consensus.rs 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357
  1. use async_std::{
  2. sync::{Arc, Mutex},
  3. task,
  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. self, 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_node_id_loop(sender: async_channel::Sender<()>, timeout: i64) -> Result<()> {
  29. loop {
  30. util::sleep(timeout as u64).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. p2p_sender: Sender,
  44. msgs_channel: Channel<T>,
  45. commits_channel: Channel<T>,
  46. datastore: DataStore<T>,
  47. seen_msgs: Arc<Mutex<FxHashMap<String, i64>>>,
  48. settings: RaftSettings,
  49. }
  50. impl<T: Decodable + Encodable + Clone> Raft<T> {
  51. pub fn new(
  52. settings: RaftSettings,
  53. seen_msgs: Arc<Mutex<FxHashMap<String, i64>>>,
  54. ) -> Result<Self> {
  55. if settings.datastore_path.to_str().is_none() {
  56. error!(target: "raft", "datastore path is incorrect");
  57. return Err(Error::ParseFailed("unable to parse pathbuf to str"))
  58. };
  59. let datastore = DataStore::new(settings.datastore_path.to_str().unwrap())?;
  60. // broadcasting channels
  61. let msgs_channel = async_channel::unbounded::<T>();
  62. let commits_channel = async_channel::unbounded::<T>();
  63. let p2p_sender = async_channel::unbounded::<NetMsg>();
  64. let id = match datastore.id.get_last()? {
  65. Some(_id) => _id,
  66. None => {
  67. let id = NodeId(gen_id(30));
  68. datastore.id.insert(&id)?;
  69. id
  70. }
  71. };
  72. let role = Role::Follower;
  73. Ok(Self {
  74. id,
  75. role,
  76. current_leader: NodeId("".into()),
  77. votes_received: vec![],
  78. sent_length: MapLength(FxHashMap::default()),
  79. acked_length: MapLength(FxHashMap::default()),
  80. nodes: Arc::new(Mutex::new(FxHashMap::default())),
  81. last_term: 0,
  82. p2p_sender,
  83. msgs_channel,
  84. commits_channel,
  85. datastore,
  86. seen_msgs,
  87. settings,
  88. })
  89. }
  90. ///
  91. /// Run raft consensus and wait stop_signal channel to terminate
  92. ///
  93. pub async fn run(
  94. &mut self,
  95. p2p: net::P2pPtr,
  96. p2p_recv_channel: async_channel::Receiver<NetMsg>,
  97. executor: Arc<Executor<'_>>,
  98. stop_signal: async_channel::Receiver<()>,
  99. ) -> Result<()> {
  100. let p2p_send_task = executor.spawn(p2p_send_loop(self.p2p_sender.1.clone(), p2p.clone()));
  101. let prune_seen_messages_task = executor.spawn(prune_map::<String>(
  102. self.seen_msgs.clone(),
  103. self.settings.prun_messages_duration,
  104. ));
  105. let prune_nodes_id_task = executor
  106. .spawn(prune_map::<NodeId>(self.nodes.clone(), self.settings.prun_nodes_ids_duration));
  107. let (node_id_sx, node_id_rv) = async_channel::unbounded::<()>();
  108. let send_node_id_loop_task =
  109. executor.spawn(send_node_id_loop(node_id_sx, self.settings.node_id_timeout));
  110. let mut rng = rand::thread_rng();
  111. let broadcast_msg_rv = self.msgs_channel.1.clone();
  112. loop {
  113. let timeout = if self.role == Role::Leader {
  114. self.settings.heartbeat_timeout
  115. } else {
  116. rng.gen_range(0..self.settings.timeout) + self.settings.timeout
  117. };
  118. let timeout = Duration::from_millis(timeout);
  119. let result: Result<()>;
  120. select! {
