Raft implementation in c language (with code)

C language Raft implementation (with code)

1. Introduction

This article introduces a simple Raft implementation. If you have read the Raft paper, you will find it easier to read this Raft implementation, because the details of the implementation are described in great detail in the Raft paper. The engineering implementation is basically to re-express the description in the Raft paper in programming language. This is the biggest advantage of Raft over Paxos, that is, it is easy to understand and implement. The Raft implementation introduced in this article is coded in C language. Except for the log compression function, other features are implemented. The member change mechanism is also relatively simple, and only supports one configuration change at a time. For the principles of Raft, you can read the Raft paper and "Raft Understanding".

2. Basic concepts of Raft

2.1 Status

raft has three states: Leader, Candidate and Follower. The transitions between these three states are shown in the figure below. Only the Leader has the right to process customer requests and copy logs to Followers. Candidate is an intermediate state between Follower and Leader. When there is no Leader in the cluster, Follower enters the Candidate state and initiates a vote to the cluster. The Follower that obtains the majority of votes will become the Leader.

2.2 Messages

In order to improve the understandability of the protocol, Raft has simplified the setting of message types and only has the following two requests. [indent]
requestVote Initiates a voting request. Candidate's request when initiating a vote. It is received and processed by other Followers and Candidates in the cluster.
appendEntries Add log request. The request issued by the Leader when adding logs to the Follower. [/indent]

2.3 Term number

The term number term is used in the Raft protocol to indicate the old and new relationship of time. This term value within the term of each Leader is Unchanging, it is absolutely different among different Leaders and increases monotonically with time. If the term of request A is larger than that of request B, request B is out of date.

3.Raft implementation

3.1 Protocol

First introduce four important data structures, corresponding to the ones mentioned above requestVote and appendEntries request and reply.

/** requestVote 请求投票
   * 竞选者Candidate去竞选Leader时发送给其它node的投票请求。
   * 其它Leader或者Candidate收到term比自己大的投票请求时，会自动变成Follower*/
typedef struct
{
    /** 当前任期号，通过任期号的大小与其它Candidate竞争Leader */
    int term;

    /** 竞选者的id */
    int candidate_id;

    /** 竞选者本地保存的最新一条日志的index */
    int last_log_idx;

    /** 竞选者本地保存的最新一条日志的任期号*/
    int last_log_term;
} msg_requestvote_t;


/** 投票请求的回复response.
  * 该response主要是给返回某个node是否接收了Candidate的投票请求. */
typedef struct
{
    /** node的任期号，Candidate根据投票结果和node的任期号来更新自己的任期号 */
    int term;

    /** 投票结果，如果node给Candidate投票则为true */
    int vote_granted;
} msg_requestvote_response_t;

/**  添加日志请求.
  * Follower可以从该消息中知道哪些日志可以安全地提交到状态机FSM中去。
  * Leader可以将该消息作为心跳消息定期发送。
  * 旧的Leader和Candidate收到该消息后可能会自动变成Follower */
typedef struct
{
    /** Leader当前的任期号 */
    int term;

    /** 最新日志的前一条日志的index，用于Follower确认与Leader的日志完全一致 */
    int prev_log_idx;

    /** 最新日志的前一条日志的任期号term */
    int prev_log_term;

    /** leader当前已经确认提交到状态机FSM的日志索引index，这意味着Follower也可以安全地将该索引index以前的日志提交 */
    int leader_commit;

    /** 这条添加日志消息携带的日志条数，该实现中最多只有一条 */
    int n_entries;

    /** 这条添加日志消息中携带的日志数组 */
    msg_entry_t* entries;
} msg_appendentries_t;

/** 添加日志回复.
 * 旧的Leader或Candidate收到该消息会变成Follower *／
typedef struct
{
    /** 当前任期号 */
    int term;

    /** node成功添加日志时返回ture，即prev_log_index和prev_log_term都比对成功。否则返回false */
    int success;

    /* 下面两个字段不是Raft论文中规定的字段:
    /* 用来优化日志追加过程，以加速日志的追加。Raft原文中的追加过程是一次只能追加一条日志*/

    /** 处理添加日志请求后本地的最大日志索引 */
    int current_idx;

