protocols/gossipsub: Implement unsub backoff spec changes (#2403)

Implements the changes specified by
libp2p/specs#383.

Co-authored-by: Max Inden <[email protected]>

protocols/gossipsub/CHANGELOG.md

@@ -13,11 +13,15 @@
 
 - Fix `GossipsubConfigBuilder::build()` requiring `&self` to live for `'static` (see [PR 2409])
 
+- Implement Unsubscribe backoff as per [libp2p specs PR 383] (see [PR 2403]).
+
 [PR 2346]: https://github.com/libp2p/rust-libp2p/pull/2346
 [PR 2339]: https://github.com/libp2p/rust-libp2p/pull/2339
 [PR 2327]: https://github.com/libp2p/rust-libp2p/pull/2327
 [PR 2408]: https://github.com/libp2p/rust-libp2p/pull/2408
 [PR 2409]: https://github.com/libp2p/rust-libp2p/pull/2409
+[PR 2403]: https://github.com/libp2p/rust-libp2p/pull/2403
+[libp2p specs PR 383]: https://github.com/libp2p/specs/pull/383
 
 # 0.34.0 [2021-11-16]
 

protocols/gossipsub/src/behaviour.rs

@@ -1049,6 +1049,7 @@ where
         topic_hash: &TopicHash,
         peer: &PeerId,
         do_px: bool,
+        on_unsubscribe: bool,
     ) -> GossipsubControlAction {
         if let Some((peer_score, ..)) = &mut self.peer_score {
             peer_score.prune(peer, topic_hash.clone());
@@ -1088,14 +1089,19 @@ where
             Vec::new()
         };
 
+        let backoff = if on_unsubscribe {
+            self.config.unsubscribe_backoff()
+        } else {
+            self.config.prune_backoff()
+        };
+
         // update backoff
-        self.backoffs
-            .update_backoff(topic_hash, peer, self.config.prune_backoff());
+        self.backoffs.update_backoff(topic_hash, peer, backoff);
 
         GossipsubControlAction::Prune {
             topic_hash: topic_hash.clone(),
             peers,
-            backoff: Some(self.config.prune_backoff().as_secs()),
+            backoff: Some(backoff.as_secs()),
         }
     }
 
@@ -1111,7 +1117,9 @@ where
             for peer in peers {
                 // Send a PRUNE control message
                 debug!("LEAVE: Sending PRUNE to peer: {:?}", peer);
-                let control = self.make_prune(topic_hash, &peer, self.config.do_px());
+                let on_unsubscribe = true;
+                let control =
+                    self.make_prune(topic_hash, &peer, self.config.do_px(), on_unsubscribe);
                 Self::control_pool_add(&mut self.control_pool, peer, control);
 
                 // If the peer did not previously exist in any mesh, inform the handler
@@ -1487,9 +1495,10 @@ where
 
         if !to_prune_topics.is_empty() {
             // build the prune messages to send
+            let on_unsubscribe = false;
             let prune_messages = to_prune_topics
                 .iter()
-                .map(|t| self.make_prune(t, peer_id, do_px))
+                .map(|t| self.make_prune(t, peer_id, do_px, on_unsubscribe))
                 .collect();
             // Send the prune messages to the peer
             debug!(
@@ -2598,6 +2607,9 @@ where
             // NOTE: In this case a peer has been added to a topic mesh, and removed from another.
             // It therefore must be in at least one mesh and we do not need to inform the handler
             // of its removal from another.
+
+            // The following prunes are not due to unsubscribing.
+            let on_unsubscribe = false;
             if let Some(topics) = to_prune.remove(&peer) {
                 let mut prunes = topics
                     .iter()
@@ -2606,6 +2618,7 @@ where
                             topic_hash,
                             &peer,
                             self.config.do_px() && !no_px.contains(&peer),
+                            on_unsubscribe,
                         )
                     })
                     .collect::<Vec<_>>();
@@ -2630,13 +2643,16 @@ where
         }
 
         // handle the remaining prunes
+        // The following prunes are not due to unsubscribing.
+        let on_unsubscribe = false;
         for (peer, topics) in to_prune.iter() {
             let mut remaining_prunes = Vec::new();
             for topic_hash in topics {
                 let prune = self.make_prune(
                     topic_hash,
                     peer,
                     self.config.do_px() && !no_px.contains(peer),
+                    on_unsubscribe,
                 );
                 remaining_prunes.push(prune);
                 // inform the handler

