func (ec *EtcdCoordinator) parseTask(resp *etcd.Response) metafora.Task {
// Sanity check / test path invariant
if !strings.HasPrefix(resp.Node.Key, ec.taskPath) {
metafora.Errorf("%s received task from outside task path: %s", ec.name, resp.Node.Key)
return nil
}
key := strings.Trim(resp.Node.Key, "/") // strip leading and trailing /s
parts := strings.Split(key, "/")
// Pickup new tasks
if newActions[resp.Action] && len(parts) == 3 && resp.Node.Dir {
// Make sure it's not already claimed before returning it
for _, n := range resp.Node.Nodes {
if strings.HasSuffix(n.Key, OwnerMarker) {
metafora.Debugf("%s ignoring task as it's already claimed: %s", ec.name, parts[2])
return nil
}
}
metafora.Debugf("%s received new task: %s", ec.name, parts[2])
props := ""
for _, n := range resp.Node.Nodes {
if strings.HasSuffix(n.Key, "/"+PropsKey) {
props = n.Value
break
}
}
return ec.newTask(parts[2], props)
}
if newActions[resp.Action] && len(parts) == 4 && parts[3] == PropsKey {
metafora.Debugf("%s received task with properties: %s", ec.name, parts[2])
return ec.newTask(parts[2], resp.Node.Value)
}
// If a claim key is removed, try to claim the task
if releaseActions[resp.Action] && len(parts) == 4 && parts[3] == OwnerMarker {
metafora.Debugf("%s received released task: %s", ec.name, parts[2])
// Sadly we need to fail parsing this task if there's an error getting the
// props file as trying to claim a task without properly knowing its
// properties could cause major issues.
parts[3] = PropsKey
propsnode, err := ec.client.Get(path.Join(parts...), unsorted, notrecursive)
if err != nil {
if ee, ok := err.(*etcd.EtcdError); ok && ee.ErrorCode == EcodeKeyNotFound {
// No props file
return ec.newTask(parts[2], "")
}
metafora.Errorf("%s error getting properties while handling %s", ec.name, parts[2])
return nil
}
return ec.newTask(parts[2], propsnode.Node.Value)
}
// Ignore any other key events (_metafora keys, task deletion, etc.)
return nil
}
// remove tells a single task's refresher to stop and blocks until the task is
// handled.
func (m *taskManager) remove(taskID string, done bool) {
m.taskL.Lock()
states, ok := m.tasks[taskID]
if !ok {
m.taskL.Unlock()
metafora.Debugf("Cannot remove task %s from refresher: not present.", taskID)
return
}
select {
case <-states.release:
// already stopping
case <-states.done:
// already stopping
default:
if done {
close(states.done)
} else {
close(states.release)
}
}
m.taskL.Unlock()
// Block until task is released/deleted to prevent races on shutdown where
// the process could exit before all tasks are released.
<-states.finished
}
func (ec *EtcdCoordinator) upsertDir(path string, ttl uint64) {
//hidden etcd key that isn't visible to ls commands on the directory,
// you have to know about it to find it :). I'm using it to add some
// info about when the cluster's schema was setup.
pathMarker := path + "/" + MetadataKey
_, err := ec.client.Get(path, unsorted, notrecursive)
if err == nil {
return
}
etcdErr, ok := err.(*etcd.EtcdError)
if ok && etcdErr.ErrorCode == EcodeKeyNotFound {
_, err := ec.client.CreateDir(path, ttl)
if err != nil {
metafora.Debugf("Error trying to create directory. path:[%s] error:[ %v ]", path, err)
}
host, _ := os.Hostname()
metadata := struct {
Host string `json:"host"`
CreatedTime string `json:"created"`
ownerValue
}{
Host: host,
CreatedTime: time.Now().String(),
ownerValue: ownerValue{Node: ec.conf.Name},
}
metadataB, _ := json.Marshal(metadata)
metadataStr := string(metadataB)
ec.client.Create(pathMarker, metadataStr, ttl)
}
}
// Init is called once by the consumer to provide a Logger to Coordinator
// implementations.
