Merge 6fb21d1424 into c1f465e009

Cargo.lock

@@ -844,6 +844,7 @@ dependencies = [
 name = "channel"
 version = "0.1.0"
 dependencies = [
+ "metrics",
  "tokio",
 ]
 
@@ -4538,6 +4539,8 @@ dependencies = [
  "futures-core",
  "futures-util",
  "jsonrpsee",
+ "lazy_static",
+ "metrics",
  "serde_json",
  "shared_types",
  "storage",

common/channel/Cargo.toml

@@ -5,3 +5,4 @@ edition = "2021"
 
 [dependencies]
 tokio = { version = "1.19.2", features = ["sync", "time"] }
+metrics = { workspace = true }

common/channel/src/channel.rs

@@ -1,6 +1,8 @@
 use crate::error::Error;
-use std::time::Duration;
+use metrics::{register_meter_with_group, Counter, CounterUsize, Histogram, Meter, Sample};
-use tokio::sync::mpsc::error::TryRecvError;
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+use tokio::sync::mpsc::error::{SendError, TryRecvError};
 use tokio::sync::{mpsc, oneshot};
 use tokio::time::timeout;
 
@@ -19,56 +21,139 @@ pub struct Channel<N, Req, Res> {
 }
 
 impl<N, Req, Res> Channel<N, Req, Res> {
-    pub fn unbounded() -> (Sender<N, Req, Res>, Receiver<N, Req, Res>) {
+    pub fn unbounded(name: &str) -> (Sender<N, Req, Res>, Receiver<N, Req, Res>) {
         let (sender, receiver) = mpsc::unbounded_channel();
-        (Sender { chan: sender }, Receiver { chan: receiver })
+
+        let metrics_group = format!("common_channel_{}", name);
+        let metrics_queued = CounterUsize::register_with_group(metrics_group.as_str(), "size");
+
+        (
+            Sender {
+                chan: sender,
+                metrics_send_qps: register_meter_with_group(metrics_group.as_str(), "send"),
+                metrics_queued: metrics_queued.clone(),
+                metrics_timeout: CounterUsize::register_with_group(
+                    metrics_group.as_str(),
+                    "timeout",
+                ),
+            },
+            Receiver {
+                chan: receiver,
+                metrics_recv_qps: register_meter_with_group(metrics_group.as_str(), "recv"),
+                metrics_queued,
+                metrics_queue_latency: Sample::ExpDecay(0.015).register_with_group(
+                    metrics_group.as_str(),
+                    "latency",
+                    1024,
+                ),
+            },
+        )
+    }
+}
+
+enum TimedMessage<N, Req, Res> {
+    Notification(Instant, N),
+    Request(Instant, Req, ResponseSender<Res>),
+}
+
+impl<N, Req, Res> From<Message<N, Req, Res>> for TimedMessage<N, Req, Res> {
+    fn from(value: Message<N, Req, Res>) -> Self {
+        match value {
+            Message::Notification(n) => TimedMessage::Notification(Instant::now(), n),
+            Message::Request(req, res) => TimedMessage::Request(Instant::now(), req, res),
+        }
+    }
+}
+
+impl<N, Req, Res> TimedMessage<N, Req, Res> {
+    fn into_message(self) -> (Instant, Message<N, Req, Res>) {
+        match self {
+            TimedMessage::Notification(since, n) => (since, Message::Notification(n)),
+            TimedMessage::Request(since, req, res) => (since, Message::Request(req, res)),
+        }
     }
 }
 
 pub struct Sender<N, Req, Res> {
-    chan: mpsc::UnboundedSender<Message<N, Req, Res>>,
+    chan: mpsc::UnboundedSender<TimedMessage<N, Req, Res>>,
+
+    metrics_send_qps: Arc<dyn Meter>,
+    metrics_queued: Arc<dyn Counter<usize>>,
+    metrics_timeout: Arc<dyn Counter<usize>>,
 }
 
 impl<N, Req, Res> Clone for Sender<N, Req, Res> {
     fn clone(&self) -> Self {
         Sender {
             chan: self.chan.clone(),
+            metrics_send_qps: self.metrics_send_qps.clone(),
+            metrics_queued: self.metrics_queued.clone(),
+            metrics_timeout: self.metrics_timeout.clone(),
         }
     }
 }
 
 impl<N, Req, Res> Sender<N, Req, Res> {
     pub fn notify(&self, msg: N) -> Result<(), Error<N, Req, Res>> {
-        self.chan
+        self.send(Message::Notification(msg))
-            .send(Message::Notification(msg))
-            .map_err(|e| Error::SendError(e))
     }
 
     pub async fn request(&self, request: Req) -> Result<Res, Error<N, Req, Res>> {
         let (sender, receiver) = oneshot::channel();
 
