// handleMaster performs one loop of master locking.
// on success it returns <master>, nil
// on error it returns "", err
// in situations where you should try again due to concurrent state changes (e.g. another actor simultaneously acquiring the lock)
// it returns "", nil
func (e *etcdMasterElector) handleMaster(path, id string, ttl uint64) (string, error) {
res, err := e.etcd.Get(path, false, false)
// Unexpected error, bail out
if err != nil && !etcdstorage.IsEtcdNotFound(err) {
return "", err
}
// There is no master, try to become the master.
if err != nil && etcdstorage.IsEtcdNotFound(err) {
return e.becomeMaster(path, id, ttl)
}
// This should never happen.
if res.Node == nil {
return "", fmt.Errorf("unexpected response: %#v", res)
}
// We're not the master, just return the current value
if res.Node.Value != id {
return res.Node.Value, nil
}
// We are the master, try to extend out lease
return e.extendMaster(path, id, ttl, res)
}
// acquireOrRenewLease either races to acquire a new master lease, or update the existing master's lease
// returns true if we have the lease, and an error if one occurs.
// TODO: use the master election utility once it is merged in.
func (c *config) acquireOrRenewLease(etcdClient *etcd.Client) (bool, error) {
result, err := etcdClient.Get(c.key, false, false)
if err != nil {
if etcdstorage.IsEtcdNotFound(err) {
// there is no current master, try to become master, create will fail if the key already exists
_, err := etcdClient.Create(c.key, c.whoami, c.ttl)
if err != nil {
return false, err
}
c.lastLease = time.Now()
return true, nil
}
return false, err
}
if result.Node.Value == c.whoami {
glog.Infof("key already exists, we are the master (%s)", result.Node.Value)
// we extend our lease @ 1/2 of the existing TTL, this ensures the master doesn't flap around
if result.Node.Expiration.Sub(time.Now()) < time.Duration(c.ttl/2)*time.Second {
_, err := etcdClient.CompareAndSwap(c.key, c.whoami, c.ttl, c.whoami, result.Node.ModifiedIndex)
if err != nil {
return false, err
}
}
c.lastLease = time.Now()
return true, nil
}
glog.Infof("key already exists, the master is %s, sleeping.", result.Node.Value)
return false, nil
}
// InterpretDeleteError converts a generic etcd error on a delete
// operation into the appropriate API error.
func InterpretDeleteError(err error, kind, name string) error {
switch {
case etcdstorage.IsEtcdNotFound(err):
return errors.NewNotFound(kind, name)
default:
return err
}
}
func (s *SchedulerServer) fetchFrameworkID(client tools.EtcdClient) (*mesos.FrameworkID, error) {
if s.FailoverTimeout > 0 {
if response, err := client.Get(meta.FrameworkIDKey, false, false); err != nil {
if !etcdstorage.IsEtcdNotFound(err) {
return nil, fmt.Errorf("unexpected failure attempting to load framework ID from etcd: %v", err)
}
log.V(1).Infof("did not find framework ID in etcd")
} else if response.Node.Value != "" {
log.Infof("configuring FrameworkInfo with Id found in etcd: '%s'", response.Node.Value)
return mutil.NewFrameworkID(response.Node.Value), nil
}
} else {
//TODO(jdef) this seems like a totally hackish way to clean up the framework ID
if _, err := client.Delete(meta.FrameworkIDKey, true); err != nil {
if !etcdstorage.IsEtcdNotFound(err) {
return nil, fmt.Errorf("failed to delete framework ID from etcd: %v", err)
}
log.V(1).Infof("nothing to delete: did not find framework ID in etcd")
}
}
return nil, nil
}
// Refresh reloads the RangeAllocation from etcd.
func (e *Etcd) Refresh() (*api.RangeAllocation, error) {
e.lock.Lock()
defer e.lock.Unlock()
existing := &api.RangeAllocation{}
if err := e.storage.Get(e.baseKey, existing, false); err != nil {
if etcdstorage.IsEtcdNotFound(err) {
return nil, nil
}
return nil, etcderr.InterpretGetError(err, e.kind, "")
}
return existing, nil
}