func (db *ETCDDB) batchDeleteNodes(keys []string, logger lager.Logger) {
if len(keys) == 0 {
return
}
works := []func(){}
for _, key := range keys {
key := key
works = append(works, func() {
logger.Info("deleting", lager.Data{"key": key})
_, err := db.client.Delete(key, true)
if err != nil {
logger.Error("failed-to-delete", err, lager.Data{
"key": key,
})
}
})
}
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
logger.Error("failed-to-create-throttler", err)
}
throttler.Work()
return
}
func (db *ETCDDB) batchDeleteTasks(taskGuids []string, logger lager.Logger) {
if len(taskGuids) == 0 {
return
}
works := []func(){}
for _, taskGuid := range taskGuids {
taskGuid := taskGuid
works = append(works, func() {
_, err := db.client.Delete(taskGuid, true)
if err != nil {
logger.Error("failed-to-delete", err, lager.Data{
"task_guid": taskGuid,
})
}
})
}
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
logger.Error("failed-to-create-throttler", err)
}
throttler.Work()
return
}
func (db *ETCDDB) ResolveConvergence(logger lager.Logger, desiredLRPs map[string]*models.DesiredLRP, changes *models.ConvergenceChanges) ([]*auctioneer.LRPStartRequest, []*models.ActualLRPKeyWithSchedulingInfo, []*models.ActualLRPKey) {
startRequests := newStartRequests(desiredLRPs)
for _, actual := range changes.StaleUnclaimedActualLRPs {
startRequests.Add(logger, &actual.ActualLRPKey)
}
works := []func(){}
keysToRetire := make([]*models.ActualLRPKey, len(changes.ActualLRPsForExtraIndices))
for i, actual := range changes.ActualLRPsForExtraIndices {
keysToRetire[i] = &actual.ActualLRPKey
}
keysWithMissingCells := []*models.ActualLRPKeyWithSchedulingInfo{}
for _, actual := range changes.ActualLRPsWithMissingCells {
desiredLRP, ok := desiredLRPs[actual.ProcessGuid]
if !ok {
logger.Debug("actual-with-missing-cell-no-desired")
continue
}
schedInfo := desiredLRP.DesiredLRPSchedulingInfo()
key := &models.ActualLRPKeyWithSchedulingInfo{
Key: &actual.ActualLRPKey,
SchedulingInfo: &schedInfo,
}
keysWithMissingCells = append(keysWithMissingCells, key)
}
for _, actualKey := range changes.ActualLRPKeysForMissingIndices {
works = append(works, db.resolveActualsWithMissingIndices(logger, desiredLRPs[actualKey.ProcessGuid], actualKey, startRequests))
}
for _, actual := range changes.RestartableCrashedActualLRPs {
works = append(works, db.resolveRestartableCrashedActualLRPS(logger, actual, startRequests))
}
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
logger.Error("failed-constructing-throttler", err, lager.Data{"max_workers": db.convergenceWorkersSize, "num_works": len(works)})
return nil, nil, nil
}
logger.Debug("waiting-for-lrp-convergence-work")
throttler.Work()
logger.Debug("done-waiting-for-lrp-convergence-work")
return startRequests.Slice(), keysWithMissingCells, keysToRetire
}
func (db *ETCDDB) deleteLeaves(logger lager.Logger, keys []string) error {
works := []func(){}
for _, key := range keys {
key := key
works = append(works, func() {
_, err := db.client.DeleteDir(key)
if err != nil {
logger.Error("failed-deleting-leaf-node", err, lager.Data{"key": key})
}
})
}
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
return err
}
throttler.Work()
return nil
}
func (h *DesiredLRPHandler) createUnclaimedActualLRPs(logger lager.Logger, keys []*models.ActualLRPKey) []int {
count := len(keys)
createdIndicesChan := make(chan int, count)
works := make([]func(), count)
logger = logger.Session("create-unclaimed-actual-lrp")
for i, key := range keys {
key := key
works[i] = func() {
logger.Info("starting", lager.Data{"actual_lrp_key": key})
actualLRPGroup, err := h.actualLRPDB.CreateUnclaimedActualLRP(logger, key)
if err != nil {
logger.Info("failed", lager.Data{"actual_lrp_key": key, "err_message": err.Error()})
} else {
go h.actualHub.Emit(models.NewActualLRPCreatedEvent(actualLRPGroup))
createdIndicesChan <- int(key.Index)
}
}
}
throttlerSize := h.updateWorkersCount
throttler, err := workpool.NewThrottler(throttlerSize, works)
if err != nil {
logger.Error("failed-constructing-throttler", err, lager.Data{"max_workers": throttlerSize, "num_works": len(works)})
return []int{}
}
go func() {
throttler.Work()
close(createdIndicesChan)
}()
createdIndices := make([]int, 0, count)
for createdIndex := range createdIndicesChan {