-        self.chan
+        self.send(Message::Request(request, sender))?;
-            .send(Message::Request(request, sender))
-            .map_err(|e| Error::SendError(e))?;
 
         timeout(DEFAULT_REQUEST_TIMEOUT, receiver)
             .await
-            .map_err(|_| Error::TimeoutError)?
+            .map_err(|_| {
+                self.metrics_timeout.inc(1);
+                Error::TimeoutError
+            })?
             .map_err(|e| Error::RecvError(e))
     }
+
+    fn send(&self, message: Message<N, Req, Res>) -> Result<(), Error<N, Req, Res>> {
+        match self.chan.send(message.into()) {
+            Ok(()) => {
+                self.metrics_send_qps.mark(1);
+                self.metrics_queued.inc(1);
+                Ok(())
+            }
+            Err(e) => {
+                let (_, msg) = e.0.into_message();
+                Err(Error::SendError(SendError(msg)))
+            }
+        }
+    }
 }
 
 pub struct Receiver<N, Req, Res> {
-    chan: mpsc::UnboundedReceiver<Message<N, Req, Res>>,
+    chan: mpsc::UnboundedReceiver<TimedMessage<N, Req, Res>>,
+
+    metrics_recv_qps: Arc<dyn Meter>,
+    metrics_queued: Arc<dyn Counter<usize>>,
+    metrics_queue_latency: Arc<dyn Histogram>,
 }
 
 impl<N, Req, Res> Receiver<N, Req, Res> {
     pub async fn recv(&mut self) -> Option<Message<N, Req, Res>> {
-        self.chan.recv().await
+        let data = self.chan.recv().await?;
+        Some(self.on_recv_data(data))
     }
 
     pub fn try_recv(&mut self) -> Result<Message<N, Req, Res>, TryRecvError> {
-        self.chan.try_recv()
+        let data = self.chan.try_recv()?;
+        Ok(self.on_recv_data(data))
+    }
+
+    fn on_recv_data(&self, data: TimedMessage<N, Req, Res>) -> Message<N, Req, Res> {
+        self.metrics_recv_qps.mark(1);
+        self.metrics_queued.dec(1);
+
+        let (since, msg) = data.into_message();
+        self.metrics_queue_latency.update_since(since);
+
+        msg
     }
 }
 
@@ -91,7 +176,7 @@ mod tests {
 
     #[tokio::test]
     async fn request_response() {
-        let (tx, mut rx) = Channel::<Notification, Request, Response>::unbounded();
+        let (tx, mut rx) = Channel::<Notification, Request, Response>::unbounded("test");
 
         let task1 = async move {
             match rx.recv().await.expect("not dropped") {

node/log_entry_sync/Cargo.toml

@@ -22,3 +22,5 @@ contract-interface = { path = "../../common/contract-interface" }
 futures-core = "0.3.28"
 futures-util = "0.3.28"
 thiserror = "1.0.44"
+lazy_static = "1.4.0"
+metrics = { workspace = true }

node/log_entry_sync/src/sync_manager/metrics.rs

@@ -0,0 +1,7 @@
+use std::sync::Arc;
+
+use metrics::{register_timer, Timer};
+
+lazy_static::lazy_static! {
+    pub static ref STORE_PUT_TX: Arc<dyn Timer> = register_timer("log_entry_sync_store_put_tx");
+}

node/log_entry_sync/src/sync_manager/mod.rs

@@ -11,7 +11,7 @@ use std::collections::BTreeMap;
 use std::fmt::Debug;
 use std::future::Future;
 use std::sync::Arc;
-use std::time::Duration;
+use std::time::{Duration, Instant};
 use storage::log_store::{tx_store::BlockHashAndSubmissionIndex, Store};
 use task_executor::{ShutdownReason, TaskExecutor};
 use tokio::sync::broadcast;
@@ -358,7 +358,11 @@ impl LogSyncManager {
     }
 
     async fn put_tx_inner(&mut self, tx: Transaction) -> bool {
-        if let Err(e) = self.store.put_tx(tx.clone()) {
+        let start_time = Instant::now();
+        let result = self.store.put_tx(tx.clone());
+        metrics::STORE_PUT_TX.update_since(start_time);
+
+        if let Err(e) = result {
             error!("put_tx error: e={:?}", e);
             false
         } else {
@@ -458,3 +462,4 @@ pub(crate) mod config;
 mod data_cache;
 mod log_entry_fetcher;
 mod log_query;
+mod metrics;

