Replica {                    // a replica of a partition
  broker_id   : int
  partition   : Partition
  isLocal     : Boolean      // is this replica local to this broker
  log         : Log          // local log associated with this replica
  hw          : long         // offset of the last committed message
  leo         : long         // log end offset
}

Partition {                         //a partition in a topic
  topic            : string
  partition_id     : int
  leader           : Replica        // the leader replica of this partition
  commitQ          : Queue          // produce requests pending commit at the leader
  AR               : Set[Replica]   // replicas assigned to this partition
  ISR              : Set[Replica]   // In-sync replica set, maintained at the leader
  CUR              : Set[Replica]   // Catch-up replica set, maintained at the leader
  RAR              : Set[Replica]   // Reassigned replica set, maintained at the leader
}

brokerStartup()
{
  create /brokers/state/[broker_id] path if it doesn’t already exist
  register the state change handler to listen on child change ZK notifications on /brokers/state/[broker_id]
  register session expiration listener
  drain the state change queue 
  get replica info from ZK and compute AR, a list of replicas assigned to this broker
  for each r in AR
  {
      subscribe to leader changes for the r’s partition
      startReplica(r)
  }
  // broker startup procedure is complete. Register is broker id in ZK to announce the availability of this broker
  register its broker_id in /brokers/ids/[broker_id] in ZK
}

leaderElection(r: Replica)
  read the current ISR and AR for r.partition.partition_id from ZK
  if( (r in AR) && (ISR is empty || r in ISR) ) 
  {
     wait for PreferredReplicaTime if r is not the preferred replica
     if(successfully write r as the current leader of r.partition in ZK)
       becomeLeader(r, ISR, CUR)
  }
} 

onLeaderChange() 
{
   if(broker_id is registered under /brokers/topics/[topic]/[partition_id]/replicas)
     leaderElection(r)
}

stateChangeListener() {  
// listens to state change requests issued by the leader and acts on those

    drain the state change queue
    read next state change request
    Let r be the replica that the state change request is sent for.
    // this should not happen
    Throw an error if r is not hosted on this broker                
    Let requestEpoch be the epoch of the state change request
    if(closeReplicaRequest) 
    { 
	 closeReplica(r)
      // we don’t need to check epoch here to be able to handle delete topic/delete partition for dead brokers.
    }
    if(startReplicaRequest) 
    { 
       Let latestPartitionEpoch be the latest epoch for this partition, got by reading /brokers/topics/[topic]/[partition]/leader=[broker_id],[epoch] 
       if(leader for r.partition doesn’t exist) {
	   // this can only happen for new topics or new partitions for existing topics
         startReplica(r)         
       }else if(requestEpoch == latestPartitionEpoch) {
         // this is to ensure that if a follower is slow, and reads a state change request queued up by a previous leader, it ignores the request
	    startReplica(r)
       }
    }
}

onPartitionsReassigned() 
{
  if(this broker is the leader for [partition_id])
  {
  	p.RAR = the new replicas from /brokers/partitions_reassigned/[topic]/[partition_id]
     AR = /brokers/topics/[topic]/[partition_id]/replicas
     newReplicas = p.RAR - AR
     for(newReplica <- newReplicas)
       sendStateChange(“start-replica”, newReplica.broker_id, epoch)
  }
}

sendStateChange(stateChange, followerBrokerId, leaderEpoch)
{
   stateChangeQ = new StateChangeQueue(“/brokers/state/followerBrokerId”)
   stateChangeRequest = new StateChangeRequest(stateChange, leaderEpoch)
   // check if the state change Q is full. This can happen if a broker is offline for a long time
   if(stateChangeQ.isFull) {
      // this operation retains only one close-replica request for a partition, the one with the latest epoch. This is to ensure that an offline broker, on startup, will delete old topics and partitions, which it hosted before going offline. You don’t have to retain any start-replica requests for a partition
      stateChangeQ.shrink
      // if the queue is still full, log an error
      throw new FollowerStateChangeQueueFull
   }
   stateChangeQ.put(stateChangeRequest)
} 

startReplica(r: Replica) {
  if(broker_id not in /brokers/topics/[r.topic]/[r.partition]/replicas)
    throw NotReplicaForPartitionException()
  if( r's log is not already started) {
     do local recovery of r's log
     r.hw = min(last checkpointed HW for r, r.leo)
     register a leader-change listener on partition r.partition.partition_id
  }
   if( a leader does not exist for partition r.partition.partition_id in ZK)
      leaderElection(r)
   else {
     if(this broker is not the leader, then it is a follower since it is in the AR list for this partition)
      becomeFollower(r)
   }
}

