func (self *NsqdCoordinator) putMessageOnSlave(coord *TopicCoordinator, logData CommitLogData, msg *nsqd.Message) *CoordErr {
var logMgr *TopicCommitLogMgr
var topic *nsqd.Topic
var queueEnd nsqd.BackendQueueEnd
checkDupOnSlave := func(tc *coordData) bool {
if coordLog.Level() >= levellogger.LOG_DETAIL {
topicName := tc.topicInfo.Name
coordLog.Debugf("pub on slave : %v, msg %v", topicName, msg.ID)
}
logMgr = tc.logMgr
if logMgr.IsCommitted(logData.LogID) {
coordLog.Infof("pub the already committed log id : %v", logData.LogID)
return true
}
return false
}
doLocalWriteOnSlave := func(tc *coordData) *CoordErr {
var localErr error
topicName := tc.topicInfo.Name
partition := tc.topicInfo.Partition
topic, localErr = self.localNsqd.GetExistingTopic(topicName, partition)
if localErr != nil {
coordLog.Infof("pub on slave missing topic : %v", topicName)
// leave the isr and try re-sync with leader
return &CoordErr{localErr.Error(), RpcErrTopicNotExist, CoordSlaveErr}
}
if topic.GetTopicPart() != partition {
coordLog.Errorf("topic on slave has different partition : %v vs %v", topic.GetTopicPart(), partition)
return &CoordErr{ErrLocalTopicPartitionMismatch.String(), RpcErrTopicNotExist, CoordSlaveErr}
}
topic.Lock()
queueEnd, localErr = topic.PutMessageOnReplica(msg, nsqd.BackendOffset(logData.MsgOffset))
topic.Unlock()
if localErr != nil {
coordLog.Errorf("put message on slave failed: %v", localErr)
return &CoordErr{localErr.Error(), RpcCommonErr, CoordSlaveErr}
}
return nil
}
doLocalCommit := func() error {
localErr := logMgr.AppendCommitLog(&logData, true)
if localErr != nil {
coordLog.Errorf("write commit log on slave failed: %v", localErr)
return localErr
}
topic.Lock()
topic.UpdateCommittedOffset(queueEnd)
topic.Unlock()
return nil
}
doLocalExit := func(err *CoordErr) {
if err != nil {
coordLog.Infof("slave put message %v error: %v", logData, err)
}
}
return self.doWriteOpOnSlave(coord, checkDupOnSlave, doLocalWriteOnSlave, doLocalCommit, doLocalExit)
}
func (self *NsqdCoordinator) putMessagesOnSlave(coord *TopicCoordinator, logData CommitLogData, msgs []*nsqd.Message) *CoordErr {
if len(msgs) == 0 {
return ErrPubArgError
}
if logData.LogID != int64(msgs[0].ID) {
return ErrPubArgError
}
var logMgr *TopicCommitLogMgr
// this last log id should be used on slave to avoid the slave switch
// override the leader's prev mpub message id.
// While slave is chosen as leader, the next id should be larger than the last logid.
// Because the mpub maybe already committed after the leader is down, the new leader should begin
// with the last message id + 1 for next message.
lastMsgLogID := int64(msgs[len(msgs)-1].ID)
if logData.LastMsgLogID != lastMsgLogID {
return ErrPubArgError
}
var queueEnd nsqd.BackendQueueEnd
var topic *nsqd.Topic
checkDupOnSlave := func(tc *coordData) bool {
if coordLog.Level() >= levellogger.LOG_DETAIL {
topicName := tc.topicInfo.Name
coordLog.Debugf("pub on slave : %v, msg count: %v", topicName, len(msgs))
}
logMgr = tc.logMgr
if logMgr.IsCommitted(logData.LogID) {
coordLog.Infof("put the already committed log id : %v", logData.LogID)
return true
}
return false
}
doLocalWriteOnSlave := func(tc *coordData) *CoordErr {
var localErr error
var start time.Time
checkCost := coordLog.Level() >= levellogger.LOG_DEBUG
if self.enableBenchCost {
checkCost = true
}
if checkCost {
start = time.Now()
}
topicName := tc.topicInfo.Name
partition := tc.topicInfo.Partition
topic, localErr = self.localNsqd.GetExistingTopic(topicName, partition)
if localErr != nil {
coordLog.Infof("pub on slave missing topic : %v", topicName)
// leave the isr and try re-sync with leader
return &CoordErr{localErr.Error(), RpcErrTopicNotExist, CoordSlaveErr}
}
topic.Lock()
var cost time.Duration
if checkCost {
cost = time.Now().Sub(start)
if cost > time.Millisecond {
coordLog.Infof("prepare write on slave local cost :%v", cost)
}
}
queueEnd, localErr = topic.PutMessagesOnReplica(msgs, nsqd.BackendOffset(logData.MsgOffset))
if checkCost {
cost2 := time.Now().Sub(start)
if cost2 > time.Millisecond {
coordLog.Infof("write local on slave cost :%v, %v", cost, cost2)
}
}
topic.Unlock()
if localErr != nil {
