consensus.rs 12 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373
  1. use async_std::{
  2. sync::{Arc, Mutex},
  3. task::sleep,
  4. };
  5. use std::time::Duration;
  6. use async_executor::Executor;
  7. use chrono::Utc;
  8. use futures::{select, FutureExt};
  9. use fxhash::FxHashMap;
  10. use log::{debug, error, warn};
  11. use rand::{rngs::OsRng, Rng, RngCore};
  12. use crate::{
  13. net,
  14. util::{
  15. gen_id,
  16. serial::{deserialize, serialize, Decodable, Encodable},
  17. },
  18. Error, Result,
  19. };
  20. use super::{
  21. p2p_send_loop,
  22. primitives::{
  23. BroadcastMsgRequest, Channel, Log, LogRequest, LogResponse, Logs, MapLength, NetMsg,
  24. NetMsgMethod, NodeId, NodeIdMsg, Role, Sender, VoteRequest, VoteResponse,
  25. },
  26. prune_map, DataStore, RaftSettings,
  27. };
  28. async fn send_loop(sender: async_channel::Sender<()>, timeout: Duration) -> Result<()> {
  29. loop {
  30. sleep(timeout).await;
  31. sender.send(()).await?;
  32. }
  33. }
  34. pub struct Raft<T> {
  35. id: NodeId,
  36. pub(super) role: Role,
  37. pub(super) current_leader: NodeId,
  38. pub(super) votes_received: Vec<NodeId>,
  39. pub(super) sent_length: MapLength,
  40. pub(super) acked_length: MapLength,
  41. pub(super) nodes: Arc<Mutex<FxHashMap<NodeId, i64>>>,
  42. pub(super) last_term: u64,
  43. pub(super) last_heartbeat: i64,
  44. p2p_sender: Sender,
  45. msgs_channel: Channel<T>,
  46. commits_channel: Channel<T>,
  47. datastore: DataStore<T>,
  48. seen_msgs: Arc<Mutex<FxHashMap<String, i64>>>,
  49. pub(super) settings: RaftSettings,
  50. pending_msgs: Vec<T>,
  51. }
  52. impl<T: Decodable + Encodable + Clone> Raft<T> {
  53. pub fn new(
  54. settings: RaftSettings,
  55. seen_msgs: Arc<Mutex<FxHashMap<String, i64>>>,
  56. ) -> Result<Self> {
  57. if settings.datastore_path.to_str().is_none() {
  58. error!(target: "raft", "datastore path is incorrect");
  59. return Err(Error::ParseFailed("unable to parse pathbuf to str"))
  60. };
  61. let datastore = DataStore::new(settings.datastore_path.to_str().unwrap())?;
  62. // broadcasting channels
  63. let msgs_channel = async_channel::unbounded::<T>();
  64. let commits_channel = async_channel::unbounded::<T>();
  65. let p2p_sender = async_channel::unbounded::<NetMsg>();
  66. let id = match datastore.id.get_last()? {
  67. Some(_id) => _id,
  68. None => {
  69. let id = NodeId(gen_id(30));
  70. datastore.id.insert(&id)?;
  71. id
  72. }
  73. };
  74. let role = Role::Follower;
  75. Ok(Self {
  76. id,
  77. role,
  78. current_leader: NodeId("".into()),
  79. votes_received: vec![],
  80. sent_length: MapLength(FxHashMap::default()),
  81. acked_length: MapLength(FxHashMap::default()),
  82. nodes: Arc::new(Mutex::new(FxHashMap::default())),
  83. last_term: 0,
  84. last_heartbeat: Utc::now().timestamp(),
  85. p2p_sender,
  86. msgs_channel,
  87. commits_channel,
  88. datastore,
  89. seen_msgs,
  90. settings,
  91. pending_msgs: vec![],
  92. })
  93. }
  94. ///
  95. /// Run raft consensus and wait stop_signal channel to terminate
  96. ///
  97. pub async fn run(
  98. &mut self,
  99. p2p: net::P2pPtr,
  100. p2p_recv_channel: async_channel::Receiver<NetMsg>,
  101. executor: Arc<Executor<'_>>,
  102. stop_signal: async_channel::Receiver<()>,
  103. ) -> Result<()> {
  104. let p2p_send_task = executor.spawn(p2p_send_loop(self.p2p_sender.1.clone(), p2p.clone()));
  105. let prune_seen_messages_task = executor
  106. .spawn(prune_map::<String>(self.seen_msgs.clone(), self.settings.prun_duration));
  107. let prune_nodes_id_task =
  108. executor.spawn(prune_map::<NodeId>(self.nodes.clone(), self.settings.prun_duration));
  109. let mut rng = rand::thread_rng();
  110. let (id_sx, id_rv) = async_channel::unbounded::<()>();