  121. m = p2p_recv_channel.recv().fuse() => result = self.handle_method(m?).await,
  122. m = broadcast_msg_rv.recv().fuse() => result = self.broadcast_msg(&m?,None).await,
  123. _ = node_id_rv.recv().fuse() => result = self.send_node_id_msg().await,
  124. _ = task::sleep(timeout).fuse() => {
  125. result = if self.role == Role::Leader {
  126. self.send_heartbeat().await
  127. }else {
  128. self.send_vote_request().await
  129. };
  130. },
  131. _ = stop_signal.recv().fuse() => break,
  132. }
  133. match result {
  134. Ok(_) => {}
  135. Err(e) => warn!(target: "raft", "warn: {}", e),
  136. }
  137. }
  138. warn!(target: "raft", "Raft Terminating...");
  139. p2p_send_task.cancel().await;
  140. prune_seen_messages_task.cancel().await;
  141. prune_nodes_id_task.cancel().await;
  142. send_node_id_loop_task.cancel().await;
  143. self.datastore.flush().await?;
  144. Ok(())
  145. }
  146. ///
  147. /// Return async receiver channel which can be used to receive T Messages
  148. /// from raft consensus
  149. ///
  150. pub fn receiver(&self) -> async_channel::Receiver<T> {
  151. self.commits_channel.1.clone()
  152. }
  153. ///
  154. /// Return async sender channel which can be used to broadcast T Messages
  155. /// to raft consensus
  156. ///
  157. pub fn sender(&self) -> async_channel::Sender<T> {
  158. self.msgs_channel.0.clone()
  159. }
  160. ///
  161. /// Return the raft node id
  162. ///
  163. pub fn id(&self) -> NodeId {
  164. self.id.clone()
  165. }
  166. async fn send_node_id_msg(&self) -> Result<()> {
  167. let node_id_msg = serialize(&NodeIdMsg { id: self.id.clone() });
  168. self.send(None, &node_id_msg, NetMsgMethod::NodeIdMsg, None).await?;
  169. Ok(())
  170. }
  171. async fn broadcast_msg(&mut self, msg: &T, msg_id: Option<u64>) -> Result<()> {
  172. if self.role == Role::Leader {
  173. let msg = serialize(msg);
  174. let log = Log { msg, term: self.current_term()? };
  175. self.push_log(&log)?;
  176. self.acked_length.insert(&self.id, self.logs_len());
  177. } else {
  178. let b_msg = BroadcastMsgRequest(serialize(msg));
  179. self.send(
  180. Some(self.current_leader.clone()),
  181. &serialize(&b_msg),
  182. NetMsgMethod::BroadcastRequest,
  183. msg_id,
  184. )
  185. .await?;
  186. }
  187. debug!(target: "raft", "Role: {:?} Id: {:?}, broadcast a msg id: {:?} ", self.role, self.id, msg_id);
  188. Ok(())
  189. }
  190. async fn handle_method(&mut self, msg: NetMsg) -> Result<()> {
  191. match msg.method {
  192. NetMsgMethod::LogResponse => {
  193. let lr: LogResponse = deserialize(&msg.payload)?;
  194. self.receive_log_response(lr).await?;
  195. }
  196. NetMsgMethod::LogRequest => {
  197. let lr: LogRequest = deserialize(&msg.payload)?;
  198. self.receive_log_request(lr).await?;
  199. }
  200. NetMsgMethod::VoteResponse => {
  201. let vr: VoteResponse = deserialize(&msg.payload)?;
  202. self.receive_vote_response(vr).await?;
  203. }
  204. NetMsgMethod::VoteRequest => {
  205. let vr: VoteRequest = deserialize(&msg.payload)?;
  206. self.receive_vote_request(vr).await?;
  207. }
  208. NetMsgMethod::BroadcastRequest => {
  209. let vr: BroadcastMsgRequest = deserialize(&msg.payload)?;
  210. let d: T = deserialize(&vr.0)?;