func (ec *EtcdCoordinator) Init(cordCtx metafora.CoordinatorContext) error {
metafora.Debugf("Initializing coordinator with namespace: %s and etcd cluster: %s",
ec.conf.Namespace, strings.Join(ec.client.GetCluster(), ", "))
ec.cordCtx = cordCtx
ec.upsertDir(ec.conf.Namespace, foreverTTL)
ec.upsertDir(ec.taskPath, foreverTTL)
if _, err := ec.client.CreateDir(ec.nodePath, ec.conf.NodeTTL); err != nil {
return err
}
// Create etcd client for task manager
tmc, err := newEtcdClient(ec.conf.Hosts)
if err != nil {
return err
}
// Start goroutine to heartbeat node key in etcd
go ec.nodeRefresher()
ec.upsertDir(ec.commandPath, foreverTTL)
ec.taskManager = newManager(cordCtx, tmc, ec.taskPath, ec.conf.Name, ec.conf.ClaimTTL)
return nil
}
// SubmitTask creates a new task in etcd
func (mc *mclient) SubmitTask(task metafora.Task) error {
fullpath := path.Join(mc.tskPath(task.ID()), PropsKey)
buf, err := json.Marshal(task)
if err != nil {
return err
}
if _, err := mc.etcd.Create(fullpath, string(buf), foreverTTL); err != nil {
return err
}
metafora.Debugf("task %s submitted: %s", task.ID(), fullpath)
return nil
}
// watch will return either an etcd Response or an error. Two errors returned
// by this method should be treated specially:
//
// 1. etcd.ErrWatchStoppedByUser - the coordinator has closed, exit
// accordingly
//
// 2. restartWatchError - the specified index is too old, try again with a
// newer index
func (ec *EtcdCoordinator) watch(c *etcd.Client, path string, index uint64, stop chan bool) (*etcd.Response, error) {
const recursive = true
for {
// Start the blocking watch after the last response's index.
rawResp, err := protectedRawWatch(c, path, index+1, recursive, nil, stop)
if err != nil {
if err == etcd.ErrWatchStoppedByUser {
// This isn't actually an error, the stop chan was closed. Time to stop!
return nil, err
}
// This is probably a canceled request panic
// Wait a little bit, then continue as normal
// Can be removed after Go 1.5 is released
if ispanic(err) {
time.Sleep(250 * time.Millisecond)
continue
}
// Other RawWatch errors should be retried forever. If the node refresher
// also fails to communicate with etcd it will close the coordinator,
// closing ec.stop in the process which will cause this function to with
// ErrWatchStoppedByUser.
metafora.Errorf("%s Retrying after unexpected watch error: %v", path, err)
transport.CloseIdleConnections() // paranoia; let's get fresh connections on errors.
continue
}
if len(rawResp.Body) == 0 {
// This is a bug in Go's HTTP + go-etcd + etcd which causes the
// connection to timeout perdiocally and need to be restarted *after*
// closing idle connections.
transport.CloseIdleConnections()
continue
}
resp, err := rawResp.Unmarshal()
if err != nil {
if ee, ok := err.(*etcd.EtcdError); ok {
if ee.ErrorCode == EcodeExpiredIndex {
metafora.Debugf("%s Too many events have happened since index was updated. Restarting watch.", ec.taskPath)
// We need to retrieve all existing tasks to update our index
// without potentially missing some events.
return nil, restartWatchError
}
}
metafora.Errorf("%s Unexpected error unmarshalling etcd response: %+v", ec.taskPath, err)
return nil, err
}
return resp, nil
}
}
// SubmitCommand creates a new command for a particular nodeId, the
// command has a random name and is added to the particular nodeId
// directory in etcd.
func (mc *mclient) SubmitCommand(node string, command metafora.Command) error {
cmdPath := mc.cmdPath(node)
body, err := command.Marshal()
if err != nil {
// This is either a bug in metafora or someone implemented their own
// command incorrectly.
return err
}
if _, err := mc.etcd.AddChild(cmdPath, string(body), foreverTTL); err != nil {
metafora.Errorf("Error submitting command: %s to node: %s", command, node)
return err
}
metafora.Debugf("Submitted command: %s to node: %s", string(body), node)
return nil
}
// apply a message to cause a state transition. Returns false if the state
// transition is invalid.