  111. let (heartbeat_sx, heartbeat_rv) = async_channel::unbounded::<()>();
  112. let (timeout_sx, timeout_rv) = async_channel::unbounded::<()>();
  113. let id_timeout = Duration::from_secs(self.settings.id_timeout);
  114. let send_id_task = executor.spawn(send_loop(id_sx, id_timeout));
  115. let heartbeat_timeout = Duration::from_millis(self.settings.heartbeat_timeout);
  116. let send_heartbeat_task = executor.spawn(send_loop(heartbeat_sx, heartbeat_timeout));
  117. let timeout =
  118. Duration::from_secs(rng.gen_range(0..self.settings.timeout) + self.settings.timeout);
  119. let send_timeout_task = executor.spawn(send_loop(timeout_sx, timeout));
  120. let broadcast_msg_rv = self.msgs_channel.1.clone();
  121. loop {
  122. let mut result = select! {
  123. m = p2p_recv_channel.recv().fuse() => self.handle_method(m?).await,
  124. m = broadcast_msg_rv.recv().fuse() => self.broadcast_msg(&m?,None).await,
  125. _ = id_rv.recv().fuse() => self.send_id_msg().await,
  126. _ = heartbeat_rv.recv().fuse() => self.send_heartbeat().await,
  127. _ = timeout_rv.recv().fuse() => self.send_vote_request().await,
  128. _ = stop_signal.recv().fuse() => break,
  129. };
  130. // send pending messages
  131. if !self.pending_msgs.is_empty() && self.role != Role::Candidate {
  132. let pending_msgs = self.pending_msgs.clone();
  133. for m in &pending_msgs {
  134. result = self.broadcast_msg(m, None).await;
  135. }
  136. self.pending_msgs = vec![];
  137. }
  138. if let Err(e) = result {
  139. warn!(target: "raft", "warn: {}", e);
  140. }
  141. }
  142. warn!(target: "raft", "Raft Terminating...");
  143. p2p_send_task.cancel().await;
  144. prune_seen_messages_task.cancel().await;
  145. prune_nodes_id_task.cancel().await;
  146. send_id_task.cancel().await;
  147. send_heartbeat_task.cancel().await;
  148. send_timeout_task.cancel().await;
  149. self.datastore.flush().await?;
  150. Ok(())
  151. }
  152. ///
  153. /// Return async receiver channel which can be used to receive T Messages
  154. /// from raft consensus
  155. ///
  156. pub fn receiver(&self) -> async_channel::Receiver<T> {
  157. self.commits_channel.1.clone()
  158. }
  159. ///
  160. /// Return async sender channel which can be used to broadcast T Messages
  161. /// to raft consensus
  162. ///
  163. pub fn sender(&self) -> async_channel::Sender<T> {
  164. self.msgs_channel.0.clone()
  165. }
  166. ///
  167. /// Return the raft node id
  168. ///
  169. pub fn id(&self) -> NodeId {
  170. self.id.clone()
  171. }
  172. async fn send_id_msg(&self) -> Result<()> {
  173. let id_msg = serialize(&NodeIdMsg { id: self.id.clone() });
  174. self.send(None, &id_msg, NetMsgMethod::NodeIdMsg, None).await?;
  175. Ok(())
  176. }
  177. async fn broadcast_msg(&mut self, msg: &T, msg_id: Option<u64>) -> Result<()> {
  178. match self.role {
  179. Role::Leader => {
  180. let msg = serialize(msg);
  181. let log = Log { msg, term: self.current_term()? };
  182. self.push_log(&log)?;
  183. self.acked_length.insert(&self.id, self.logs_len());
  184. }
  185. Role::Follower => {
  186. let b_msg = BroadcastMsgRequest(serialize(msg));
  187. self.send(
  188. Some(self.current_leader.clone()),
  189. &serialize(&b_msg),
  190. NetMsgMethod::BroadcastRequest,
  191. msg_id,
  192. )
  193. .await?;
  194. }
  195. Role::Candidate => {
  196. self.pending_msgs.push(msg.clone());
  197. }
  198. }
  199. debug!(target: "raft", "Role: {:?} Id: {:?}, broadcast a msg id: {:?} ", self.role, self.id, msg_id);
  200. Ok(())
  201. }
  202. async fn handle_method(&mut self, msg: NetMsg) -> Result<()> {
  203. match msg.method {
  204. NetMsgMethod::LogResponse => {
  205. let lr: LogResponse = deserialize(&msg.payload)?;
  206. self.receive_log_response(lr).await?;
  207. }
  208. NetMsgMethod::LogRequest => {
