func (self *NsqdCoordinator) PutMessagesToCluster(topic *nsqd.Topic,
msgs []*nsqd.Message) (nsqd.MessageID, nsqd.BackendOffset, int32, error) {
var commitLog CommitLogData
topicName := topic.GetTopicName()
partition := topic.GetTopicPart()
coord, checkErr := self.getTopicCoord(topicName, partition)
if checkErr != nil {
return nsqd.MessageID(commitLog.LogID), nsqd.BackendOffset(commitLog.MsgOffset), commitLog.MsgSize, checkErr.ToErrorType()
}
var queueEnd nsqd.BackendQueueEnd
var logMgr *TopicCommitLogMgr
doLocalWrite := func(d *coordData) *CoordErr {
topic.Lock()
logMgr = d.logMgr
id, offset, writeBytes, totalCnt, qe, localErr := topic.PutMessagesNoLock(msgs)
queueEnd = qe
topic.Unlock()
if localErr != nil {
coordLog.Warningf("put batch messages 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 = int64(msgs[len(msgs)-1].ID)
commitLog.MsgOffset = int64(offset)
commitLog.MsgSize = writeBytes
// This MsgCnt is the total count until now (include the current written batch message count)
commitLog.MsgCnt = totalCnt
commitLog.MsgNum = int32(len(msgs))
return nil
}
doLocalExit := func(err *CoordErr) {
if err != nil {
coordLog.Infof("topic %v PutMessagesToCluster error: %v", topic.GetFullName(), 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.ResetBackendEndNoLock(nsqd.BackendOffset(commitLog.MsgOffset), commitLog.MsgCnt-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.PutMessages(&tcData.topicLeaderSession, &tcData.topicInfo, commitLog, msgs)
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 nsqd.MessageID(commitLog.LogID), nsqd.BackendOffset(commitLog.MsgOffset), commitLog.MsgSize, err
}