    /** 从添加日志请求中接受的第一条日志索引 */
    int first_idx;
} msg_appendentries_response_t;

3.2 Two important abstractions

##raft_server_private_t This structure is the abstraction in Raft’s implementation. Saves the status and all data required during the operation of the Raft protocol.

typedef struct {
    /* 所有服务器比较固定的状态: */

    /* 服务器最后一次知道的任期号（初始化为 0，持续递增） */
    int current_term;

    /* 记录在当前分期内给哪个Candidate投过票，
       */
    int voted_for;

    /* 日志条目集；每一个条目包含一个用户状态机执行的指令，和收到时的任期号 */
    void* log;

    /* 变动比较频繁的变量: */

    /* 已知的最大的已经被提交的日志条目的索引值 */
    int commit_idx;

    /* 最后被应用到状态机的日志条目索引值（初始化为 0，持续递增） */
    int last_applied_idx;

    /* 三种状态：follower/leader/candidate */
    int state;

    /* 计时器，周期函数每次执行时会递增改值 */
    int timeout_elapsed;

    raft_node_t* nodes;
    int num_nodes;

    int election_timeout;
    int request_timeout;

    /* 保存Leader的信息，没有Leader时为NULL */
    raft_node_t* current_leader;

    /* callbacks，由调用该raft实现的调用者来实现，网络IO和持久存储
     * 都由调用者在callback中实现 */
    raft_cbs_t cb;
    void* udata;

    /* 自己的信息 */
    raft_node_t* node;

    /* 该raft实现每次只进行一个服务器的配置更改，该变量记录raft server
     * 是否正在进行配置更改*/
    int voting_cfg_change_log_idx;
} raft_server_private_t;

raft_node_private_t The abstract body of the machine node in the cluster contains information on other machines that needs to be saved during the running of the raft protocol

typedef struct
{
    void* udata;  /*一般保存与其它机器的连接信息，由使用者决定怎么实现连接*/

    int next_idx; /*对于每一个服务器，需要发送给他的下一个日志条目的索引值（初始化为领导人最后索引值加一）*/
    int match_idx; /*对于每一个服务器，已经复制给他的日志的最高索引值*/

    int flags; /*有三种取值，是相或的关系 1:该机器有给我投票 2:该机器有投票权  3: 该机器有最新的日志*/

    int id; /*机器对应的id值，这个每台机器在全局都是唯一的*/
} raft_node_private_t;

3.3 Raft protocol Process

Periodic function Raft needs to do some things periodically. For example, the Leader needs to periodically append logs to other servers, so that servers with lagging logs have a chance to catch up; All servers need to periodically apply confirmed commit logs to the state machine, etc. raft_periodic function is a function that is called periodically in the raft implementation, and the calling period is 1000ms. The machine will do different things in this function in different states. Leader periodically synchronizes logs to Follower. The Follower periodically detects whether it has not received a heartbeat packet from the Leader within a specific period of time. If so, it becomes a Candidate and starts voting for the Leader. Both Leader and Follower will periodically commit submitted logs to the state machine FSM.

/** raft周期性执行的函数,实现raft中的定时器以及定期应用日志到状态机
  */
int raft_periodic(raft_server_t* me_, int msec_since_last_period)
{
    raft_server_private_t* me = (raft_server_private_t*)me_;

    /* 选举计时器；Follower每次收到Leader的心跳后会重置清0，Leader每次发送日志也会清0 */
    me->timeout_elapsed += msec_since_last_period;

    /* Leader周期性地向Follower同步日志 */
    if (me->state == RAFT_STATE_LEADER) 
    {
        if (me->request_timeout <= me->timeout_elapsed)
            raft_send_appendentries_all(me_);
    }
    /* Follower检测选举计时器是否超时 */
    else if (me->election_timeout <= me->timeout_elapsed)
    {
        if (1 < me->num_nodes)
            raft_election_start(me_);
    }