func apply(cur *State, m *Message) (*State, bool) {
//XXX Is a linear scan of all rules really the best option here?
for _, trans := range Rules {
if trans.Event == m.Code && trans.From == cur.Code {
metafora.Debugf("Transitioned %s", trans)
if m.Err != nil {
// Append errors from message
cur.Errors = append(cur.Errors, Err{Time: time.Now(), Err: m.Err.Error()})
}
// New State + Message's Until + Combined Errors
return &State{Code: trans.To, Until: m.Until, Errors: cur.Errors}, true
}
}
return cur, false
}
// Run the state machine enabled handler. Loads the initial state and passes
// control to the internal stateful handler passing commands from the command
// listener into the handler's commands chan.
func (s *stateMachine) Run() (done bool) {
// Multiplex external (Stop) messages and internal ones
s.cmds = make(chan *Message)
go func() {
for {
select {
case m := <-s.cl.Receive():
if !m.Valid() {
metafora.Warnf("Ignoring invalid command: %q", m)
continue
}
select {
case s.cmds <- m:
case <-s.stopped:
return
}
case <-s.stopped:
return
}
}
}()
// Stop the command listener and internal message multiplexer when Run exits
defer func() {
s.cl.Stop()
s.stop()
}()
tid := s.task.ID()
// Load the initial state
state, err := s.ss.Load(s.task)
if err != nil {
// A failure to load the state for a task is *fatal* - the task will be
// unscheduled and requires operator intervention to reschedule.
metafora.Errorf("task=%q could not load initial state. Marking done! Error: %v", tid, err)
return true
}
if state == nil {
// Note to StateStore implementors: This should not happen! Either state or
// err must be non-nil. This code is simply to prevent a nil pointer panic.
metafora.Errorf("statestore %T returned nil state and err for task=%q - unscheduling")
return true
}
if state.Code.Terminal() {
metafora.Warnf("task=%q in terminal state %s - exiting.", tid, state.Code)
return true
}
s.setState(state) // for introspection/debugging
// Main Statemachine Loop
done = false
for {
// Enter State
metafora.Debugf("task=%q in state %s", tid, state.Code)
msg := s.exec(state)
// Apply Message
newstate, ok := apply(state, msg)
if !ok {
metafora.Warnf("task=%q Invalid state transition=%q returned by task. Old state=%q", tid, msg.Code, state.Code)
msg = ErrorMessage(err)
if newstate, ok = apply(state, msg); !ok {
metafora.Errorf("task=%q Unable to transition to error state! Exiting with state=%q", tid, state.Code)
return state.Code.Terminal()
}
}
metafora.Infof("task=%q transitioning %s --> %s --> %s", tid, state, msg, newstate)
// Save state
if err := s.ss.Store(s.task, newstate); err != nil {
metafora.Errorf("task=%q Unable to persist state=%q. Unscheduling.", tid, newstate.Code)
return true
}
// Set next state and loop if non-terminal
state = newstate
// Expose the state for introspection
s.setState(state)
// Exit and unschedule task on terminal state.
if state.Code.Terminal() {
return true
}
// Release messages indicate the task should exit but not unschedule.
if msg.Code == Release {
return false
}
// Alternatively Stop() may have been called but the handler may not have
// returned the Release message. Always exit if we've been told to Stop()
// even if the handler has returned a different Message.
select {
case <-s.stopped:
return false
default:
}
}
}
// add starts refreshing a given key+value pair for a task asynchronously.
func (m *taskManager) add(task metafora.Task) bool {
tid := task.ID()
// Attempt to claim the node
key, value := m.ownerNode(tid)
resp, err := m.client.Create(key, value, m.ttl)
if err != nil {
etcdErr, ok := err.(*etcd.EtcdError)
if !ok || etcdErr.ErrorCode != EcodeNodeExist {
metafora.Errorf("Claim of %s failed with an unexpected error: %v", key, err)
} else {
metafora.Debugf("Claim of %s failed, already claimed", key)
}
return false
}
index := resp.Node.CreatedIndex
// lytics/metafora#124 - the successful create above may have resurrected a
// deleted (done) task. Compare the CreatedIndex of the directory with the
// CreatedIndex of the claim key, if they're equal this claim ressurected a
// done task and should cleanup.
