apply.md

Raft Apply

Apply is the primary operation provided by raft. A client calls raft.Apply to apply a command to the FSM. A command will first be committed, i.e., durably stored on a quorum of raft nodes. Then, the committed command is applied to fsm.

This sequence diagram shows the steps involved in a raft.Apply operation. Each box across the top is a separate thread. The name in the box identifies the state of the peer (leader or follower) and the thread (<peer state>:<thread name>). When there are multiple copies of the thread, it is indicated with (each peer).

sequenceDiagram
   autonumber
 
   participant client
   participant leadermain as leader:main
   participant leaderfsm as leader:fsm
   participant leaderreplicate as leader:replicate (each peer)
   participant followermain as follower:main (each peer)
   participant followerfsm as follower:fsm (each peer)
 
   client-)leadermain: applyCh to dispatchLogs
   leadermain->>leadermain: store logs to disk
 
   leadermain-)leaderreplicate: triggerCh
   leaderreplicate-->>followermain: Transport.AppendEntries RPC
 
   followermain->>followermain: store logs to disk
 
   opt leader commit index is ahead of peer commit index
       followermain-)followerfsm: fsmMutateCh <br>apply committed logs
       followerfsm->>followerfsm: fsm.Apply
   end
 
   followermain-->>leaderreplicate: respond success=true
   leaderreplicate->>leaderreplicate: update commitment
 
   opt quorum commit index has increased
       leaderreplicate-)leadermain: commitCh
       leadermain-)leaderfsm: fsmMutateCh
       leaderfsm->>leaderfsm: fsm.Apply
       leaderfsm-)client: future.respond
   end

Loading

Following is the description of each step as shown in the above diagram

1. The raft node handles the raft.Apply call by creating a new log entry and send the entry to the applyCh channel.

2. If the node is not a leader, the method will return an error of ErrNotLeader. Otherwise, the main loop of the leader node calls raft.dispatchLogs to write the log entry locally.

3. raft.dispatchLogs also sends a notification to the f.triggerCh of each follower (map[ServerID]*followerReplication) to start replicating log entries to the followers.

4. For each follower, the leader has started a long running routine (replicate) to replicates log entries. On receiving a log entry to the triggerCh, the replicate routine makes the Transport.AppendEntries RPC call to do the replication. The log entries to be replicated are from the follower's nextIndex to min(nextIndex + maxAppendEntries, leader's lastIndex). Another parameter to AppendEntries is the LeaderCommitIndex. Following is some examples:

AppendEntries(Log: 1..5, LeaderCommitIndex: 0)   // Replicating log entries 1..5, 
                                                 // the leader hasn't committed any log entry;
AppendEntries(Log: 6..8, LeaderCommitIndex: 4)   // Replicating log entries 6..8,
                                                 // log 0..4 are committed after the leader receives
                                                 // a quorum of responses
AppendEntries(Log: 9, LeaderCommitIndex: 8)      // Replicating log entry 9,
                                                 // log 5..8 are committed.
AppendEntries(Log: , LeaderCommitIndex: 9)       // no new log, bumping the commit index
                                                 // to let the follower stay up to date of the
                                                 // latest committed entries

5. The follower which receives the appendEntries RPC calls invokes raft.appendEntries to handle the request. It appends any new entries to the local log store.

6. In the same method on the follower as step 5, if the LeaderCommitIndex > this follower's commitIndex, the follower updates it's commitIndex to min(LeaderCommitIndex, index of its last log entries). In the first AppendEntries call of the above example, the follower won't update it's commitIndex, because LeaderCommitIndex is 0. The last RPC call doesn't contain any new log, whereas the follower will update its commitIndex to 9.

Further, the follower start processLogs to send all the committed entries that haven't been applied to fsm (fsmMutateCh <- batch). Otherwise (i.e., commitIndex <= lastApplied), the appendEntries RPC call returns success.

Therefore, it's possible that a very small window of time exists when all followers have committed the log to disk, the write has been realized in the FSM of the leader but the followers have not yet applied the log to their FSM.

7. The peer applies the committed entries to the FSM.

8. If all went well, the follower responds success (resp.Success = true) to the appendEntries RPC call.

9. On receiving the successful response from Transport.AppendEntries, the leader needs to update the fsm based on the replicated log entries. Specifically, the leader finds the highest log entry index that has been replicated to a quorum of the servers ( if quorumMatchIndex > c.commitIndex), update commitIndex to that index, and notify through the commitCh channel.

10. The leader receives the notification on the r.leaderState.commitCh channel and starts grouping the entries that can be applied to the fsm.

11. processLogs applies all the committed entries that haven't been applied by batching the log entries and forwarding them through the fsmMutateCh channel to fsm.

12. The actual place applying the committed log entries is in the main loop of runFSM().

13. After the log entries that contains the client req are applied to the fsm, the fsm module will set the responses to the client request (req.future.respond(nil)). From the client's point of view, the future returned by raft.Apply should now be unblocked and calls to Error() or Response() should return the data at this point.