support follower read (#5051)

* raftstore,server: support follower read

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* raftstore: fix condition check of read on replica

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* raftstore: follower read waits for apply index reaches read index

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* add a test of waiting for read index

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* fix test_wait_for_apply_index

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* dec pending reads count after follower handle read index cmd

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* update comments

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* remove unused file

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* fix test_wait_for_apply_index

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* update comments

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* update test_wait_for_apply_index

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

* update dependency 'kvproto'

Signed-off-by: 5kbpers <tangminghua@pingcap.com>

src/raftstore/store/fsm/apply.rs

@@ -1097,6 +1097,10 @@ impl ApplyDelegate {
         ctx: &ApplyContext,
         req: &RaftCmdRequest,
     ) -> Result<(RaftCmdResponse, ApplyResult)> {
+        fail_point!("on_apply_write_cmd", self.id() == 3, |_| {
+            unimplemented!();
+        });
+
         let requests = req.get_requests();
         let mut responses = Vec::with_capacity(requests.len());
 

src/raftstore/store/fsm/peer.rs

@@ -2254,7 +2254,15 @@ impl<'a, T: Transport, C: PdClient> PeerFsmDelegate<'a, T, C> {
         // ReadIndex can be processed on the replicas.
         let is_read_index_request =
             request.len() == 1 && request[0].get_cmd_type() == CmdType::ReadIndex;
-        if !(self.fsm.peer.is_leader() || is_read_index_request) {
+        let mut read_only = true;
+        for r in msg.get_requests() {
+            match r.get_cmd_type() {
+                CmdType::Get | CmdType::Snap | CmdType::ReadIndex => (),
+                _ => read_only = false,
+            }
+        }
+        let allow_follower_read = read_only && msg.get_header().get_follower_read();
+        if !(self.fsm.peer.is_leader() || is_read_index_request || allow_follower_read) {
             self.ctx.raft_metrics.invalid_proposal.not_leader += 1;
             let leader = self.fsm.peer.get_peer_from_cache(leader_id);
             self.fsm.group_state = GroupState::Chaos;

src/raftstore/store/peer.rs

@@ -1009,6 +1009,14 @@ impl Peer {
             && !self.is_merging()
     }
 
+    fn ready_to_handle_unsafe_follower_read(&self, read_index: u64) -> bool {
+        // Wait until the follower applies all values before the read. There is still a
+        // problem if the leader applies fewer values than the follower, the follower read
+        // could get a newer value, and after that, the leader may read a stale value,
+        // which violates linearizability.
+        self.get_store().applied_index() >= read_index && !self.is_splitting() && !self.is_merging()
+    }
+
     #[inline]
     fn is_splitting(&self) -> bool {
         self.last_committed_split_idx > self.get_store().applied_index()
@@ -1354,18 +1362,24 @@ impl Peer {
                     && read.cmds[0].0.get_requests().len() == 1
                     && read.cmds[0].0.get_requests()[0].get_cmd_type() == CmdType::ReadIndex;
 
-                if !is_read_index_request {
-                    let term = self.term();
-                    // Only read index request is valid.
-                    for (_, cb) in read.cmds.drain(..) {
-                        apply::notify_stale_req(term, cb);
-                    }
-                } else {
+                let term = self.term();
+                if is_read_index_request {
                     for (req, cb) in read.cmds.drain(..) {
                         cb.invoke_read(self.handle_read(ctx, req, true, read.read_index));
                     }
+                    self.pending_reads.ready_cnt -= 1;
+                } else if self.ready_to_handle_unsafe_follower_read(read.read_index.unwrap()) {
+                    for (req, cb) in read.cmds.drain(..) {
+                        if req.get_header().get_follower_read() {
+                            cb.invoke_read(self.handle_read(ctx, req, true, read.read_index));
+                        } else {
+                            apply::notify_stale_req(term, cb);
+                        }
+                    }
+                    self.pending_reads.ready_cnt -= 1;
+                } else {
+                    self.pending_reads.reads.push_front(read);
                 }
-                self.pending_reads.ready_cnt -= 1;
             }
         }
     }

src/storage/kv/raftkv.rs

@@ -171,6 +171,7 @@ impl<S: RaftStoreRouter> RaftKv<S> {
             header.set_term(ctx.get_term());
         }
         header.set_sync_log(ctx.get_sync_log());
+        header.set_follower_read(ctx.get_follower_read());
         header
     }
 

tests/failpoints/cases/mod.rs

@@ -3,6 +3,7 @@
 mod test_bootstrap;
 mod test_conf_change;
 mod test_coprocessor;
+mod test_follower_read;
 mod test_merge;
 mod test_pending_peers;
 mod test_snap;

