// nodeRefresher is in charge of keeping the node entry in etcd alive. If it's
// unable to communicate with etcd it must shutdown the coordinator.
//
// watch retries on errors and taskmgr calls Lost(task) on tasks it can't
// refresh, so it's up to nodeRefresher to cause the coordinator to close if
// it's unable to communicate with etcd.
func (ec *EtcdCoordinator) nodeRefresher() {
ttl := ec.conf.NodeTTL >> 1 // have some leeway before ttl expires
if ttl < 1 {
metafora.Warnf("%s Dangerously low NodeTTL: %d", ec.name, ec.conf.NodeTTL)
ttl = 1
}
// Create a local etcd client since it's not threadsafe, but don't bother
// checking for errors at this point.
client, _ := newEtcdClient(ec.conf.Hosts)
for {
// Deadline for refreshes to finish by or the coordinator closes.
deadline := time.Now().Add(time.Duration(ec.conf.NodeTTL) * time.Second)
select {
case <-ec.stop:
return
case <-time.After(time.Duration(ttl) * time.Second):
if err := ec.refreshBy(client, deadline); err != nil {
// We're in a bad state; shut everything down
metafora.Errorf("Unable to refresh node key before deadline %s. Last error: %v", deadline, err)
ec.Close()
}
}
}
}
func (e *etcdClusterState) NodeTaskCount() (map[string]int, error) {
state := map[string]int{}
// First initialize state with nodes as keys
resp, err := e.client.Get(e.nodePath, unsorted, recursive)
if err != nil {
return nil, err
}
if resp == nil || resp.Node == nil {
metafora.Warnf("balancer received empty response from GET %s", e.nodePath)
return state, nil
}
for _, node := range resp.Node.Nodes {
state[path.Base(node.Key)] = 0
}
// Then count how many tasks each node has
resp, err = e.client.Get(e.taskPath, unsorted, recursive)
if err != nil {
return nil, err
}
// No current tasks
if resp == nil {
return state, nil
}
// Get the list of all claimed work, create a map of the counts and
// node values
// We ignore tasks which have no claims
for _, task := range resp.Node.Nodes {
for _, claim := range task.Nodes {
if path.Base(claim.Key) == OwnerMarker {
val := ownerValue{}
if err := json.Unmarshal([]byte(claim.Value), &val); err == nil {
// We want to only include those nodes which were initially included,
// as some nodes may be shutting down, etc, and should not be counted
_, ok := state[val.Node]
if ok {
state[val.Node]++
}
}
}
}
}
return state, nil
}
func newManager(ctx metafora.CoordinatorContext, client client, path, nodeID string, ttl uint64) *taskManager {
if ttl == 0 {
panic("refresher: TTL must be > 0")
}
// refresh more often than strictly necessary to be safe
interval := time.Duration((ttl>>1)+(ttl>>2)) * time.Second
if ttl == 1 {
interval = 750 * time.Millisecond
metafora.Warnf("Dangerously low TTL: %d; consider raising.", ttl)
}
return &taskManager{
ctx: ctx,
client: client,
tasks: make(map[string]taskStates),
path: path,
node: nodeID,
ttl: ttl,
interval: interval,
}
}
// refreshBy retries refreshing the node key until the deadline is reached.
func (ec *EtcdCoordinator) refreshBy(c *etcd.Client, deadline time.Time) (err error) {
for time.Now().Before(deadline) {
// Make sure we shouldn't exit
select {
case <-ec.stop:
return err
default:
}
_, err = c.UpdateDir(ec.nodePath, ec.conf.NodeTTL)
if err == nil {
// It worked!
return nil
}
metafora.Warnf("Unexpected error updating node key: %v", err)
transport.CloseIdleConnections() // paranoia; let's get fresh connections on errors.
time.Sleep(500 * time.Millisecond) // rate limit retries a bit
}
// Didn't get a successful response before deadline, exit with error
return err
}
// execute non-terminal states
func (s *stateMachine) exec(state *State) *Message {
switch state.Code {
case Runnable:
// Runnable passes control to the stateful handler
return run(s.h, s.task, s.cmds)
case Paused:
// Paused until a message arrives
return <-s.cmds
case Sleeping:
// Sleeping until the specified time (or a message)
if state.Until == nil {
metafora.Warnf("task=%q told to sleep without a time. Resuming.", s.task.ID())
return RunMessage()
}
dur := state.Until.Sub(time.Now())
metafora.Infof("task=%q sleeping for %s", s.task.ID(), dur)
timer := time.NewTimer(dur)
select {
case <-timer.C:
return RunMessage()
case msg := <-s.cmds:
timer.Stop()
// Checkpoint & Release are special cases that shouldn't affect sleep
// time, so maintain it across the state transition
if msg.Code == Checkpoint || msg.Code == Release {
msg.Until = state.Until
}
return msg
}
case Fault:
// Special case where we potentially trim the current state to keep
// errors from growing without bound.
var msg *Message
msg, state.Errors = s.errHandler(s.task, state.Errors)
return msg
default:
panic("invalid state: " + state.String())
}
}
// 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
}
func main() {
mlvl := metafora.LogLevelInfo
hostname, _ := os.Hostname()
peers := flag.String("etcd", "http://127.0.0.1:2379", "comma delimited etcd peer list")
namespace := flag.String("namespace", "koalemos", "metafora namespace")
name := flag.String("name", hostname, "node name or empty for automatic")
loglvl := flag.String("log", mlvl.String(), "set log level: [debug], info, warn, error")
flag.Parse()
hosts := strings.Split(*peers, ",")
etcdc := etcd.NewClient(hosts)
switch strings.ToLower(*loglvl) {
case "debug":
mlvl = metafora.LogLevelDebug
case "info":
mlvl = metafora.LogLevelInfo
case "warn":
mlvl = metafora.LogLevelWarn
case "error":
mlvl = metafora.LogLevelError
default:
metafora.Warnf("Invalid log level %q - using %s", *loglvl, mlvl)
}
metafora.SetLogLevel(mlvl)
conf := m_etcd.NewConfig(*name, *namespace, hosts)
// Replace NewTask func with one that returns a *koalemos.Task
conf.NewTaskFunc = func(id, value string) metafora.Task {
t := koalemos.NewTask(id)
if value == "" {
return t
}
if err := json.Unmarshal([]byte(value), t); err != nil {
metafora.Errorf("Unable to unmarshal task %s: %v", t.ID(), err)
return nil
}
return t
}
hfunc := makeHandlerFunc(etcdc)
ec, err := m_etcd.NewEtcdCoordinator(conf)
if err != nil {
metafora.Errorf("Error creating etcd coordinator: %v", err)
}
bal := m_etcd.NewFairBalancer(conf)
c, err := metafora.NewConsumer(ec, hfunc, bal)
if err != nil {
metafora.Errorf("Error creating consumer: %v", err)
os.Exit(2)
}
metafora.Infof(
"Starting koalsmosd with etcd=%s; namespace=%s; name=%s; loglvl=%s",
*peers, conf.Namespace, conf.Name, mlvl)
consumerRunning := make(chan struct{})
go func() {
defer close(consumerRunning)
c.Run()
}()
sigC := make(chan os.Signal, 1)
signal.Notify(sigC, os.Interrupt, os.Kill, syscall.SIGTERM)
select {
case s := <-sigC:
metafora.Infof("Received signal %s, shutting down", s)
case <-consumerRunning:
metafora.Warn("Consumer exited. Shutting down.")
}
c.Shutdown()
metafora.Info("Shutdown")
}