar.runnerHooks = append(ar.runnerHooks, &allocFailingPrestartHook{})
go ar.Run()
defer destroy(ar)
select {
case <-ar.WaitCh():
case <-time.After(10 * time.Second):
require.Fail(t, "alloc.waitCh wasn't closed")
}
testutil.WaitForResult(func() (bool, error) {
s := ar.AllocState()
if s.ClientStatus != structs.AllocClientStatusFailed {
return false, fmt.Errorf("expected complete, got %s", s.ClientStatus)
}
return true, nil
}, func(err error) {
require.NoError(t, err)
})
// once job finishes, it shouldn't run again
require.False(t, ar.shouldRun())
initialRunEvents := ar.AllocState().TaskStates[task.Name].Events
require.Len(t, initialRunEvents, 2)
ls, ts, err := conf.StateDB.GetTaskRunnerState(alloc.ID, task.Name)
require.NoError(t, err)
require.NotNil(t, ls)
require.NotNil(t, ts)
require.Equal(t, structs.TaskStateDead, ts.State)
require.True(t, ts.Failed)
// TR waitCh must be closed too!
select {
case <-ar.tasks[task.Name].WaitCh():
case <-time.After(10 * time.Second):
require.Fail(t, "tr.waitCh wasn't closed")
}
}
type allocFailingPrestartHook struct{}
func (*allocFailingPrestartHook) Name() string { return "failing_prestart" }
func (*allocFailingPrestartHook) Prerun() error {
return fmt.Errorf("failing prestart hooks")
}