logIndex, lastLogOffset, lastLog, _ := logMgr.GetLastCommitLogOffsetV2()
coordLog.Errorf("put messages on slave failed: %v, slave last logid: %v, data: %v:%v, %v",
localErr, logMgr.GetLastCommitLogID(), logIndex, lastLogOffset, lastLog)
return &CoordErr{localErr.Error(), RpcCommonErr, CoordSlaveErr}
}
return nil
}
doLocalCommit := func() error {
localErr := logMgr.AppendCommitLog(&logData, true)
if localErr != nil {
coordLog.Errorf("write commit log on slave failed: %v", localErr)
return localErr
}
topic.Lock()
topic.UpdateCommittedOffset(queueEnd)
topic.Unlock()
return nil
}
doLocalExit := func(err *CoordErr) {
if err != nil {
coordLog.Warningf("failed to batch put messages on slave: %v", err)
}
}
return self.doWriteOpOnSlave(coord, checkDupOnSlave, doLocalWriteOnSlave, doLocalCommit,
doLocalExit)
}
func (self *NsqdCoordinator) PutMessageToCluster(topic *nsqd.Topic,
body []byte, traceID uint64) (nsqd.MessageID, nsqd.BackendOffset, int32, nsqd.BackendQueueEnd, error) {
var commitLog CommitLogData
var queueEnd nsqd.BackendQueueEnd
msg := nsqd.NewMessage(0, body)
msg.TraceID = traceID
topicName := topic.GetTopicName()
partition := topic.GetTopicPart()
coord, checkErr := self.getTopicCoord(topicName, partition)
if checkErr != nil {
return msg.ID, nsqd.BackendOffset(commitLog.MsgOffset), commitLog.MsgSize, queueEnd, checkErr.ToErrorType()
}
var logMgr *TopicCommitLogMgr
doLocalWrite := func(d *coordData) *CoordErr {
logMgr = d.logMgr
topic.Lock()
id, offset, writeBytes, qe, localErr := topic.PutMessageNoLock(msg)
queueEnd = qe
topic.Unlock()
if localErr != nil {
coordLog.Warningf("put message to local failed: %v", localErr)
return &CoordErr{localErr.Error(), RpcNoErr, CoordLocalErr}
}
commitLog.LogID = int64(id)
// epoch should not be changed.
// leader epoch change means leadership change, leadership change
// need disable write which should hold the write lock.
// However, we are holding write lock while doing the cluster write replication.
commitLog.Epoch = d.GetTopicEpochForWrite()
commitLog.LastMsgLogID = commitLog.LogID
commitLog.MsgOffset = int64(offset)
commitLog.MsgSize = writeBytes
commitLog.MsgCnt = queueEnd.TotalMsgCnt()
commitLog.MsgNum = 1
return nil
}
doLocalExit := func(err *CoordErr) {
if err != nil {
coordLog.Infof("topic %v PutMessageToCluster msg %v error: %v", topic.GetFullName(), msg, err)
if coord.IsWriteDisabled() {
topic.DisableForSlave()
}
}
}
doLocalCommit := func() error {
localErr := logMgr.AppendCommitLog(&commitLog, false)
if localErr != nil {
coordLog.Errorf("topic : %v, Generator %v failed write commit log : %v, logmgr: %v, %v",
topic.GetFullName(), topic.GetMsgGenerator(), localErr, logMgr.pLogID, logMgr.nLogID)
}
topic.Lock()
topic.UpdateCommittedOffset(queueEnd)
topic.Unlock()
return localErr
}
doLocalRollback := func() {
coordLog.Warningf("failed write begin rollback : %v, %v", topic.GetFullName(), commitLog)
topic.Lock()
topic.RollbackNoLock(nsqd.BackendOffset(commitLog.MsgOffset), 1)
topic.Unlock()
}
doRefresh := func(d *coordData) *CoordErr {
logMgr = d.logMgr
if d.GetTopicEpochForWrite() != commitLog.Epoch {
coordLog.Warningf("write epoch changed during write: %v, %v", d.GetTopicEpochForWrite(), commitLog)
return ErrEpochMismatch
}
self.requestNotifyNewTopicInfo(d.topicInfo.Name, d.topicInfo.Partition)
return nil
}
doSlaveSync := func(c *NsqdRpcClient, nodeID string, tcData *coordData) *CoordErr {
// should retry if failed, and the slave should keep the last success write to avoid the duplicated
putErr := c.PutMessage(&tcData.topicLeaderSession, &tcData.topicInfo, commitLog, msg)
if putErr != nil {
coordLog.Infof("sync write to replica %v failed: %v. put offset:%v, logmgr: %v, %v",
nodeID, putErr, commitLog, logMgr.pLogID, logMgr.nLogID)
}
return putErr
}
handleSyncResult := func(successNum int, tcData *coordData) bool {
if successNum == len(tcData.topicInfo.ISR) {
return true
}
return false
}
clusterErr := self.doSyncOpToCluster(true, coord, doLocalWrite, doLocalExit, doLocalCommit, doLocalRollback,
doRefresh, doSlaveSync, handleSyncResult)
var err error
if clusterErr != nil {
err = clusterErr.ToErrorType()
}
return msg.ID, nsqd.BackendOffset(commitLog.MsgOffset), commitLog.MsgSize, queueEnd, err
}