resp, err = m.client.Get(m.taskPath(tid), unsorted, notrecursive)
if err != nil {
// Erroring here is BAD as we may have resurrected a done task, and because
// of this failure there's no way to tell. The claim will eventually
// timeout and the task will get reclaimed.
metafora.Errorf("Error retrieving task path %q after claiming %q: %v", m.taskPath(tid), tid, err)
return false
}
if resp.Node.CreatedIndex == index {
metafora.Debugf("Task %s resurrected due to claim/done race. Re-deleting.", tid)
if _, err = m.client.Delete(m.taskPath(tid), recursive); err != nil {
// This is as bad as it gets. We *know* we resurrected a task, but we
// failed to re-delete it.
metafora.Errorf("Task %s was resurrected and could not be removed! %s should be manually removed. Error: %v",
tid, m.taskPath(tid), err)
}
// Regardless of whether or not the delete succeeded, never treat
// resurrected tasks as claimed.
return false
}
// Claim successful, start the refresher
metafora.Debugf("Claim successful: %s", key)
done := make(chan struct{})
release := make(chan struct{})
finished := make(chan struct{})
m.taskL.Lock()
m.tasks[tid] = taskStates{done: done, release: release, finished: finished}
m.taskL.Unlock()
metafora.Debugf("Starting claim refresher for task %s", tid)
go func() {
defer func() {
m.taskL.Lock()
delete(m.tasks, tid)
m.taskL.Unlock()
close(finished)
}()
for {
select {
case <-time.After(m.interval):
// Try to refresh the claim node (0 index means compare by value)
if _, err := m.client.CompareAndSwap(key, value, m.ttl, value, 0); err != nil {
metafora.Errorf("Error trying to update task %s ttl: %v", tid, err)
m.ctx.Lost(task)
// On errors, don't even try to Delete as we're in a bad state
return
}
case <-done:
metafora.Debugf("Deleting directory for task %s as it's done.", tid)
const recursive = true
if _, err := m.client.Delete(m.taskPath(tid), recursive); err != nil {
metafora.Errorf("Error deleting task %s while stopping: %v", tid, err)
}
return
case <-release:
metafora.Debugf("Deleting claim for task %s as it's released.", tid)
// Not done, releasing; just delete the claim node
if _, err := m.client.CompareAndDelete(key, value, 0); err != nil {
metafora.Warnf("Error releasing task %s while stopping: %v", tid, err)
}
return
}
}
}()
return true
}
// Delete a task
func (mc *mclient) DeleteTask(taskId string) error {
fullpath := mc.tskPath(taskId)
_, err := mc.etcd.Delete(fullpath, recursive)
metafora.Debugf("task %s deleted: %s", taskId, fullpath)
return err
}
// Fair balancer shouldn't consider a shutting-down node
// See https://github.com/lytics/metafora/issues/92
func TestFairBalancerShutdown(t *testing.T) {
coord1, conf1 := setupEtcd(t)
conf2 := conf1.Copy()
conf2.Name = "node2"
coord2, _ := NewEtcdCoordinator(conf2)
cli := NewClient(conf1.Namespace, conf1.Hosts)
// This handler always returns immediately
h1 := metafora.SimpleHandler(func(task metafora.Task, stop <-chan bool) bool {
metafora.Debugf("H1 Starting %s", task.ID())
<-stop
