C implementation of the Raft consensus protocol, BSD licensed.

See raft.h for full documentation.

See ticketd for real life use of this library.

Networking is out of scope for this project. The implementor will need to do all the plumbing. The library doesn't assume a network layer with ordering or duplicate detection. This means you could use UDP for transmission.

There are no dependencies, however https://github.com/willemt/linked-List-queue is required for testing.

make tests

We use the following methods to ensure that the library is safe:

• A simulator (virtraft2) is used to test the Raft invariants on unreliable networks
• Fuzzing/property-based-testing via Hypothesis
• All bugs have regression tests
• Many unit tests
• Usage

This cluster simulator checks the following:

• Log Matching (servers must have matching logs)
• State Machine Safety (applied entries have the same ID)
• Election Safety (only one valid leader per term)
• Current Index Validity (does the current index have an existing entry?)
• Entry ID Monotonicity (entries aren't appended out of order)
• Committed entry popping (committed entries are not popped from the log)
• Log Accuracy (does the server's log match mirror an independent log?)
• Deadlock detection (does the cluster continuously make progress?)

Chaos generated by virtraft2:

• Random bi-directional partitions between nodes
• Message dropping
• Message duplication
• Membership change injection
• Random compactions

Run the simulator using:

make test_virtraft

virtraft2 succeeds virtraft

The source has been amalgamated into a single raft.h header file. Use clib to download the source into your project's deps folder, ie:

brew install clib
clib install willemt/raft_amalgamation

The file is stored in the deps folder like below:

deps/raft/raft.h

See ticketd for an example of how to integrate with this library.

If you don't have access to coroutines it's easiest to use two separate threads - one for handling Raft peer traffic, and another for handling client traffic.

Be aware that this library is not thread safe. You will need to ensure that the library's functions are called exclusively.

Instantiate a new Raft server using raft_new.

void* raft = raft_new();

We tell the Raft server what the cluster configuration is by using the raft_add_node function. For example, if we have 5 servers¹ in our cluster, we call raft_add_node 5² times.

raft_add_node(raft, connection_user_data, node_id, peer_is_self);

Where:

• connection_user_data is a pointer to user data.
• peer_is_self is boolean indicating that this is the current server's server index.
• node_id is the unique integer ID of the node. Peers use this to identify themselves. This SHOULD be a random integer.

We need to call raft_periodic at periodic intervals.

raft_periodic(raft, 1000);

Example using a libuv timer:

static void __periodic(uv_timer_t* handle)
{
    raft_periodic(sv->raft, PERIOD_MSEC);
}

uv_timer_t *periodic_req;
periodic_req = malloc(sizeof(uv_timer_t));
periodic_req->data = sv;
uv_timer_init(&peer_loop, periodic_req);
uv_timer_start(periodic_req, __periodic, 0, 1000);

Our Raft application receives log entries from the client.

When this happens we need to:

• Redirect the client to the Raft cluster leader (if necessary)
• Append the entry to our log
• Block until the log entry has been committed³

Append the entry to our log

We call raft_recv_entry when we want to append the entry to the log.

msg_entry_response_t response;
e = raft_recv_entry(raft,  &entry, &response);

You should populate the entry struct with the log entry the client has sent. After the call completes the response parameter is populated and can be used by the raft_msg_entry_response_committed function to check if the log entry has been committed or not.

Blocking until the log entry has been committed

When the server receives a log entry from the client, it has to block until the entry is committed. This is necessary as our Raft server has to replicate the log entry with the other peers of the Raft cluster.

The raft_recv_entry function does not block! This means you will need to implement the blocking functionality yourself.

Example below is from the ticketd client thread. This shows that we need to block on client requests. ticketd does the blocking by waiting on a conditional, which is signalled by the peer thread. The separate thread is responsible for handling traffic between Raft peers.

msg_entry_response_t response;

e = raft_recv_entry(sv->raft, &entry, &response);
if (0 != e)
    return h2oh_respond_with_error(req, 500, "BAD");

/* block until the entry is committed */
int done = 0;
do {
    uv_cond_wait(&sv->appendentries_received, &sv->raft_lock);
    e = raft_msg_entry_response_committed(sv->raft, &r);
    switch (e)
    {
        case 0:
            /* not committed yet */
            break;
        case 1:
            done = 1;
            uv_mutex_unlock(&sv->raft_lock);
            break;
        case -1:
            uv_mutex_unlock(&sv->raft_lock);
            return h2oh_respond_with_error(req, 400, "TRY AGAIN");
    }
} while (!done);

Example from ticketd of the peer thread. When an appendentries response is received from a Raft peer, we signal to the client thread that an entry might be committed.

e = raft_recv_appendentries_response(sv->raft, conn->node, &m.aer);
uv_cond_signal(&sv->appendentries_received);

Redirecting the client to the leader

When we receive an entry log from the client it's possible we might not be a leader.

