func (self *NsqdCoordinator) SetChannelConsumeOffsetToCluster(ch *nsqd.Channel, queueOffset int64, cnt int64, force bool) error {
topicName := ch.GetTopicName()
partition := ch.GetTopicPart()
coord, checkErr := self.getTopicCoord(topicName, partition)
if checkErr != nil {
return checkErr.ToErrorType()
}
var syncOffset ChannelConsumerOffset
syncOffset.AllowBackward = true
syncOffset.VCnt = cnt
syncOffset.VOffset = queueOffset
doLocalWrite := func(d *coordData) *CoordErr {
err := ch.SetConsumeOffset(nsqd.BackendOffset(queueOffset), cnt, force)
if err != nil {
if err != nsqd.ErrSetConsumeOffsetNotFirstClient {
coordLog.Infof("failed to set the consume offset: %v, err:%v", queueOffset, err)
return &CoordErr{err.Error(), RpcNoErr, CoordLocalErr}
}
coordLog.Debugf("the consume offset: %v can only be set by the first client", queueOffset)
return ErrLocalSetChannelOffsetNotFirstClient
}
return nil
}
doLocalExit := func(err *CoordErr) {}
doLocalCommit := func() error {
return nil
}
doLocalRollback := func() {}
doRefresh := func(d *coordData) *CoordErr {
return nil
}
doSlaveSync := func(c *NsqdRpcClient, nodeID string, tcData *coordData) *CoordErr {
if ch.IsEphemeral() {
return nil
}
rpcErr := c.UpdateChannelOffset(&tcData.topicLeaderSession, &tcData.topicInfo, ch.GetName(), syncOffset)
if rpcErr != nil {
coordLog.Infof("sync channel(%v) offset to replica %v failed: %v, offset: %v", ch.GetName(),
nodeID, rpcErr, syncOffset)
}
return rpcErr
}
handleSyncResult := func(successNum int, tcData *coordData) bool {
if successNum == len(tcData.topicInfo.ISR) {
return true
}
return false
}
clusterErr := self.doSyncOpToCluster(false, coord, doLocalWrite, doLocalExit, doLocalCommit, doLocalRollback,
doRefresh, doSlaveSync, handleSyncResult)
if clusterErr != nil {
return clusterErr.ToErrorType()
}
return nil
}
func (self *NsqdCoordinator) updateChannelOffsetOnSlave(tc *coordData, channelName string, offset ChannelConsumerOffset) *CoordErr {
topicName := tc.topicInfo.Name
partition := tc.topicInfo.Partition
if !tc.IsMineISR(self.myNode.GetID()) {
return ErrTopicWriteOnNonISR
}
if coordLog.Level() >= levellogger.LOG_DETAIL {
coordLog.Debugf("got update channel(%v) offset on slave : %v", channelName, offset)
}
coord, coordErr := self.getTopicCoord(topicName, partition)
if coordErr != nil {
return ErrMissingTopicCoord
}
topic, localErr := self.localNsqd.GetExistingTopic(topicName, partition)
if localErr != nil {
coordLog.Warningf("slave missing topic : %v", topicName)
// TODO: leave the isr and try re-sync with leader
return &CoordErr{localErr.Error(), RpcCommonErr, CoordSlaveErr}
}
if topic.GetTopicPart() != partition {
coordLog.Errorf("topic on slave has different partition : %v vs %v", topic.GetTopicPart(), partition)
return ErrLocalMissingTopic
}
var ch *nsqd.Channel
ch, localErr = topic.GetExistingChannel(channelName)
// if a new channel on slave, we should set the consume offset by force
if localErr != nil {
offset.AllowBackward = true
ch = topic.GetChannel(channelName)
coordLog.Infof("slave init the channel : %v, %v, offset: %v", topic.GetTopicName(), channelName, ch.GetConfirmed())
}
if ch.IsEphemeral() {
coordLog.Errorf("ephemeral channel %v should not be synced on slave", channelName)
}
currentEnd := ch.GetChannelEnd()
if nsqd.BackendOffset(offset.VOffset) > currentEnd.Offset() {
coordLog.Debugf("update channel(%v) consume offset exceed end %v on slave : %v", channelName, offset, currentEnd)
// cache the offset (using map?) to reduce the slave channel flush.
