京公网安备11010502063487号follower 落后或者数据不一致时,leader 怎么让它追上?文档说有 DIFF、TRUNC、SNAP 三种。这篇逐一制造:略落后、数据丢光、旧 leader 带着未提交的写回来,看 leader 日志里实际选了哪种。
[DIFF] 落后 10 个事务的 follower 重启:追上用时 570 ms
leader: Synchronizing with Learner sid: 1 maxCommittedLog=0x10000000c minCommittedLog=0x100000001 lastProcessedZxid=0x10000000c peerLastZxid=0x10000
leader: Sending DIFF zxid=0x10000000c for peer sid: 1
follower: Peer state changed: following - synchronization - diff 清空数据目录后重启,追上用时 624 ms,znode 数 1005
leader: Synchronizing with Learner sid: 1 maxCommittedLog=0x1000003e9 minCommittedLog=0x1000001f6 lastProcessedZxid=0x1000003e9 peerLastZxid=0x0
leader: Sending snapshot last zxid of peer is 0x0, zxid of leader is 0x1000003e9, send zxid of db as 0x1000003e9, 1 concurrent snapshot sync, snapsh
follower: Peer state changed: following - synchronization - snap 旧 leader 被隔离后,发给它的 create /extra:ready = True (它收下并写进了自己的事务日志,但凑不齐多数派)
隔离 4 秒后各节点: {'z42t-1': 'follower', 'z42t-2': 'leader', 'z42t-3': '不服务'}
多数派一侧的新 leader: z42t-2 | /extra 在多数派一侧存在吗: False
网络恢复后各节点: {'z42t-1': 'follower', 'z42t-2': 'leader', 'z42t-3': 'follower'}
新 leader: Synchronizing with Learner sid: 3 maxCommittedLog=0x200000002 minCommittedLog=0x100000001 lastProcessedZxid=0x200000002 peerLastZxid=0x10000
新 leader: Cannot send TRUNC to peer sid: 3 peer zxid is from different epoch
新 leader: Unhandled scenario for peer sid: 3 fall back to use snapshot
新 leader: Sending snapshot last zxid of peer is 0x100000003, zxid of leader is 0x200000002, send zxid of db as 0x200000002, 1 concurrent snapshot sync
旧 leader(现 follower): Peer state changed: following - synchronization - snap
旧 leader 上读 /extra: None | 读 /after: b'2'
当初那个 create /extra 调用的结果: ConnectionLoss隔离是用我写的 TCP 代理实现的(节点间通信悄悄丢包)。旧 leader 收下 /extra 之后,因为没有多数派 ack,这条事务从未提交;新 leader 的 epoch 更高,旧 leader 日志里的多余事务在重新加入时被快照覆盖。