If we aren't currently the leader of the raft cluster, we MUST send a redirect error message to the client. This is so that the client can connect directly to the leader in future connections. This enables future requests to be faster (ie. no redirects are required after the first redirect until the leader changes).

We use the raft_get_current_leader function to check who is the current leader.

Example of ticketd sending a 301 HTTP redirect response:

/* redirect to leader if needed */
raft_node_t* leader = raft_get_current_leader_node(sv->raft);
if (!leader)
{
    return h2oh_respond_with_error(req, 503, "Leader unavailable");
}
else if (raft_node_get_id(leader) != sv->node_id)
{
    /* send redirect */
    peer_connection_t* conn = raft_node_get_udata(leader);
    char leader_url[LEADER_URL_LEN];
    static h2o_generator_t generator = { NULL, NULL };
    static h2o_iovec_t body = { .base = "", .len = 0 };
    req->res.status = 301;
    req->res.reason = "Moved Permanently";
    h2o_start_response(req, &generator);
    snprintf(leader_url, LEADER_URL_LEN, "http://%s:%d/",
             inet_ntoa(conn->addr.sin_addr), conn->http_port);
    h2o_add_header(&req->pool,
                   &req->res.headers,
                   H2O_TOKEN_LOCATION,
                   leader_url,
                   strlen(leader_url));
    h2o_send(req, &body, 1, 1);
    return 0;
}

You provide your callbacks to the Raft server using raft_set_callbacks.

The following callbacks MUST be implemented: send_requestvote, send_appendentries, applylog, persist_vote, persist_term, log_offer, and log_pop.

Example of function callbacks being set:

raft_cbs_t raft_callbacks = {
    .send_requestvote            = __send_requestvote,
    .send_appendentries          = __send_appendentries,
    .applylog                    = __applylog,
    .persist_vote                = __persist_vote,
    .persist_term                = __persist_term,
    .log_offer                   = __raft_logentry_offer,
    .log_poll                    = __raft_logentry_poll,
    .log_pop                     = __raft_logentry_pop,
    .log                         = __raft_log,
};

char* user_data = "test";

raft_set_callbacks(raft, &raft_callbacks, user_data);

send_requestvote()

For this callback we have to serialize a msg_requestvote_t struct, and then send it to the peer identified by node.

Example from ticketd showing how the callback is implemented:

static int __send_requestvote(
    raft_server_t* raft,
    void *udata,
    raft_node_t* node,
    msg_requestvote_t* m
    )
{
    peer_connection_t* conn = raft_node_get_udata(node);

    uv_buf_t bufs[1];
    char buf[RAFT_BUFLEN];
    msg_t msg = {
        .type              = MSG_REQUESTVOTE,
        .rv                = *m
    };
    __peer_msg_serialize(tpl_map("S(I$(IIII))", &msg), bufs, buf);
    int e = uv_try_write(conn->stream, bufs, 1);
    if (e < 0)
        uv_fatal(e);
    return 0;
}

send_appendentries()

For this callback we have to serialize a msg_appendentries_t struct, and then send it to the peer identified by node. This struct is more complicated to serialize because the m->entries array might be populated.

Example from ticketd showing how the callback is implemented:

static int __send_appendentries(
    raft_server_t* raft,
    void *user_data,
    raft_node_t* node,
    msg_appendentries_t* m
    )
{
    uv_buf_t bufs[3];

    peer_connection_t* conn = raft_node_get_udata(node);

    char buf[RAFT_BUFLEN], *ptr = buf;
    msg_t msg = {
        .type              = MSG_APPENDENTRIES,
        .ae                = {
            .term          = m->term,
            .prev_log_idx  = m->prev_log_idx,
            .prev_log_term = m->prev_log_term,
            .leader_commit = m->leader_commit,
            .n_entries     = m->n_entries
        }
    };
    ptr += __peer_msg_serialize(tpl_map("S(I$(IIIII))", &msg), bufs, ptr);

    /* appendentries with payload */
    if (0 < m->n_entries)
    {
        tpl_bin tb = {
            .sz   = m->entries[0].data.len,
            .addr = m->entries[0].data.buf
        };

