Skip to content

Latest commit

 

History

History
76 lines (51 loc) · 3.12 KB

TODO.md

File metadata and controls

76 lines (51 loc) · 3.12 KB
Raw

Feature TODO list

Apply committed log

  • In leader node, we should apply log entries immediately after that we check majoritory polls and change committed value.
  • In each applying, we reply ack to client immediately, before processing next log entry.
  • In follower node, we should apply log entries right after commit index change.
  • (DONE)

Log persistence

WAL journal

  • Node can recover from the journal file.
    • WAL journal monotonically increases. The trunking of conflict mem_log should write WAL journal record to mark deletion of trunked log records.
    • Raft role change should write journal to persistent node role and term information.

Leader Election

Make sure the candidate with the latest log elected

  • Need to add check in follower handling on vote request. (DONE)

Client

Client register.

  • Per client command sequence counter should be maintained in STM. (DONE)
  • Reply clients in 2 places:
    1. Before we put client request into the _client_req_queue, reject it because of not leader, or the queue is full.
    2. After applied the command. (DONE)
  • In tick loop, we reply client by put result into the per client ret_queue. (DONE)
  • If leader crushed before reply, new leader node may not have the ret_queue for this client. To solve this, the easiest way is to let client re-send the request.
  • Command sequence counter mechanism makes sure the applied (by the previous term follower, now the new leader) command is dropped. (DONE)

Client Session

  • Add a session mechanism to client. The register command creates a client session. The client session has a idle timeout. After timeout, Raft nodes can recycle the resource allocated to this client.

Client Update Request

Client Query

  • Linearizability of read.
    • See https://aphyr.com/posts/316-call-me-maybe-etcd-and-consul and etcd-io/etcd#741
    • Accumulate read reqs before the leader node make sure it get the qurom, that means to wait for a cycle of heartbeat.
    • Leader -> Follower change should clear accumulated read reqs. It can return error to client with leader hint.
    • (DONE, without returning error to client when leader switch)

Client Side Retry

  • Discover leader according to fixed configurations.
  • Because we maintain the TCP connection after register, we would hit exception if leader crush. Need to handle exception and discover leader again. After finding the new leader, retry request.
  • Also need to set timeout when waiting request reply, as the leader may hang for unknown reason, but without breaking the TCP connection.

Leader Hint

  • Each member should know the leader id of current term.
  • Return leader hint if the node is not leader.
  • (DONE)

State Machine

In Memory State Machine persistence

  • Snapshot/Recovery

SQLite as storage backend

Integration test framework

In memory log optimization

  • Remove old enought log entry from left end.
  • Add index to log entry position to support more efficient access of mem_log.

Message packaging efficiency

  • Refactor message fields to use int not string.