TiDB Binlog 源码阅读系列文章（三）Pump client 介绍

// Interfaces exported by Pump.
service Pump {
    // Writes a binlog to the local file on the pump machine.
    // A response with an empty errmsg is returned if the binlog is written successfully.
    rpc WriteBinlog(WriteBinlogReq) returns (WriteBinlogResp) {}

    // Sends binlog stream from a given location.
    rpc PullBinlogs(PullBinlogReq) returns (stream PullBinlogResp) {}
}

// Binlog contains all the changes in a transaction, which can be used to reconstruct SQL statement, then export to
// other systems.
message Binlog {
    optional BinlogType    tp             = 1 [(gogoproto.nullable) = false];

    // start_ts is used in Prewrite, Commit and Rollback binlog Type.
    // It is used for pairing prewrite log to commit log or rollback log.
    optional int64         start_ts       = 2 [(gogoproto.nullable) = false];

    // commit_ts is used only in binlog type Commit.
    optional int64         commit_ts      = 3 [(gogoproto.nullable) = false];

    // prewrite key is used only in Prewrite binlog type.
    // It is the primary key of the transaction, is used to check that the transaction is
    // commited or not if it failed to pair to commit log or rollback log within a time window.
    optional bytes         prewrite_key   = 4;

    // prewrite_data is marshalled from PrewriteData type,
    // we do not need to unmarshal prewrite data before the binlog have been successfully paired.
    optional bytes         prewrite_value = 5;

    // ddl_query is the original ddl statement query.
    optional bytes         ddl_query      = 6;

    // ddl_job_id is used for DDL Binlog.
    // If ddl_job_id is setted, this is a DDL Binlog and ddl_query contains the DDL query.
    optional int64         ddl_job_id     = 7 [(gogoproto.nullable) = false];
}

Pump 在运行时会将自己的状态信息上报到 PD（etcd）中，并且定时更新自己的状态。在创建 Pump client 的时候，会 [首先从 PD（etcd）中获取所有的 Pump 状态信息](https://github.com/pingcap/tidb-tools/blob/c969908e6130dfbdb4ab80fb84f275df2a6fd877/tidb-binlog/pump_client/client.go#L227)，根据 Pump 状态是否为 Online 初步判断 Pump 为 avaliable 或者 unavailable。然后 Pump client 会 [watch](https://github.com/pingcap/tidb-tools/blob/c969908e6130dfbdb4ab80fb84f275df2a6fd877/tidb-binlog/pump_client/client.go#L478) etcd 中的 Pump 状态变更，及时更新内存中维护的 Pump 状态。

对于每个 Pump，在 Pump client 中都维护了一个变量 [`ErrNum`](https://github.com/pingcap/tidb-tools/blob/c969908e6130dfbdb4ab80fb84f275df2a6fd877/tidb-binlog/pump_client/pump.go#L70) 来记录该 Pump 写 binlog 的失败次数，当 ErrNum 超过一定的阈值，则判断 [该 Pump 不可用](https://github.com/pingcap/tidb-tools/blob/c969908e6130dfbdb4ab80fb84f275df2a6fd877/tidb-binlog/pump_client/pump.go#L174)，如果写 binlog 成功，则 [重置 `ErrNum`](https://github.com/pingcap/tidb-tools/blob/c969908e6130dfbdb4ab80fb84f275df2a6fd877/tidb-binlog/pump_client/pump.go#L162)。

在某些情况下，比如网络抖动，可能会导致 Pump 写 binlog 失败，因此该 Pump 被 Pump client 判断状态为 unavailable，但是当网络恢复后，该 Pump 仍然可以提供写 binlog 服务。Pump client 实现了 [detect](https://github.com/pingcap/tidb-tools/blob/c969908e6130dfbdb4ab80fb84f275df2a6fd877/tidb-binlog/pump_client/client.go#L531) 机制，会定期向 unavailable 状态的 Pump 发送探活请求，如果探活请求成功，则更新 Pump 状态为 avaliable。