baidu / braft

An industrial-grade C++ implementation of RAFT consensus algorithm based on brpc, widely used inside Baidu to build highly-available distributed systems.
Apache License 2.0
3.99k stars 886 forks source link

Snapshot logic has flaws in scenarios where the disk is full #461

Open Axlgrep opened 5 months ago

Axlgrep commented 5 months ago

问题背景

我们的线上服务使用了braft框架, 在运行期间服务所在的机器磁盘使用量达到上限, 查看日志, 在磁盘达到上限的期间有部分关于snapshot的异常日志, 随后我们扩容了磁盘, 并且重启了服务, 但是故障恢复期间在log_manager.cpp内部一直校验失败, 服务无法正常启动.

F20240701 15:34:54.037248 201290 log_manager.cpp:272] Check failed: first_index_kept >= _first_log_index (9355 vs. 10709)
I20240701 15:34:54.053728 201290 server_debug.cpp:219] Catch Signal 6...

原因分析

通过查看本地目录结构, 发现最近一次成功生成的快照目录名称为snapshot_00000000000000009354, 快照的生成时间是20240701 11:11:20.936393, 说明重启之后加载该快照后应该从点位9355处的日志开始进行回放, 但是log中的第一条有效日志的编号却是10709, 说明日志有缺失(缺失的区间[9355, 10709)), 校验确实应该失败, 不允许启动成功.

然后我们去看了最后一次成功生成快照的时间点之后的日志, 发现在20240701 13:11:54.37509720240701 13:35:03.102234这个两个时间点都触发了快照, 并且当时由于磁盘已经满了, 导致产生了部分错误日志(两次快照的日志差不多, 这里截取20240701 13:11:54.375097的日志进行分析)

W20240701 13:11:54.375097   102 util.cpp:322] write falied, err: No space left on device fd: 243 offset: 0 size: 4
W20240701 13:11:54.375156   102 protobuf_file.cpp:61] write len failed, path: /data/shards/***/snapshot/temp/__raft_snapshot_meta.tmp
E20240701 13:11:54.375200   102 snapshot.cpp:74] Fail to save meta to /data/shards/***/snapshot/temp/__raft_snapshot_meta: No space left on device [28]
E20240701 13:11:54.375685   102 snapshot.cpp:276] Fail to sync, path: /data/shards/***/snapshot/temp
I20240701 13:11:54.375737   102 snapshot.cpp:519] Deleting /data/shards/***/snapshot/temp
I20240701 13:11:56.083323   102 snapshot_executor.cpp:234] node ***:127.0.0.1:8100:0 snapshot_save_done, last_included_index=10708 last_included_term=2
I20240701 13:11:56.083458   102 shard_node.h:154] Shard:*** snapshot done ok.
I20240701 13:11:56.083458   102 shard_node.h:154] Shard:tencent-vector-test2_1 snapshot done ok.
I20240701 13:11:56.088521    96 log.cpp:1145] log save_meta /data/shards/***/log/log_meta first_log_index: 10090 time: 5007
I20240701 13:11:56.088872    96 log.cpp:603] Unlinked segment `/data/shards/tencent-vector-test2_1/log/log_00000000000000009352_00000000000000009360'
...
I20240701 13:11:56.103888    96 log.cpp:603] Unlinked segment `/data/shards/tencent-vector-test2_1/log/log_00000000000000010076_00000000000000010083'

从这里可以看出来在快照期间__raft_snapshot_meta存盘操作失败了, 所以后面删除了新生成的temp快照目录(此时最近一次成功快照实际上还是snapshot_00000000000000009354), 但是快照流程还是在继续, 后续清理了区间[9352, 10083]之间的日志并且更新了log_meta文件中的点位信息到10090, 此时快照数据就已经和日志对应不上了(实际上由于日志的清理, 这里已经存在丢失数据的问题了).

函数调用链

在业务层快照逻辑完成之后框架会回调on_snapshot_save_done()函数:

int SnapshotExecutor::on_snapshot_save_done(
    ...
    int ret = st.error_code();
    ...
    if (_snapshot_storage->close(writer) != 0) {
        ret = EIO;
        LOG(WARNING) << "node " << _node->node_id() << " fail to close writer";
    }
    ...
    if (ret == 0) {
        _last_snapshot_index = meta.last_included_index();
        _last_snapshot_term = meta.last_included_term();
        lck.unlock();
        ss << "snapshot_save_done, last_included_index=" << meta.last_included_index()
           << " last_included_term=" << meta.last_included_term();
        LOG(INFO) << ss.str();
        _log_manager->set_snapshot(&meta);
        lck.lock();
    }
    ...
    return ret;
}

其中有一段逻辑是_snapshot_storage调用close进行快照元信息的存盘以及快照目录的rename操作:

int LocalSnapshotStorage::close(SnapshotWriter* writer_base,
                                bool keep_data_on_error) {
    LocalSnapshotWriter* writer = dynamic_cast<LocalSnapshotWriter*>(writer_base);
    int ret = writer->error_code();
    do {
        ...
        ret = writer->sync();
        if (ret != 0) {
            break;
        }
        ...
    } while (0);

    if (ret != 0 && !keep_data_on_error) {
        destroy_snapshot(writer->get_path());
    }
    delete writer;
    return ret != EIO ? 0 : -1;
}

writer->sync()的内部实现实际上就是将快照的元信息进行存盘操作, 在磁盘满导致内部存盘失败的场景下, 这个函数的返回值是-1, 然后跳到函数末尾判断return ret != EIO ? 0 : -1;, 这里实现实际上是有严重问题的, 只有ret的值等于EIO(5)的时候才认为出错了, 如果ret为-1, 实际上close()函数返回了0, 静默了内部错误, 导致上层认为close()函数正常执行, 并且进行了raft 日志的清理工作.

最终导致了实际上快照目录没有更新(一直是最后一次成功快照生成的目录, 也就是20240701 11:11:20.936393时刻生成的snapshot_00000000000000009354), 但是raft日志缺被清理掉了, 并且log_meta的文件中记录的点位一直后移, 造成快照目录数据和raft日志接不上的问题.

修复办法

问题分析清楚之后实际上修复办法还是比较简单的, 当writer->sync()失败之后, 将ret设置成EIO, 将错误向上层抛, 而不是静默错误让上层继续清理braft日志.