coord.consumeMgr.Lock()
cur, ok := coord.consumeMgr.channelConsumeOffset[channelName]
if !ok || cur.VOffset < offset.VOffset {
coord.consumeMgr.channelConsumeOffset[channelName] = offset
}
coord.consumeMgr.Unlock()
if offset.Flush {
topic.ForceFlush()
currentEnd = ch.GetChannelEnd()
if nsqd.BackendOffset(offset.VOffset) > currentEnd.Offset() {
offset.VOffset = int64(currentEnd.Offset())
offset.VCnt = currentEnd.TotalMsgCnt()
}
} else {
return nil
}
}
err := ch.ConfirmBackendQueueOnSlave(nsqd.BackendOffset(offset.VOffset), offset.VCnt, offset.AllowBackward)
if err != nil {
coordLog.Warningf("update local channel(%v) offset %v failed: %v, current channel end: %v, topic end: %v",
channelName, offset, err, currentEnd, topic.TotalDataSize())
if err == nsqd.ErrExiting {
return &CoordErr{err.Error(), RpcNoErr, CoordTmpErr}
}
return &CoordErr{err.Error(), RpcCommonErr, CoordSlaveErr}
}
return nil
}
func (self *NsqdCoordinator) FinishMessageToCluster(channel *nsqd.Channel, clientID int64, clientAddr string, msgID nsqd.MessageID) error {
topicName := channel.GetTopicName()
partition := channel.GetTopicPart()
coord, checkErr := self.getTopicCoord(topicName, partition)
if checkErr != nil {
return checkErr.ToErrorType()
}
var syncOffset ChannelConsumerOffset
changed := false
var confirmed nsqd.BackendQueueEnd
if channel.IsOrdered() {
if !coord.GetData().IsISRReadyForWrite() {
coordLog.Warningf("topic(%v) finish message ordered failed since no enough ISR", topicName)
coordErrStats.incWriteErr(ErrWriteQuorumFailed)
return ErrWriteQuorumFailed.ToErrorType()
}
confirmed = channel.GetConfirmed()
}
// TODO: maybe use channel to aggregate all the sync of message to reduce the rpc call.
doLocalWrite := func(d *coordData) *CoordErr {
offset, cnt, tmpChanged, localErr := channel.FinishMessage(clientID, clientAddr, msgID)
if localErr != nil {
coordLog.Infof("channel %v finish local msg %v error: %v", channel.GetName(), msgID, localErr)
changed = false
return &CoordErr{localErr.Error(), RpcNoErr, CoordLocalErr}
}
changed = tmpChanged
syncOffset.VOffset = int64(offset)
syncOffset.VCnt = cnt
return nil
}
doLocalExit := func(err *CoordErr) {}
doLocalCommit := func() error {
channel.ContinueConsumeForOrder()
return nil
}
doLocalRollback := func() {
if channel.IsOrdered() && confirmed != nil {
coordLog.Warningf("rollback channel confirm to : %v", confirmed)
// reset read to last confirmed
channel.SetConsumeOffset(confirmed.Offset(), confirmed.TotalMsgCnt(), true)
}
}
doRefresh := func(d *coordData) *CoordErr {
return nil
}
doSlaveSync := func(c *NsqdRpcClient, nodeID string, tcData *coordData) *CoordErr {
if !changed || channel.IsEphemeral() {
return nil
}
var rpcErr *CoordErr
if channel.IsOrdered() {
// if ordered, we need make sure all the consume offset is synced to all replicas
rpcErr = c.UpdateChannelOffset(&tcData.topicLeaderSession, &tcData.topicInfo, channel.GetName(), syncOffset)
} else {
c.NotifyUpdateChannelOffset(&tcData.topicLeaderSession, &tcData.topicInfo, channel.GetName(), syncOffset)
}
if rpcErr != nil {
coordLog.Infof("sync channel(%v) offset to replica %v failed: %v, offset: %v", channel.GetName(),
nodeID, rpcErr, syncOffset)
}
return rpcErr
}
handleSyncResult := func(successNum int, tcData *coordData) bool {
// we can ignore the error if this channel is not ordered. (just sync next time)
if successNum == len(tcData.topicInfo.ISR) || !channel.IsOrdered() {
return true
}
return false
}
clusterErr := self.doSyncOpToCluster(false, coord, doLocalWrite, doLocalExit, doLocalCommit, doLocalRollback,
doRefresh, doSlaveSync, handleSyncResult)
if clusterErr != nil {
return clusterErr.ToErrorType()
}
return nil
}