protocols/gossipsub/src/behaviour/tests.rs

@@ -2037,6 +2037,77 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_unsubscribe_backoff() {
+        const HEARTBEAT_INTERVAL: Duration = Duration::from_millis(100);
+        let config = GossipsubConfigBuilder::default()
+            .backoff_slack(1)
+            // ensure a prune_backoff > unsubscribe_backoff
+            .prune_backoff(Duration::from_secs(5))
+            .unsubscribe_backoff(1)
+            .heartbeat_interval(HEARTBEAT_INTERVAL)
+            .build()
+            .unwrap();
+
+        let topic = String::from("test");
+        // only one peer => mesh too small and will try to regraft as early as possible
+        let (mut gs, _, topics) = inject_nodes1()
+            .peer_no(1)
+            .topics(vec![topic.clone()])
+            .to_subscribe(true)
+            .gs_config(config)
+            .create_network();
+
+        let _ = gs.unsubscribe(&Topic::new(topic.clone()));
+
+        assert_eq!(
+            count_control_msgs(&gs, |_, m| match m {
+                GossipsubControlAction::Prune { backoff, .. } => backoff == &Some(1),
+                _ => false,
+            }),
+            1,
+            "Peer should be pruned with `unsubscribe_backoff`."
+        );
+
+        let _ = gs.subscribe(&Topic::new(topics[0].to_string()));
+
+        // forget all events until now
+        flush_events(&mut gs);
+
+        // call heartbeat
+        gs.heartbeat();
+
+        // Sleep for one second and apply 10 regular heartbeats (interval = 100ms).
+        for _ in 0..10 {
+            sleep(HEARTBEAT_INTERVAL);
+            gs.heartbeat();
+        }
+
+        // Check that no graft got created (we have backoff_slack = 1 therefore one more heartbeat
+        // is needed).
+        assert_eq!(
+            count_control_msgs(&gs, |_, m| match m {
+                GossipsubControlAction::Graft { .. } => true,
+                _ => false,
+            }),
+            0,
+            "Graft message created too early within backoff period"
+        );
+
+        // Heartbeat one more time this should graft now
+        sleep(HEARTBEAT_INTERVAL);
+        gs.heartbeat();
+
+        // check that graft got created
+        assert!(
+            count_control_msgs(&gs, |_, m| match m {
+                GossipsubControlAction::Graft { .. } => true,
+                _ => false,
+            }) > 0,
+            "No graft message was created after backoff period"
+        );
+    }
+
     #[test]
     fn test_flood_publish() {
         let config: GossipsubConfig = GossipsubConfig::default();

protocols/gossipsub/src/config.rs

@@ -77,6 +77,7 @@ pub struct GossipsubConfig {
     do_px: bool,
     prune_peers: usize,
     prune_backoff: Duration,
+    unsubscribe_backoff: Duration,
     backoff_slack: u32,
     flood_publish: bool,
     graft_flood_threshold: Duration,
@@ -276,6 +277,15 @@ impl GossipsubConfig {
         self.prune_backoff
     }
 
+    /// Controls the backoff time when unsubscribing from a topic.
+    ///
+    /// This is how long to wait before resubscribing to the topic. A short backoff period in case
+    /// of an unsubscribe event allows reaching a healthy mesh in a more timely manner. The default
+    /// is 10 seconds.
+    pub fn unsubscribe_backoff(&self) -> Duration {
+        self.unsubscribe_backoff
+    }
+
     /// Number of heartbeat slots considered as slack for backoffs. This gurantees that we wait
     /// at least backoff_slack heartbeats after a backoff is over before we try to graft. This
     /// solves problems occuring through high latencies. In particular if
@@ -421,6 +431,7 @@ impl Default for GossipsubConfigBuilder {
                 do_px: false,
                 prune_peers: 0, // NOTE: Increasing this currently has little effect until Signed records are implemented.
                 prune_backoff: Duration::from_secs(60),
+                unsubscribe_backoff: Duration::from_secs(10),
                 backoff_slack: 1,
                 flood_publish: true,
                 graft_flood_threshold: Duration::from_secs(10),
@@ -636,6 +647,16 @@ impl GossipsubConfigBuilder {
         self
     }
 
+    /// Controls the backoff time when unsubscribing from a topic.
+    ///
+    /// This is how long to wait before resubscribing to the topic. A short backoff period in case
+    /// of an unsubscribe event allows reaching a healthy mesh in a more timely manner. The default
+    /// is 10 seconds.
+    pub fn unsubscribe_backoff(&mut self, unsubscribe_backoff: u64) -> &mut Self {
+        self.config.unsubscribe_backoff = Duration::from_secs(unsubscribe_backoff);
+        self
+    }
+
     /// Number of heartbeat slots considered as slack for backoffs. This gurantees that we wait
     /// at least backoff_slack heartbeats after a backoff is over before we try to graft. This
     /// solves problems occuring through high latencies. In particular if
@@ -777,6 +798,11 @@ impl GossipsubConfigBuilder {
                 "The following inequality doesn't hold mesh_outbound_min <= self.config.mesh_n / 2",
             );
         }
+
+        if self.config.unsubscribe_backoff.as_millis() == 0 {
+            return Err("The unsubscribe_backoff parameter should be positive.");
+        }
+
         Ok(self.config.clone())
     }
 }