closeReplica(r: Replica)
{
   stop the fetcher associated with r, if one exists
   close and delete r
}

becomeFollower(r: Replica)
{
  // this is required if this replica was the last leader
  stop the commit thread, if any                  
  stop the current ReplicaFetcherThread, if any
  truncate the log to r.hw
  start a new ReplicaFetcherThread to the current leader of r, from offset r.leo
  start HW checkpoint thread for r
}

becomeLeader(r: Replica, ISR: Set[Replica], AR: Set[Replica])
{
   // get a new epoch value and write it to the leader path
   epoch = getNewEpoch()
   /brokers/topics/[r.partition.topic]/[r.partition.pid]/leader=broker_id, epoch
   stop HW checkpoint thread for r
   r.hw = r.leo // TODO: check if this should actually be r.hw = last checkpointed HW for r
   wait until every live replica in AR catches up (i.e. its leo == r.hw) or a KeepInSyncTime has passed
   r.partition.ISR = the current set of replicas in sync with r
   r.partition.CUR = AR - ISR
   write r.partition.ISR & r.partition.CUR in ZK
   r.partition.RAR = replicas in /brokers/partitions_reassigned/[topic]/[partition_id] in ZK
   r.partition.leader = r                  // this enables reads/writes to this partition on this broker
   start a commit thread on r.partition
   start HW checkpoint thread for r
    get replica info from ZK and compute AR, a list of replicas assigned to this broker
    for each r in AR {
       if(r.broker_id is registered under /brokers/ids)
          sendStateChange(“start-replica”, r.broker_id, epoch)
    }
}

createTopic(topic, numPartitions, replicationFactor, replicaAssignmentStr) 
{
   if(!cleanFailedTopicCreationAttempt(topic))
   {
      error(“Topic topic exists with live partitions”)
      exit
   }  
   if(replicaAssignmentStr == “”) {
     // assignReplicas will always assign partitions only to online brokers
     replicaAssignment = assignReplicas(topic, numPartitions, replicationFactor)
   }

   // create topic path in ZK
   create /brokers/topics/topic
   for(partition <- replicaAssignment) {
     addPartition(topic, partition.id, partition.replicas) 
   }
   // report successfully started partitions for this topic
}
waitTillStateChangeRequestConsumed(partition.replicas, timeout)
{
   register watch on state change path for each replica
   In the listener, use a condition variable to await(timeout). If it doesn’t fire return false, else return true
}

cleanFailedTopicCreationAttempts(topic) 
{
   topicsForPartitionsReassignment = ls /brokers/partitions_reassigned
   for(topic <- topicsForPartitionsReassignment)
   {
      partitionsCreated = ls /brokers/topics/topic
      for(partition <- partitionsCreated) 
      {
		if(/brokers/topics/topic/partition/replicas path exists)
          {
             error(“Cannot cleanup. Topic exists with live partition”)
             return false 
          }
      }
	 // partition paths can be safely deleted
      for(partition <- partitionsCreated)
         delete /brokers/topics/topic/partition
      delete /brokers/topics/topic
   } 
}

deleteTopic(topic) 
{
   partitionsForTopic = ls /brokers/topics/topic
   for(partition <- partitionsForTopic) {
      if(!deletePartition(topic, partition))
      {
         error(“Failed to delete partition for topic”)
         exit 
      }
   }
   // delete topic path in ZK
   delete /brokers/topics/topic
}

addPartition(topic, partition, replicas) 
{
     // write the partitions reassigned path for this create topic command
     /brokers/partitions_reassigned/topic/partition=replicas
	// start replicas for this new partition
     for(replica <- replicas)
        sendStateChange(“start-replica”, replica.brokerId, -1)
     // wait till state change request is consumed by all replicas
     if(!waitTillStateChangeRequestConsumed(partition.replicas, timeout)) 
     {
        error(“Failed to create topic partition partitionId for timeout ms”)
        exit
     }
     // create partition paths in ZK
     /brokers/topics/topic/partitionId/replicas=replicas
     delete /brokers/partitions_reassigned/topic/partitionId
}

deletePartition(topic, partition) 
{
   // empty list for partition reassignment means delete partition
   /brokers/partitions_reassigned/topic/partition=””
   // wait till replica is closed by all replicas
   if(!waitTillStateChangeRequestConsumed(partition.replicas, timeout)) 
   {
     error(“Failed to delete topic after timeout ms”)
     return false
   }
   // create partition paths in ZK
   delete /brokers/topics/topic/partitionId
   delete /brokers/partitions_reassigned/topic/partitionId
}