createdIndices = append(createdIndices, createdIndex)
}
return createdIndices
}
func (db *ETCDDB) batchCompareAndSwapTasks(tasksToCAS []compareAndSwappableTask, logger lager.Logger) error {
if len(tasksToCAS) == 0 {
return nil
}
works := []func(){}
for _, taskToCAS := range tasksToCAS {
task := taskToCAS.NewTask
task.UpdatedAt = db.clock.Now().UnixNano()
value, err := db.serializeModel(logger, task)
if err != nil {
logger.Error("failed-to-marshal", err, lager.Data{
"task_guid": task.TaskGuid,
})
continue
}
index := taskToCAS.OldIndex
works = append(works, func() {
_, err := db.client.CompareAndSwap(TaskSchemaPathByGuid(task.TaskGuid), value, NO_TTL, index)
if err != nil {
logger.Error("failed-to-compare-and-swap", err, lager.Data{
"task_guid": task.TaskGuid,
})
}
})
}
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
return err
}
throttler.Work()
return nil
}
func (db *ETCDDB) GatherAndPruneDesiredLRPs(logger lager.Logger, guids map[string]struct{}, lmc *LRPMetricCounter) (map[string]*models.DesiredLRP, error) {
desiredLRPsRoot, modelErr := db.fetchRecursiveRaw(logger, DesiredLRPComponentsSchemaRoot)
if modelErr == models.ErrResourceNotFound {
logger.Info("actual-lrp-schema-root-not-found")
return map[string]*models.DesiredLRP{}, nil
}
if modelErr != nil {
return nil, modelErr
}
schedulingInfos := map[string]*models.DesiredLRPSchedulingInfo{}
runInfos := map[string]*models.DesiredLRPRunInfo{}
var malformedSchedulingInfos, malformedRunInfos []string
var guidsLock, schedulingInfosLock, runInfosLock sync.Mutex
works := []func(){}
logger.Debug("walking-desired-lrp-components-tree")
for _, componentRoot := range desiredLRPsRoot.Nodes {
switch componentRoot.Key {
case DesiredLRPSchedulingInfoSchemaRoot:
for _, node := range componentRoot.Nodes {
node := node
works = append(works, func() {
var schedulingInfo models.DesiredLRPSchedulingInfo
err := db.deserializeModel(logger, node, &schedulingInfo)
if err != nil || schedulingInfo.Validate() != nil {
logger.Error("failed-to-deserialize-scheduling-info", err)
schedulingInfosLock.Lock()
malformedSchedulingInfos = append(malformedSchedulingInfos, node.Key)
schedulingInfosLock.Unlock()
} else {
schedulingInfosLock.Lock()
schedulingInfos[schedulingInfo.ProcessGuid] = &schedulingInfo
schedulingInfosLock.Unlock()
atomic.AddInt32(&lmc.desiredLRPs, schedulingInfo.Instances)
guidsLock.Lock()
guids[schedulingInfo.ProcessGuid] = struct{}{}
guidsLock.Unlock()
}
})
}
case DesiredLRPRunInfoSchemaRoot:
for _, node := range componentRoot.Nodes {
node := node
works = append(works, func() {
var runInfo models.DesiredLRPRunInfo
err := db.deserializeModel(logger, node, &runInfo)
if err != nil || runInfo.Validate() != nil {
runInfosLock.Lock()
malformedRunInfos = append(malformedRunInfos, node.Key)
runInfosLock.Unlock()
} else {
runInfosLock.Lock()
runInfos[runInfo.ProcessGuid] = &runInfo
runInfosLock.Unlock()
}
})
}
default:
err := fmt.Errorf("unrecognized node under desired LRPs root node: %s", componentRoot.Key)
logger.Error("unrecognized-node", err)
return nil, err
}
}
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
logger.Error("failed-to-create-throttler", err)
}
throttler.Work()
db.batchDeleteNodes(malformedSchedulingInfos, logger)
db.batchDeleteNodes(malformedRunInfos, logger)
malformedSchedulingInfosMetric.Add(uint64(len(malformedSchedulingInfos)))
malformedRunInfosMetric.Add(uint64(len(malformedRunInfos)))
logger.Debug("done-walking-desired-lrp-tree")
desireds := make(map[string]*models.DesiredLRP)
var schedInfosToDelete []string
for guid, schedulingInfo := range schedulingInfos {
runInfo, ok := runInfos[guid]
if !ok {
err := fmt.Errorf("Missing runInfo for GUID %s", guid)
logger.Error("runInfo-not-found-error", err)
schedInfosToDelete = append(schedInfosToDelete, DesiredLRPSchedulingInfoSchemaPath(guid))
} else {
desiredLRP := models.NewDesiredLRP(*schedulingInfo, *runInfo)
desireds[guid] = &desiredLRP
}
}
db.batchDeleteNodes(schedInfosToDelete, logger)
// Check to see if we have orphaned RunInfos
if len(runInfos) != len(schedulingInfos) {
var runInfosToDelete []string
for guid, runInfo := range runInfos {
// If there is no corresponding SchedulingInfo and the RunInfo has
// existed for longer than desiredLRPCreationTimeout, consider it orphaned
// and delete it.