        /* list of entries */
        tpl_node *tn = tpl_map("IIIB",
            &m->entries[0].id,
            &m->entries[0].term,
            &m->entries[0].type,
            &tb);
        size_t sz;
        tpl_pack(tn, 0);
        tpl_dump(tn, TPL_GETSIZE, &sz);
        e = tpl_dump(tn, TPL_MEM | TPL_PREALLOCD, ptr, RAFT_BUFLEN);
        assert(0 == e);
        bufs[1].len = sz;
        bufs[1].base = ptr;
        e = uv_try_write(conn->stream, bufs, 2);
        if (e < 0)
            uv_fatal(e);

        tpl_free(tn);
    }
    else
    {
        /* keep alive appendentries only */
        e = uv_try_write(conn->stream, bufs, 1);
        if (e < 0)
            uv_fatal(e);
    }

    return 0;
}

applylog()

This callback is all what is needed to interface the FSM with the Raft library. Depending on your application, you might want to save the commit_idx to disk inside this callback.

persist_vote() & persist_term()

These callbacks simply save data to disk, so that when the Raft server is rebooted it starts from a valid state. This is necessary to ensure safety.

log_offer()

For this callback the user needs to add a log entry. The log MUST be synced to disk before this callback can return.

log_poll()

For this callback the user needs to remove the eldest log entry⁴. The log MUST be synced to disk before this callback can return.

This callback only needs to be implemented to support log compaction.

log_pop()

For this callback the user needs to remove the youngest log entry⁵. The log MUST be synced to disk before this callback can return.

To receive Append Entries, Append Entries response, Request Vote, and Request Vote response messages, you need to deserialize the bytes into the message's corresponding struct.

The table below shows the structs that you need to deserialize-to or deserialize-from:

Example of how we receive an Append Entries message, and reply to it:

msg_appendentries_t ae;
msg_appendentries_response_t response;
char buf_in[1024], buf_out[1024];
size_t len_in, len_out;

read(socket, buf_in, &len_in);

deserialize_appendentries(buf_in, len_in, &ae);

e = raft_recv_requestvote(sv->raft, conn->node, &ae, &response);

serialize_appendentries_response(&response, buf_out, &len_out);

write(socket, buf_out, &len_out);

Membership changes are managed on the Raft log. You need two log entries to add a server to the cluster. While to remove you only need one log entry. There are two log entries for adding a server because we need to ensure that the new server's log is up to date before it can take part in voting.

It's highly recommended that when a node is added to the cluster that its node ID is random. This is especially important if the server was once connected to the cluster.

Adding a node

1. Append the configuration change using raft_recv_entry. Make sure the entry has the type set to RAFT_LOGTYPE_ADD_NONVOTING_NODE
2. Once node_has_sufficient_logs callback fires, append a configuration finalization log entry using raft_recv_entry. Make sure the entry has a type set to RAFT_LOGTYPE_ADD_NODE

Removing a node

1. Append the configuration change using raft_recv_entry. Make sure the entry has the type set to RAFT_LOGTYPE_REMOVE_NODE
2. Once the RAFT_LOGTYPE_REMOVE_NODE configuration change log is applied in the applylog callback we shutdown the server if it is to be removed.

Membership callback

The notify_membership_event callback can be used to track nodes as they are added and removed as a result of configuration change log entries. A typical use case is to create and destroy connections to nodes, using connection information obtained from the configuration change log entry.

The log compaction method supported is called "Snapshotting for memory-based state machines" (Ongaro, 2014)

This library does not send snapshots (ie. there are NO send_snapshot, recv_snapshot callbacks to implement). The user has to send the snapshot outside of this library. The implementor has to serialize and deserialize the snapshot.

The process works like this:

1. Begin snapshotting with raft_begin_snapshot.
2. Save the current membership details to the snapshot.
3. Save the finite state machine to the snapshot.
4. End snapshotting with raft_end_snapshot.
5. When the send_snapshot callback fires, the user must propagate the snapshot to the peer.
6. Once the peer has the snapshot, they call raft_begin_load_snapshot.
7. Peer calls raft_add_node to add nodes as per the snapshot's membership info.
8. Peer calls raft_node_set_voting to nodes as per the snapshot's membership info.
9. Peer calls raft_node_set_active to nodes as per the snapshot's membership info.
10. Finally, peer calls raft_node_set_active to nodes as per the snapshot's membership info.

When a node receives a snapshot it could reuse that snapshot itself for other nodes.

• Batch friendly interfaces - we can speed up Raft by adding new APIs that support batching many log entries
• Implementing linearizable semantics (Ongaro, 2014)
• Processing read-only queries more efficiently (Ongaro, 2014)

Ongaro, D. (2014). Consensus: bridging theory and practice. Retrieved from https://web.stanford.edu/~ouster/cgi-bin/papers/OngaroPhD.pdf