produceRequestHandler(pr: ProduceRequest)
{
   if( the request partition pr.partition doesn't have leader replica on this broker)
      throw NotLeaderException
   log = r.partition.leader.log
   append pr.messages to log
   pr.offset = log.LEO
   add pr to pr.partition.commitQ
}

while(true) {
  pr = commitQ.dequeue
  canCommit = false
  while(!canCommit) {
    canCommit = true
    for each r in ISR
       if(!offsetReached(r, pr.offset)) {
	canCommit = false
	break
       }
   if(!canCommit) {
      p.CUR.add(r)
      p.ISR.delete(r)
      write p.ISR to ZK
   }
  }
  for each c in CUR
      if(c.leo >= pr.offset) {
	p.ISR.add(c); p.CUR.delete(c); write p.ISR to ZK
      }
      checkReassignedReplicas(pr, p.RAR, p.ISR)
      checkLoadBalancing()
       r.hw = pr.offset         // increment the HW to indicate that pr is committed
      send ACK to the client that pr is committed
}

offsetReached(r: Replica, offset: Long) {
   if(r.leo becomes equal or larger than offset within KeepInSyncTime)  return true
   return false
}

checkLoadBalancing() { // see if we need to switch the leader to the preferred replica
   if(leader replica is not the preferred one & the preferred replica is in ISR) {
       delete /brokers/topics/[topic]/[partition_id]/leader  in ZK
       stop this commit thread
       stop the HW checkpoint thread
   }
 }

checkReassignedReplicas(pr: ProduceRequest, RAR: Set[Replica], ISR: Set[Replica])

{
    // see if all reassigned replicas have fully caught up and older replicas have stopped fetching, if so, switch to those replicas

    // optimization, do the check periodically

    If (every replica in RAR has its leo >= pr.offset) {
       if(!sentCloseReplica.get) {
		oldReplicas = AR - RAR
          for(oldReplica <- oldReplicas) {
             if(r.broker_id != broker_id)
               sendStateChange(“close-replica”, oldReplica.broker_id, epoch)
          }
          sentCloseReplica.set(true)
       }else {
         // close replica is already sent. Wait until the replicas are closed or probably timeout and raise error
          if(broker_id is in (AR - RAR) && (other replicas in (AR - RAR) are not in ISR anymore)) {
            // leader is not in the reassigned replicas list
             completePartitionReassignment(RAR, ISR, AR, true)
	        sentCloseReplica.set(false)            
          }
        	else if(every replica in (AR-RAR) is not in ISR anymore) {
             completePartitionReassignment(RAR, ISR, AR, false)
	        sentCloseReplica.set(false)
	     }
       }
         
}

completePartitionsReassignment(RAR: Set[Replica], ISR: Set[Replica], AR: Set[Replica], stopCommitThread: Boolean)
{
    //newly assigned replicas are in-sync, switch over to the new replicas
    //need (RAR + ISR) in case we fail right after here

    write (RAR + ISR) as the new ISR in ZK     
    update /brokers/topics/[topic]/[partition_id]/replicas in ZK with the new replicas in RAR
//triggers leader election
    delete /brokers/partitions_reassigned/[topic]/[partition_id] in ZK
    if(stopCommitThread || (broker_id is not preferred replica))
    {
      delete /brokers/topics/[topic]/[partition_id]/leader in ZK 
      stop this commit thread
    }
}

ReplicaFetchReqeust {
   topic: String
   partition_id: Int
   replica_id: Int
   offset: Long
}

ReplicaFetchResponse {
   hw: Long			// the offset of the last message committed at the leader
   messages: MessageSet	// fetched messages
}

replicaFetch (f: ReplicaFetchRequest) {	// handler for ReplicaFetchRequest at leader
  leader = getLeaderReplica(f.topic, f.partition_id)

  if(leader == null) throw NotLeaderException
   response = new ReplicaFetcherResponse
   getReplica(f.topic, f.partition_id, f.replica_id).leo = f.offset
    response.messages = fetch messages starting from f.offset from leader.log
    response.hw = leader.hw
    send response back
}

while(true) {
   send ReplicaFetchRequest to leader and get response:ReplicaFetcherResponse back
   append response.messages to r's log
   r.hw = response.hw
   advance offset in ReplicaFetchRequest
}