  211. self.broadcast_msg(&d, Some(msg.id)).await?;
  212. }
  213. NetMsgMethod::NodeIdMsg => {
  214. let node_id_msg: NodeIdMsg = deserialize(&msg.payload)?;
  215. if node_id_msg.id != self.id {
  216. self.nodes.lock().await.insert(node_id_msg.id, Utc::now().timestamp());
  217. }
  218. }
  219. }
  220. debug!(target: "raft", "Role: {:?} Id: {:?}, receive a msg with id: {} recipient_id: {:?} method: {:?} ",
  221. self.role, self.id, msg.id, &msg.recipient_id, &msg.method);
  222. Ok(())
  223. }
  224. pub(super) async fn send(
  225. &self,
  226. recipient_id: Option<NodeId>,
  227. payload: &[u8],
  228. method: NetMsgMethod,
  229. msg_id: Option<u64>,
  230. ) -> Result<()> {
  231. let random_id = if msg_id.is_some() { msg_id.unwrap() } else { OsRng.next_u64() };
  232. debug!(target: "raft","Role: {:?} Id: {:?}, send a msg with id: {} recipient_id: {:?} method: {:?} ",
  233. self.role, self.id, random_id, &recipient_id, &method);
  234. let net_msg = NetMsg { id: random_id, recipient_id, payload: payload.to_vec(), method };
  235. self.seen_msgs.lock().await.insert(random_id.to_string(), Utc::now().timestamp());
  236. self.p2p_sender.0.send(net_msg).await?;
  237. Ok(())
  238. }
  239. pub(super) fn reset_last_term(&mut self) -> Result<()> {
  240. self.last_term = 0;
  241. if let Some(log) = self.last_log()? {
  242. self.last_term = log.term;
  243. }
  244. Ok(())
  245. }
  246. pub(super) fn set_current_term(&mut self, i: &u64) -> Result<()> {
  247. self.datastore.current_term.insert(i)
  248. }
  249. pub(super) fn set_voted_for(&mut self, i: &Option<NodeId>) -> Result<()> {
  250. self.datastore.voted_for.insert(i)
  251. }
  252. pub(super) async fn push_commit(&mut self, commit: &[u8]) -> Result<()> {
  253. let commit: T = deserialize(commit)?;
  254. self.commits_channel.0.send(commit.clone()).await?;
  255. self.datastore.commits.insert(&commit)
  256. }
  257. pub(super) fn push_log(&mut self, log: &Log) -> Result<()> {
  258. self.datastore.logs.insert(log)
  259. }
  260. pub(super) fn push_logs(&mut self, logs: &Logs) -> Result<()> {
  261. self.datastore.logs.wipe_insert_all(&logs.to_vec())
  262. }
  263. pub(super) fn current_term(&self) -> Result<u64> {
  264. Ok(self.datastore.current_term.get_last()?.unwrap_or(0))
  265. }
  266. pub(super) fn voted_for(&self) -> Result<Option<NodeId>> {
  267. Ok(self.datastore.voted_for.get_last()?.flatten())
  268. }
  269. pub(super) fn commits_len(&self) -> u64 {
  270. self.datastore.commits.len()
  271. }
  272. fn logs(&self) -> Result<Logs> {
  273. Ok(Logs(self.datastore.logs.get_all()?))
  274. }
  275. pub(super) fn logs_len(&self) -> u64 {
  276. self.datastore.logs.len()
  277. }
  278. fn last_log(&self) -> Result<Option<Log>> {
  279. self.datastore.logs.get_last()
  280. }
  281. pub(super) fn get_log(&self, index: u64) -> Result<Log> {
  282. self.datastore.logs.get(index)
  283. }
  284. pub(super) fn slice_logs_from(&self, index: u64) -> Result<Option<Logs>> {
  285. let logs = self.logs()?;
  286. Ok(logs.slice_from(index))
  287. }
  288. pub(super) fn slice_logs_to(&self, index: u64) -> Result<Logs> {
  289. let logs = self.logs()?;
  290. Ok(logs.slice_to(index))
  291. }
  292. }