metafora.Debugf("H1 Stopping %s", task.ID())
return false // never done
})
// Block forever on a single task
stop2 := make(chan struct{})
stopr := make(chan chan struct{}, 1)
stopr <- stop2
h2 := metafora.SimpleHandler(func(task metafora.Task, stop <-chan bool) bool {
metafora.Debugf("H2 Starting %s", task.ID())
blockchan, ok := <-stopr
if ok {
<-blockchan
}
<-stop
metafora.Debugf("H2 Stopping %s", task.ID())
return false // never done
})
// Create two consumers
b1 := NewFairBalancer(conf1)
con1, err := metafora.NewConsumer(coord1, h1, b1)
if err != nil {
t.Fatal(err)
}
b2 := NewFairBalancer(conf2)
con2, err := metafora.NewConsumer(coord2, h2, b2)
if err != nil {
t.Fatal(err)
}
// Start the first and let it claim a bunch of tasks
go con1.Run()
defer con1.Shutdown()
cli.SubmitTask(DefaultTaskFunc("t1", ""))
cli.SubmitTask(DefaultTaskFunc("t2", ""))
cli.SubmitTask(DefaultTaskFunc("t3", ""))
cli.SubmitTask(DefaultTaskFunc("t4", ""))
cli.SubmitTask(DefaultTaskFunc("t5", ""))
cli.SubmitTask(DefaultTaskFunc("t6", ""))
time.Sleep(500 * time.Millisecond)
if len(con1.Tasks()) != 6 {
t.Fatalf("con1 should have claimed 6 tasks: %d", len(con1.Tasks()))
}
// Start the second consumer and force the 1st to rebalance
go con2.Run()
close(stopr)
// Wait for node to startup and register
time.Sleep(500 * time.Millisecond)
cli.SubmitCommand(conf1.Name, metafora.CommandBalance())
time.Sleep(2 * time.Second)
c1Tasks := con1.Tasks()
c2Tasks := con2.Tasks()
if len(c1Tasks) != 4 || len(c2Tasks) != 2 {
t.Fatalf("expected consumers to have 4|2 tasks: %d|%d", len(c1Tasks), len(c2Tasks))
}
// Make sure that balancing the other node does nothing
cli.SubmitCommand("node2", metafora.CommandBalance())
time.Sleep(2 * time.Second)
c1Tasks2 := con1.Tasks()
c2Tasks2 := con2.Tasks()
if len(c1Tasks2) != 4 || len(c2Tasks2) != 2 {
t.Fatalf("expected consumers to have 4|2 tasks: %d|%d", len(c1Tasks2), len(c2Tasks2))
}
for i := 0; i < 4; i++ {
if c1Tasks[i] != c1Tasks2[i] {
t.Errorf("task mismatch: %s != %s", c1Tasks[i], c1Tasks2[i])
}
}
for i := 0; i < 2; i++ {
if c2Tasks[i] != c2Tasks2[i] {
t.Errorf("task mismatch: %s != %s", c2Tasks[i], c2Tasks2[i])
}
}
// Second consumer should block on a single task forever
// Rebalancing the first node should then cause it to pickup all but
// one task
c2stop := make(chan struct{})
go func() {
con2.Shutdown()
close(c2stop)
}()
time.Sleep(500 * time.Millisecond)
cli.SubmitCommand(conf1.Name, metafora.CommandBalance())
time.Sleep(2 * time.Second)
c1Tasks3 := con1.Tasks()
c2Tasks3 := con2.Tasks()
if len(c1Tasks3) != 5 || len(c2Tasks3) != 1 {
t.Fatalf("Expected consumers to have 5|1 tasks: %d|%d", len(c1Tasks3), len(c2Tasks3))
}
// Now stop blocking task, rebalance and make sure the first node picked up the remaining
close(stop2)
time.Sleep(500 * time.Millisecond)
// Consumer 2 should stop now
<-c2stop
cli.SubmitCommand(conf1.Name, metafora.CommandBalance())
time.Sleep(2 * time.Second)
// con2 is out of the picture. con1 has all the tasks.
