[consensus] sync improvements to help slow nodes sync better (#15364) (…

…#15376)

* [qs] grace period before GC committed batches

* [consensus] trigger sync based on remote LI timestamp

(cherry picked from commit d672085)

Co-authored-by: Balaji Arun <[email protected]>

consensus/src/quorum_store/batch_store.rs

@@ -117,6 +117,7 @@ pub struct BatchStore {
     batch_quota: usize,
     validator_signer: ValidatorSigner,
     persist_subscribers: DashMap<HashValue, Vec<oneshot::Sender<PersistedValue>>>,
+    expiration_buffer_usecs: u64,
 }
 
 impl BatchStore {
@@ -128,6 +129,7 @@ impl BatchStore {
         db_quota: usize,
         batch_quota: usize,
         validator_signer: ValidatorSigner,
+        expiration_buffer_usecs: u64,
     ) -> Self {
         let db_clone = db.clone();
         let batch_store = Self {
@@ -142,6 +144,7 @@ impl BatchStore {
             batch_quota,
             validator_signer,
             persist_subscribers: DashMap::new(),
+            expiration_buffer_usecs,
         };
         let db_content = db_clone
             .get_all_batches()
@@ -283,15 +286,19 @@ impl BatchStore {
     // pub(crate) for testing
     #[allow(clippy::unwrap_used)]
     pub(crate) fn clear_expired_payload(&self, certified_time: u64) -> Vec<HashValue> {
-        let expired_digests = self.expirations.lock().unwrap().expire(certified_time);
+        // To help slow nodes catch up via execution without going to state sync we keep the blocks for 60 extra seconds
+        // after the expiration time. This will help remote peers fetch batches that just expired but are within their
+        // execution window.
+        let expiration_time = certified_time.saturating_sub(self.expiration_buffer_usecs);
+        let expired_digests = self.expirations.lock().unwrap().expire(expiration_time);
         let mut ret = Vec::new();
         for h in expired_digests {
             let removed_value = match self.db_cache.entry(h) {
                 Occupied(entry) => {
                     // We need to check up-to-date expiration again because receiving the same
                     // digest with a higher expiration would update the persisted value and
                     // effectively extend the expiration.
-                    if entry.get().expiration() <= certified_time {
+                    if entry.get().expiration() <= expiration_time {
                         self.persist_subscribers.remove(entry.get().digest());
                         Some(entry.remove())
                     } else {

consensus/src/quorum_store/quorum_store_builder.rs

@@ -260,6 +260,7 @@ impl InnerBuilder {
             self.config.db_quota,
             self.config.batch_quota,
             signer,
+            Duration::from_secs(60).as_micros() as u64,
         ));
         self.batch_store = Some(batch_store.clone());
         let batch_reader = Arc::new(BatchReaderImpl::new(batch_store.clone(), batch_requester));

consensus/src/quorum_store/tests/batch_store_test.rs

@@ -36,6 +36,7 @@ pub fn batch_store_for_test(memory_quota: usize) -> Arc<BatchStore> {
         2001,         // db quota
         2001,         // batch quota
         signers[0].clone(),
+        0,
     ))
 }
 

testsuite/forge-cli/src/suites/realistic_environment.rs

@@ -303,9 +303,9 @@ pub(crate) fn realistic_env_max_load_test(
         .add_system_metrics_threshold(SystemMetricsThreshold::new(
             // Check that we don't use more than 18 CPU cores for 15% of the time.
             MetricsThreshold::new(25.0, 15),
-            // Memory starts around 7GB, and grows around 1.4GB/hr in this test.
+            // Memory starts around 8GB, and grows around 1.4GB/hr in this test.
             // Check that we don't use more than final expected memory for more than 20% of the time.
-            MetricsThreshold::new_gb(7.0 + 1.4 * (duration_secs as f64 / 3600.0), 20),
+            MetricsThreshold::new_gb(8.0 + 1.4 * (duration_secs as f64 / 3600.0), 20),
         ))
         .add_no_restarts()
         .add_wait_for_catchup_s(
@@ -316,7 +316,7 @@ pub(crate) fn realistic_env_max_load_test(
         .add_latency_threshold(4.5, LatencyType::P70)
         .add_chain_progress(StateProgressThreshold {
             max_non_epoch_no_progress_secs: 15.0,
-            max_epoch_no_progress_secs: 15.0,
+            max_epoch_no_progress_secs: 16.0,
             max_non_epoch_round_gap: 4,
             max_epoch_round_gap: 4,
         });