    /* 周期性地将已经确认commit的日志应用到状态机FSM */
    if (me->last_applied_idx < me->commit_idx)
        if (-1 == raft_apply_entry(me_))
            return -1;

    return 0;
}

Become a candidate Candidate Each server in the cluster has an election timer. When a server does not receive a heartbeat from the Leader within the timer timeout period, the cluster is considered If there is no leader in the server or the leader is down, the server will become a candidate, and then initiate a vote to run for leader. The following raft_become_candidate function is the function for the server to become a candidate. The function mainly does the following things:

Increase the current term number (currentTerm)

Vote for yourself

Reset the election timeout timer

Send The RPC requesting voting is sent to all other servers

/** Follower成为Candidate执行的函数
  */
void raft_become_candidate(raft_server_t* me_)
{
    raft_server_private_t* me = (raft_server_private_t*)me_;
    int i;

    /*自增当前的任期号；给自己投票，设置自己的状态为CANDIDATE*/
    raft_set_current_term(me_, raft_get_current_term(me_) + 1);
    for (i = 0; i < me->num_nodes; i++)
        raft_node_vote_for_me(me->nodes[i], 0);
    raft_vote(me_, me->node);
    me->current_leader = NULL;
    raft_set_state(me_, RAFT_STATE_CANDIDATE);

    /* 重置选举超时计时器。为了防止多个Candidate竞争，将下一次发起投票的时间间隔设置成随机值*/
    /* TODO: this should probably be lower */
    me->timeout_elapsed = rand() % me->election_timeout;

    /*发送请求投票的 RPC 给其他所有服务器*/
    for (i = 0; i < me->num_nodes; i++)
        if (me->node != me->nodes[i] && raft_node_is_voting(me->nodes[i]))
            raft_send_requestvote(me_, me->nodes[i]);
}

Processing voting requestsThe logic of processing voting requests is mainly to determine whether to agree to vote. The basis for judgment is the term number and log information in the request The level of freshness and age, and whether you have voted for other servers with the same term number. If you have voted before, you cannot vote again. Each person only has one vote.

If term > currentTerm, switch to Follower mode.

The server that receives the voting request here may be a Leader with poor network conditions or a Candidate that has not yet had time to issue a voting request. After they receive a request with a newer term number than their own, they must unconditionally become Follower to ensure that only one Leader exists

如果term daab63a7de571b85be82bb1122a6d0ecterm即为过时)

follower是否成功添加日志，如果添加失败，则减小发给follower的日志索引nextIndex再重试；如果添加成功则更新本地记录的follower日志信息，并检查日志是否最新，如果不是最新则继续发送添加日志请求。

新机器的日志添加，详见3.4节-- 成员变更

/** 处理添加日志请求回复
  * /
int raft_recv_appendentries_response(raft_server_t* me_,
                                     raft_node_t* node,
                                     msg_appendentries_response_t* r)
{
    raft_server_private_t* me = (raft_server_private_t*)me_;

    __log(me_, node,
          "received appendentries response %s ci:%d rci:%d 1stidx:%d",
          r->success == 1 ? "SUCCESS" : "fail",
          raft_get_current_idx(me_),
          r->current_idx,
          r->first_idx);

    /* 过时的回复 -- 忽略 */
    if (r->current_idx != 0 && r->current_idx <= raft_node_get_match_idx(node))
        return 0;

    /* oh~我不是Leader */
    if (!raft_is_leader(me_))
        return -1;

    /* 回复中的term比自己的要大，说明自己是一个过时的Leader，无条件转为Follower */
    if (me->current_term < r->term)
    {
        raft_set_current_term(me_, r->term);
        raft_become_follower(me_);
        return 0;
    }
    /* 过时的回复，网络状况不好时会出现 */
    else if (me->current_term != r->term)
        return 0;

    /* stop processing, this is a node we don&#39;t have in our configuration */
    if (!node)
        return 0;

    /* 由于日志不一致导致添加日志不成功*/
    if (0 == r->success)
    {
        assert(0 <= raft_node_get_next_idx(node));