tests/failpoints/cases/test_follower_read.rs

@@ -0,0 +1,65 @@
+// Copyright 2019 TiKV Project Authors. Licensed under Apache-2.0.
+
+use std::sync::Arc;
+
+use fail;
+use std::time::Duration;
+use test_raftstore::*;
+use tikv_util::HandyRwLock;
+
+#[test]
+fn test_wait_for_apply_index() {
+    let _guard = crate::setup();
+    let mut cluster = new_server_cluster(0, 3);
+
+    // Increase the election tick to make this test case running reliably.
+    configure_for_lease_read(&mut cluster, Some(50), Some(10_000));
+    let pd_client = Arc::clone(&cluster.pd_client);
+    pd_client.disable_default_operator();
+
+    let r1 = cluster.run_conf_change();
+    cluster.must_put(b"k0", b"v0");
+    let p2 = new_peer(2, 2);
+    cluster.pd_client.must_add_peer(r1, p2.clone());
+    let p3 = new_peer(3, 3);
+    cluster.pd_client.must_add_peer(r1, p3.clone());
+    must_get_equal(&cluster.get_engine(3), b"k0", b"v0");
+
+    let region = cluster.get_region(b"k0");
+    cluster.must_transfer_leader(region.get_id(), p2.clone());
+
+    // Block all write cmd applying of Peer 3.
+    fail::cfg("on_apply_write_cmd", "sleep(2000)").unwrap();
+    cluster.must_put(b"k1", b"v1");
+    must_get_equal(&cluster.get_engine(2), b"k1", b"v1");
+
+    // Peer 3 does not apply the cmd of putting 'k1' right now, then the follower read must
+    // be blocked.
+    must_get_none(&cluster.get_engine(3), b"k1");
+    let mut request = new_request(
+        region.get_id(),
+        region.get_region_epoch().clone(),
+        vec![new_get_cf_cmd("default", b"k1")],
+        false,
+    );
+    request.mut_header().set_peer(p3.clone());
+    request.mut_header().set_follower_read(true);
+    let (cb, rx) = make_cb(&request);
+    cluster
+        .sim
+        .rl()
+        .async_command_on_node(3, request, cb)
+        .unwrap();
+    // Must timeout here
+    assert!(rx.recv_timeout(Duration::from_millis(500)).is_err());
+    fail::cfg("on_apply_write_cmd", "off").unwrap();
+
+    // After write cmd applied, the follower read will be executed.
+    match rx.recv_timeout(Duration::from_secs(3)) {
+        Ok(resp) => {
+            assert_eq!(resp.get_responses().len(), 1);
+            assert_eq!(resp.get_responses()[0].get_get().get_value(), b"v1");
+        }
+        Err(_) => panic!("follower read failed"),
+    }
+}

tests/integrations/storage/test_raftkv.rs

@@ -4,6 +4,8 @@ use kvproto::kvrpcpb::Context;
 
 use engine::IterOption;
 use engine::{CfName, CF_DEFAULT};
+use std::thread;
+use std::time;
 use test_raftstore::*;
 use tikv::storage::kv::*;
 use tikv::storage::{CFStatistics, Key};
@@ -123,6 +125,66 @@ fn test_read_index_on_replica() {
     );
 }
 
+#[test]
+fn test_read_on_replica() {
+    let count = 3;
+    let mut cluster = new_server_cluster(0, count);
+    cluster.run();
+
+    let k1 = b"k1";
+    let (k2, v2) = (b"k2", b"v2");
+    let (k3, v3) = (b"k3", b"v3");
+    let (k4, v4) = (b"k4", b"v4");
+
+    // make sure leader has been elected.
+    assert_eq!(cluster.must_get(k1), None);
+
+    let region = cluster.get_region(b"");
+    let leader = cluster.leader_of_region(region.get_id()).unwrap();
+    let leader_storage = cluster.sim.rl().storages[&leader.get_id()].clone();
+
+    let mut leader_ctx = Context::new();
+    leader_ctx.set_region_id(region.get_id());
+    leader_ctx.set_region_epoch(region.get_region_epoch().clone());
+    leader_ctx.set_peer(leader.clone());
+
+    // write some data
+    let peers = region.get_peers();
+    assert_none(&leader_ctx, &leader_storage, k2);
+    must_put(&leader_ctx, &leader_storage, k2, v2);
+
+    // read on follower
+    let mut follower_peer = None;
+    let mut follower_id = 0;
+    for p in peers {
+        if p.get_id() != leader.get_id() {
+            follower_id = p.get_id();
+            follower_peer = Some(p.clone());
+            break;
+        }
+    }
+
+    assert!(follower_peer.is_some());
+    let mut follower_ctx = Context::new();
+    follower_ctx.set_region_id(region.get_id());
+    follower_ctx.set_region_epoch(region.get_region_epoch().clone());
+    follower_ctx.set_peer(follower_peer.as_ref().unwrap().clone());
+    follower_ctx.set_follower_read(true);
+    let follower_storage = cluster.sim.rl().storages[&follower_id].clone();
+    assert_has(&follower_ctx, &follower_storage, k2, v2);
+
+    must_put(&leader_ctx, &leader_storage, k3, v3);
+    assert_has(&follower_ctx, &follower_storage, k3, v3);
+
+    cluster.stop_node(follower_id);
+    must_put(&leader_ctx, &leader_storage, k4, v4);
+    cluster.run_node(follower_id).unwrap();
+    let follower_storage = cluster.sim.rl().storages[&follower_id].clone();
+    // sleep to ensure the follower has received a heartbeat from the leader
+    thread::sleep(time::Duration::from_millis(300));
+    assert_has(&follower_ctx, &follower_storage, k4, v4);
+}
+
 fn must_put<E: Engine>(ctx: &Context, engine: &E, key: &[u8], value: &[u8]) {
     engine.put(ctx, Key::from_raw(key), value.to_vec()).unwrap();
 }