_, ok := schedulingInfos[guid]
if !ok && db.clock.Since(time.Unix(0, runInfo.CreatedAt)) > db.desiredLRPCreationTimeout {
orphanedRunInfosMetric.Add(1)
runInfosToDelete = append(runInfosToDelete, DesiredLRPRunInfoSchemaPath(guid))
}
}
db.batchDeleteNodes(runInfosToDelete, logger)
}
return desireds, nil
}
func (db *ETCDDB) gatherAndOptionallyPruneActualLRPs(logger lager.Logger, guids map[string]struct{}, doPrune bool, lmc *LRPMetricCounter) (map[string]map[int32]*models.ActualLRP, error) {
response, modelErr := db.fetchRecursiveRaw(logger, ActualLRPSchemaRoot)
if modelErr == models.ErrResourceNotFound {
logger.Info("actual-lrp-schema-root-not-found")
return map[string]map[int32]*models.ActualLRP{}, nil
}
if modelErr != nil {
return nil, modelErr
}
actuals := map[string]map[int32]*models.ActualLRP{}
var guidKeysToDelete, indexKeysToDelete []string
var actualsToDelete []string
var guidsLock, actualsLock, guidKeysToDeleteLock, indexKeysToDeleteLock,
crashingDesiredsLock, actualsToDeleteLock sync.Mutex
logger.Debug("walking-actual-lrp-tree")
works := []func(){}
crashingDesireds := map[string]struct{}{}
for _, guidGroup := range response.Nodes {
guidGroup := guidGroup
works = append(works, func() {
guidGroupWillBeEmpty := true
for _, indexGroup := range guidGroup.Nodes {
indexGroupWillBeEmpty := true
for _, actualNode := range indexGroup.Nodes {
actual := new(models.ActualLRP)
err := db.deserializeModel(logger, actualNode, actual)
if err != nil {
actualsToDeleteLock.Lock()
actualsToDelete = append(actualsToDelete, actualNode.Key)
actualsToDeleteLock.Unlock()
continue
}
err = actual.Validate()
if err != nil {
actualsToDeleteLock.Lock()
actualsToDelete = append(actualsToDelete, actualNode.Key)
actualsToDeleteLock.Unlock()
continue
}
indexGroupWillBeEmpty = false
guidGroupWillBeEmpty = false
switch actual.State {
case models.ActualLRPStateUnclaimed:
atomic.AddInt32(&lmc.unclaimedLRPs, 1)
case models.ActualLRPStateClaimed:
atomic.AddInt32(&lmc.claimedLRPs, 1)
case models.ActualLRPStateRunning:
atomic.AddInt32(&lmc.runningLRPs, 1)
case models.ActualLRPStateCrashed:
crashingDesiredsLock.Lock()
crashingDesireds[actual.ProcessGuid] = struct{}{}
crashingDesiredsLock.Unlock()
atomic.AddInt32(&lmc.crashedActualLRPs, 1)
}
guidsLock.Lock()
guids[actual.ProcessGuid] = struct{}{}
guidsLock.Unlock()
if path.Base(actualNode.Key) == ActualLRPInstanceKey {
actualsLock.Lock()
if actuals[actual.ProcessGuid] == nil {
actuals[actual.ProcessGuid] = map[int32]*models.ActualLRP{}
}
actuals[actual.ProcessGuid][actual.Index] = actual
actualsLock.Unlock()
}
}
if indexGroupWillBeEmpty {
indexKeysToDeleteLock.Lock()
indexKeysToDelete = append(indexKeysToDelete, indexGroup.Key)
indexKeysToDeleteLock.Unlock()
}
}
if guidGroupWillBeEmpty {
guidKeysToDeleteLock.Lock()
guidKeysToDelete = append(guidKeysToDelete, guidGroup.Key)
guidKeysToDeleteLock.Unlock()
}
})
}
logger.Debug("done-walking-actual-lrp-tree")
throttler, err := workpool.NewThrottler(db.convergenceWorkersSize, works)
if err != nil {
logger.Error("failed-to-create-throttler", err)
}
throttler.Work()
if doPrune {
logger.Info("deleting-invalid-actual-lrps", lager.Data{"num_lrps": len(actualsToDelete)})
db.batchDeleteNodes(actualsToDelete, logger)