        /* 将nextIdex减*/
        int next_idx = raft_node_get_next_idx(node);
        assert(0 <= next_idx);
        /* Follower的日志数量还远远少于Leader，将nextIdex设为回复中的current_idx+1和Leader
         * 当前索引中较小的一个，一般回复中的current_idx+1会比较小*/
        if (r->current_idx < next_idx - 1)
            raft_node_set_next_idx(node, min(r->current_idx + 1, raft_get_current_idx(me_)));
        /* Follower的日志数量和Leader差不多，但是比对前一条日志时失败，这种情况将next_idx减1
         * 重试*/
        else
            raft_node_set_next_idx(node, next_idx - 1);

        /* 使用更新后的nextIdx重新发送添加日志请求 */
        raft_send_appendentries(me_, node);
        return 0;
    }

    assert(r->current_idx <= raft_get_current_idx(me_));

    /* 下面处理添加日志请求的情况 */
    /* 更新本地记录的Follower的日志情况 */
    raft_node_set_next_idx(node, r->current_idx + 1);
    raft_node_set_match_idx(node, r->current_idx);

    /* 如果是新加入的机器，则判断它的日志是否是最新，如果达到了最新，则赋予它投票权，
     * 这里逻辑的详细解释在第3.4节 -- 成员变更*/
    if (!raft_node_is_voting(node) &&
        -1 == me->voting_cfg_change_log_idx &&
        raft_get_current_idx(me_) <= r->current_idx + 1 &&
        me->cb.node_has_sufficient_logs &&
        0 == raft_node_has_sufficient_logs(node)
        )
    {
        raft_node_set_has_sufficient_logs(node);
        me->cb.node_has_sufficient_logs(me_, me->udata, node);
    }

    /* 如果一条日志回复成功的数量超过一半，则将日志提交commit，即允许应用到状态机 */
    int votes = 1; /* include me */
    int point = r->current_idx;
    int i;
    for (i = 0; i < me->num_nodes; i++)
    {
        if (me->node == me->nodes[i] || !raft_node_is_voting(me->nodes[i]))
            continue;

        int match_idx = raft_node_get_match_idx(me->nodes[i]);

        if (0 < match_idx)
        {
            raft_entry_t* ety = raft_get_entry_from_idx(me_, match_idx);
            /*如果follower已经添加了索引大于等于r->current_idx的日志，则vote加1*/
            if (ety->term == me->current_term && point <= match_idx)
                votes++;
        }
    }

    /* 投票数大于所有服务器的一半，则将日志提交 */
    if (me->num_nodes / 2 < votes && raft_get_commit_idx(me_) < point)
        raft_set_commit_idx(me_, point);

    /* 如果follower的日志还没有最新，那么继续发送添加日志请求 */
    if (raft_get_entry_from_idx(me_, raft_node_get_next_idx(node)))
        raft_send_appendentries(me_, node);

    /* periodic applies committed entries lazily */

    return 0;
}

3.3 成员变更

成员的变更都是以日志的形式下发的。添加的新成员分两阶段进行，第一阶段中新成员没有有投票权，但是有接收日志的权力；当它的日志同步到最新后就进入到第二阶段，由Leader赋予投票权，从而成为集群中完整的一员。删除成员相对比较简单，所有服务器收到删除成员的日志后，立马将该成员的信息从本地抹除。

添加成员过程

1. 管理员向Leader发送添加成员命令
   

2. Leader添加一条 RAFT_LOGTYPE_ADD_NONVOTING_NODE日志，即添加没有投票权的服务器。该日志与其它普通日志一样同步给集群中其它服务器。收到该日志的服务器在本地保存该新成员的信息。
   

3. 当新成员的日志同步到最新后，Leader添加一条 RAFT_LOGTYPE_ADD_NODE日志，即有投票权的服务器，同样地，该日志与其它普通日志一样同步给集群中其它服务器。收到该日志的服务器在本地保存该新成员的信息，以后的投票活动会将新成员考虑进去。

删除成员过程

1. 管理员向Leader发送删除成员命令。
   

2. Leader添加一条 RAFT_LOGTYPE_REMOVE_NODE 日志，并跟普通日志一样同步给其它服务器。收到该日志的服务器立即将被成员信息从本地删除。

感谢大家的阅读，希望大家收益多多。

推荐教程：《C语言》

The above is the detailed content of Raft implementation in c language (with code). For more information, please follow other related articles on the PHP Chinese website!