c1Tasks4 := con1.Tasks()
c2Tasks4 := con2.Tasks()
if len(c1Tasks4) != 6 || len(c2Tasks4) != 0 {
t.Fatalf("Expected consumers to have 6|0 tasks: %d|%d", len(c1Tasks4), len(c2Tasks4))
}
}
func TestFairBalancer(t *testing.T) {
t.Parallel()
coord1, conf1 := setupEtcd(t)
conf2 := conf1.Copy()
conf2.Name = "coord2"
coord2, _ := NewEtcdCoordinator(conf2)
cli := NewClient(conf1.Namespace, conf1.Hosts)
h := metafora.SimpleHandler(func(task metafora.Task, stop <-chan bool) bool {
metafora.Debugf("Starting %s", task.ID())
<-stop
metafora.Debugf("Stopping %s", task.ID())
return false // never done
})
// Create two consumers
b1 := NewFairBalancer(conf1)
con1, err := metafora.NewConsumer(coord1, h, b1)
if err != nil {
t.Fatal(err)
}
b2 := NewFairBalancer(conf2)
con2, err := metafora.NewConsumer(coord2, h, b2)
if err != nil {
t.Fatal(err)
}
// Start the first and let it claim a bunch of tasks
go con1.Run()
defer con1.Shutdown()
cli.SubmitTask(DefaultTaskFunc("t1", ""))
cli.SubmitTask(DefaultTaskFunc("t2", ""))
cli.SubmitTask(DefaultTaskFunc("t3", ""))
cli.SubmitTask(DefaultTaskFunc("t4", ""))
cli.SubmitTask(DefaultTaskFunc("t5", ""))
cli.SubmitTask(DefaultTaskFunc("t6", ""))
time.Sleep(1 * time.Second)
if len(con1.Tasks()) != 6 {
t.Fatalf("con1 should have claimed 6 tasks: %d", len(con1.Tasks()))
}
// Start the second consumer and force the 1st to rebalance
go con2.Run()
defer con2.Shutdown()
// Wait for node to startup and register
time.Sleep(1 * time.Second)
cli.SubmitCommand(conf1.Name, metafora.CommandBalance())
time.Sleep(2 * time.Second)
c1Tasks := con1.Tasks()
c2Tasks := con2.Tasks()
if len(c1Tasks) != 4 || len(c2Tasks) != 2 {
t.Fatalf("expected consumers to have 4|2 tasks: %d|%d", len(c1Tasks), len(c2Tasks))
}
// Finally make sure that balancing the other node does nothing
cli.SubmitCommand("node2", metafora.CommandBalance())
time.Sleep(2 * time.Second)
c1Tasks2 := con1.Tasks()
c2Tasks2 := con2.Tasks()
if len(c1Tasks2) != 4 || len(c2Tasks2) != 2 {
t.Fatalf("expected consumers to have 4|2 tasks: %d|%d", len(c1Tasks2), len(c2Tasks2))
}
for i := 0; i < 4; i++ {
if c1Tasks[i] != c1Tasks2[i] {
t.Errorf("task mismatch: %s != %s", c1Tasks[i], c1Tasks2[i])
}
}
for i := 0; i < 2; i++ {
if c2Tasks[i] != c2Tasks2[i] {
t.Errorf("task mismatch: %s != %s", c2Tasks[i], c2Tasks2[i])
}
}
}
func (c *cmdrListener) watcher() {
var index uint64
var ok bool
startWatch:
resp, err := c.cli.Get(c.path, notrecursive, unsorted)
if err != nil {
if ee, ok := err.(*etcd.EtcdError); ok && ee.ErrorCode == EcodeKeyNotFound {
// No command found; this is normal. Grab index and skip to watching
index = ee.Index
goto watchLoop
}
metafora.Errorf("Error GETting %s - sending error to stateful handler: %v", c.path, err)
c.sendErr(err)
return
}
if index, ok = c.sendMsg(resp); !ok {
return
}
watchLoop:
for {
rr, err := protectedRawWatch(c.cli, c.path, index, notrecursive, nil, c.stop)
if err != nil {
if err == etcd.ErrWatchStoppedByUser {
return
}
// This is probably a canceled request panic
// Wait a little bit, then continue as normal
// Can be removed after Go 1.5 is released
if ispanic(err) {
continue
}
metafora.Errorf("Error watching %s - sending error to stateful handler: %v", c.path, err)
c.sendErr(err)
return
}
if len(rr.Body) == 0 {
// This is a bug in Go's HTTP + go-etcd + etcd which causes the
// connection to timeout perdiocally and need to be restarted *after*
// closing idle connections.
transport.CloseIdleConnections()
continue watchLoop
}
resp, err := rr.Unmarshal()
if err != nil {
if ee, ok := err.(*etcd.EtcdError); ok {
if ee.ErrorCode == EcodeExpiredIndex {
goto startWatch
}
}
metafora.Errorf("Error watching %s - sending error to stateful handler: %v", c.path, err)
c.sendErr(err)
return
}
metafora.Debugf("Received command via %s -- sending to statemachine", c.path)
if index, ok = c.sendMsg(resp); !ok {
return
}
}
}