actualLRPsDeleted.Add(uint64(len(actualsToDelete)))
logger.Info("deleting-empty-actual-indices", lager.Data{"num_indices": len(indexKeysToDelete)})
err = db.deleteLeaves(logger, indexKeysToDelete)
if err != nil {
logger.Error("failed-deleting-empty-actual-indices", err, lager.Data{"num_indices": len(indexKeysToDelete)})
} else {
logger.Info("succeeded-deleting-empty-actual-indices", lager.Data{"num_indices": len(indexKeysToDelete)})
}
logger.Info("deleting-empty-actual-guids", lager.Data{"num_guids": len(guidKeysToDelete)})
err = db.deleteLeaves(logger, guidKeysToDelete)
if err != nil {
logger.Error("failed-deleting-empty-actual-guids", err, lager.Data{"num_guids": len(guidKeysToDelete)})
} else {
logger.Info("succeeded-deleting-empty-actual-guids", lager.Data{"num_guids": len(guidKeysToDelete)})
}
}
lmc.crashingDesiredLRPs = int32(len(crashingDesireds))
return actuals, nil
}
func (h *LRPConvergenceController) ConvergeLRPs(logger lager.Logger) error {
logger = h.logger.Session("converge-lrps")
var err error
logger.Debug("listing-cells")
var cellSet models.CellSet
cellSet, err = h.serviceClient.Cells(logger)
if err == models.ErrResourceNotFound {
logger.Info("no-cells-found")
cellSet = models.CellSet{}
} else if err != nil {
logger.Error("failed-listing-cells", err)
// convergence should run again later
return nil
}
logger.Debug("succeeded-listing-cells")
startRequests, keysWithMissingCells, keysToRetire := h.db.ConvergeLRPs(logger, cellSet)
retireLogger := logger.WithData(lager.Data{"retiring_lrp_count": len(keysToRetire)})
works := []func(){}
for _, key := range keysToRetire {
key := key
works = append(works, func() {
h.retirer.RetireActualLRP(retireLogger, key.ProcessGuid, key.Index)
})
}
errChan := make(chan *models.Error, 1)
startRequestLock := &sync.Mutex{}
for _, key := range keysWithMissingCells {
key := key
works = append(works, func() {
before, after, err := h.db.UnclaimActualLRP(logger, key.Key)
if err == nil {
h.actualHub.Emit(models.NewActualLRPChangedEvent(before, after))
startRequest := auctioneer.NewLRPStartRequestFromSchedulingInfo(key.SchedulingInfo, int(key.Key.Index))
startRequestLock.Lock()
startRequests = append(startRequests, &startRequest)
startRequestLock.Unlock()
} else {
bbsErr := models.ConvertError(err)
if bbsErr.GetType() != models.Error_Unrecoverable {
return
}
logger.Error("unrecoverable-error", bbsErr)
select {
case errChan <- bbsErr:
default:
}
}
})
}
var throttler *workpool.Throttler
throttler, err = workpool.NewThrottler(h.convergenceWorkersSize, works)
if err != nil {
logger.Error("failed-constructing-throttler", err, lager.Data{"max_workers": h.convergenceWorkersSize, "num_works": len(works)})
return nil
}
retireLogger.Debug("retiring-actual-lrps")
throttler.Work()
retireLogger.Debug("done-retiring-actual-lrps")
select {
case err := <-errChan:
return err
default:
}
startLogger := logger.WithData(lager.Data{"start_requests_count": len(startRequests)})
if len(startRequests) > 0 {
startLogger.Debug("requesting-start-auctions")
err = h.auctioneerClient.RequestLRPAuctions(logger, startRequests)
if err != nil {
startLogger.Error("failed-to-request-starts", err, lager.Data{"lrp_start_auctions": startRequests})
}
startLogger.Debug("done-requesting-start-auctions")
}
return nil
}