  209. self.last_heartbeat = Utc::now().timestamp();
  210. let lr: LogRequest = deserialize(&msg.payload)?;
  211. self.receive_log_request(lr).await?;
  212. }
  213. NetMsgMethod::VoteResponse => {
  214. let vr: VoteResponse = deserialize(&msg.payload)?;
  215. self.receive_vote_response(vr).await?;
  216. }
  217. NetMsgMethod::VoteRequest => {
  218. let vr: VoteRequest = deserialize(&msg.payload)?;
  219. self.receive_vote_request(vr).await?;
  220. }
  221. NetMsgMethod::BroadcastRequest => {
  222. let vr: BroadcastMsgRequest = deserialize(&msg.payload)?;
  223. let d: T = deserialize(&vr.0)?;
  224. self.broadcast_msg(&d, Some(msg.id)).await?;
  225. }
  226. NetMsgMethod::NodeIdMsg => {
  227. let node_id_msg: NodeIdMsg = deserialize(&msg.payload)?;
  228. if node_id_msg.id != self.id {
  229. self.nodes.lock().await.insert(node_id_msg.id, Utc::now().timestamp());
  230. }
  231. }
  232. }
  233. debug!(target: "raft", "Role: {:?} Id: {:?}, receive a msg with id: {} recipient_id: {:?} method: {:?} ",
  234. self.role, self.id, msg.id, &msg.recipient_id, &msg.method);
  235. Ok(())
  236. }
  237. pub(super) async fn send(
  238. &self,
  239. recipient_id: Option<NodeId>,
  240. payload: &[u8],
  241. method: NetMsgMethod,
  242. msg_id: Option<u64>,
  243. ) -> Result<()> {
  244. let random_id = if msg_id.is_some() { msg_id.unwrap() } else { OsRng.next_u64() };
  245. debug!(target: "raft","Role: {:?} Id: {:?}, send a msg with id: {} recipient_id: {:?} method: {:?} ",
  246. self.role, self.id, random_id, &recipient_id, &method);
  247. let net_msg = NetMsg { id: random_id, recipient_id, payload: payload.to_vec(), method };
  248. self.seen_msgs.lock().await.insert(random_id.to_string(), Utc::now().timestamp());
  249. self.p2p_sender.0.send(net_msg).await?;
  250. Ok(())
  251. }
  252. pub(super) fn reset_last_term(&mut self) -> Result<()> {
  253. self.last_term = 0;
  254. if let Some(log) = self.last_log()? {
  255. self.last_term = log.term;
  256. }
  257. Ok(())
  258. }
  259. pub(super) fn set_current_term(&mut self, i: &u64) -> Result<()> {
  260. self.datastore.current_term.insert(i)
  261. }
  262. pub(super) fn set_voted_for(&mut self, i: &Option<NodeId>) -> Result<()> {
  263. self.datastore.voted_for.insert(i)
  264. }
  265. pub(super) async fn push_commit(&mut self, commit: &[u8]) -> Result<()> {
  266. let commit: T = deserialize(commit)?;
  267. self.commits_channel.0.send(commit.clone()).await?;
  268. self.datastore.commits.insert(&commit)
  269. }
  270. pub(super) fn push_log(&mut self, log: &Log) -> Result<()> {
  271. self.datastore.logs.insert(log)
  272. }
  273. pub(super) fn push_logs(&mut self, logs: &Logs) -> Result<()> {
  274. self.datastore.logs.wipe_insert_all(&logs.to_vec())
  275. }
  276. pub(super) fn current_term(&self) -> Result<u64> {
  277. Ok(self.datastore.current_term.get_last()?.unwrap_or(0))
  278. }
  279. pub(super) fn voted_for(&self) -> Result<Option<NodeId>> {
  280. Ok(self.datastore.voted_for.get_last()?.flatten())
  281. }
  282. pub(super) fn commits_len(&self) -> u64 {
  283. self.datastore.commits.len()
  284. }
  285. fn logs(&self) -> Result<Logs> {
  286. Ok(Logs(self.datastore.logs.get_all()?))
  287. }
  288. pub(super) fn logs_len(&self) -> u64 {
  289. self.datastore.logs.len()
  290. }
  291. fn last_log(&self) -> Result<Option<Log>> {
  292. self.datastore.logs.get_last()
  293. }
  294. pub(super) fn get_log(&self, index: u64) -> Result<Log> {
  295. self.datastore.logs.get(index)
  296. }
  297. pub(super) fn slice_logs_from(&self, index: u64) -> Result<Option<Logs>> {
  298. let logs = self.logs()?;
  299. Ok(logs.slice_from(index))
  300. }
  301. pub(super) fn slice_logs_to(&self, index: u64) -> Result<Logs> {
  302. let logs = self.logs()?;
  303. Ok(logs.slice_to(index))
  304. }
  305. }