k3s/vendor/go.etcd.io/etcd/raft/design.md
2019-12-15 23:28:25 -07:00

5.5 KiB
Raw Blame History

Progress

Progress represents a followers progress in the view of the leader. Leader maintains progresses of all followers, and sends replication message to the follower based on its progress.

replication message is a msgApp with log entries.

A progress has two attribute: match and next. match is the index of the highest known matched entry. If leader knows nothing about followers replication status, match is set to zero. next is the index of the first entry that will be replicated to the follower. Leader puts entries from next to its latest one in next replication message.

A progress is in one of the three state: probe, replicate, snapshot.

                            +--------------------------------------------------------+          
                            |                  send snapshot                         |          
                            |                                                        |          
                  +---------+----------+                                  +----------v---------+
              +--->       probe        |                                  |      snapshot      |
              |   |  max inflight = 1  <----------------------------------+  max inflight = 0  |
              |   +---------+----------+                                  +--------------------+
              |             |            1. snapshot success                                    
              |             |               (next=snapshot.index + 1)                           
              |             |            2. snapshot failure                                    
              |             |               (no change)                                         
              |             |            3. receives msgAppResp(rej=false&&index>lastsnap.index)
              |             |               (match=m.index,next=match+1)                        
receives msgAppResp(rej=true)                                                                   
(next=match+1)|             |                                                                   
              |             |                                                                   
              |             |                                                                   
              |             |   receives msgAppResp(rej=false&&index>match)                     
              |             |   (match=m.index,next=match+1)                                    
              |             |                                                                   
              |             |                                                                   
              |             |                                                                   
              |   +---------v----------+                                                        
              |   |     replicate      |                                                        
              +---+  max inflight = n  |                                                        
                  +--------------------+                                                        

When the progress of a follower is in probe state, leader sends at most one replication message per heartbeat interval. The leader sends replication message slowly and probing the actual progress of the follower. A msgHeartbeatResp or a msgAppResp with reject might trigger the sending of the next replication message.

When the progress of a follower is in replicate state, leader sends replication message, then optimistically increases next to the latest entry sent. This is an optimized state for fast replicating log entries to the follower.

When the progress of a follower is in snapshot state, leader stops sending any replication message.

A newly elected leader sets the progresses of all the followers to probe state with match = 0 and next = last index. The leader slowly (at most once per heartbeat) sends replication message to the follower and probes its progress.

A progress changes to replicate when the follower replies with a non-rejection msgAppResp, which implies that it has matched the index sent. At this point, leader starts to stream log entries to the follower fast. The progress will fall back to probe when the follower replies a rejection msgAppResp or the link layer reports the follower is unreachable. We aggressively reset next to match+1 since if we receive any msgAppResp soon, both match and next will increase directly to the index in msgAppResp. (We might end up with sending some duplicate entries when aggressively reset next too low. see open question)

A progress changes from probe to snapshot when the follower falls very far behind and requires a snapshot. After sending msgSnap, the leader waits until the success, failure or abortion of the previous snapshot sent. The progress will go back to probe after the sending result is applied.

Flow Control

  1. limit the max size of message sent per message. Max should be configurable. Lower the cost at probing state as we limit the size per message; lower the penalty when aggressively decreased to a too low next

  2. limit the # of in flight messages < N when in replicate state. N should be configurable. Most implementation will have a sending buffer on top of its actual network transport layer (not blocking raft node). We want to make sure raft does not overflow that buffer, which can cause message dropping and triggering a bunch of unnecessary resending repeatedly.