node/router/src/libp2p_event_handler.rs

@@ -815,7 +815,7 @@ mod tests {
             let runtime = TestRuntime::default();
             let (network_globals, keypair) = Context::new_network_globals();
             let (network_send, network_recv) = mpsc::unbounded_channel();
-            let (sync_send, sync_recv) = channel::Channel::unbounded();
+            let (sync_send, sync_recv) = channel::Channel::unbounded("test");
             let (chunk_pool_send, _chunk_pool_recv) = mpsc::unbounded_channel();
             let store = LogManager::memorydb(LogConfig::default()).unwrap();
             Self {

node/rpc/src/admin/impl.rs

@@ -4,7 +4,7 @@ use crate::{error, Context};
 use futures::prelude::*;
 use jsonrpsee::core::async_trait;
 use jsonrpsee::core::RpcResult;
-use metrics::DEFAULT_REGISTRY;
+use metrics::{DEFAULT_GROUPING_REGISTRY, DEFAULT_REGISTRY};
 use network::{multiaddr::Protocol, Multiaddr};
 use std::collections::{BTreeMap, HashMap};
 use std::net::IpAddr;
@@ -266,6 +266,21 @@ impl RpcServer for RpcServerImpl {
             }
         }
 
+        for (group_name, metrics) in DEFAULT_GROUPING_REGISTRY.read().get_all() {
+            for (metric_name, metric) in metrics.iter() {
+                let name = format!("{}.{}", group_name, metric_name);
+                match &maybe_prefix {
+                    Some(prefix) if !name.starts_with(prefix) => {}
+                    _ => {
+                        result.insert(
+                            name,
+                            format!("{} {}", metric.get_type(), metric.get_value()),
+                        );
+                    }
+                }
+            }
+        }
+
         Ok(result)
     }
 }

node/sync/src/auto_sync/batcher.rs

@@ -1,7 +1,7 @@
-use crate::{controllers::SyncState, Config, SyncRequest, SyncResponse, SyncSender};
+use crate::{controllers::SyncState, SyncRequest, SyncResponse, SyncSender};
 use anyhow::{bail, Result};
 use serde::{Deserialize, Serialize};
-use std::{collections::HashSet, fmt::Debug, sync::Arc};
+use std::{collections::HashSet, fmt::Debug, sync::Arc, time::Duration};
 use storage_async::Store;
 use tokio::sync::RwLock;
 
@@ -15,18 +15,23 @@ pub enum SyncResult {
 /// Supports to sync files concurrently.
 #[derive(Clone)]
 pub struct Batcher {
-    pub(crate) config: Config,
     capacity: usize,
+    find_peer_timeout: Duration,
     tasks: Arc<RwLock<HashSet<u64>>>, // files to sync
     store: Store,
     sync_send: SyncSender,
 }
 
 impl Batcher {
-    pub fn new(config: Config, capacity: usize, store: Store, sync_send: SyncSender) -> Self {
+    pub fn new(
+        capacity: usize,
+        find_peer_timeout: Duration,
+        store: Store,
+        sync_send: SyncSender,
+    ) -> Self {
         Self {
-            config,
             capacity,
+            find_peer_timeout,
             tasks: Default::default(),
             store,
             sync_send,
@@ -128,7 +133,7 @@ impl Batcher {
 
             // finding peers timeout
             Some(SyncState::FindingPeers { origin, .. })
-                if origin.elapsed() > self.config.find_peer_timeout =>
+                if origin.elapsed() > self.find_peer_timeout =>
             {
                 debug!(%tx_seq, "Terminate file sync due to finding peers timeout");
                 self.terminate_file_sync(tx_seq, false).await;
@@ -137,7 +142,7 @@ impl Batcher {
 
             // connecting peers timeout
             Some(SyncState::ConnectingPeers { origin, .. })
-                if origin.elapsed() > self.config.find_peer_timeout =>
+                if origin.elapsed() > self.find_peer_timeout =>
             {
                 debug!(%tx_seq, "Terminate file sync due to connecting peers timeout");
                 self.terminate_file_sync(tx_seq, false).await;

node/sync/src/auto_sync/batcher_random.rs

@@ -22,6 +22,7 @@ pub struct RandomBatcherState {
 
 #[derive(Clone)]
 pub struct RandomBatcher {
+    config: Config,
     batcher: Batcher,
     sync_store: Arc<SyncStore>,
 }
@@ -34,7 +35,13 @@ impl RandomBatcher {
         sync_store: Arc<SyncStore>,
     ) -> Self {
         Self {
-            batcher: Batcher::new(config, config.max_random_workers, store, sync_send),
+            config,
+            batcher: Batcher::new(
+                config.max_random_workers,
+                config.random_find_peer_timeout,
+                store,
+                sync_send,
+            ),
             sync_store,
         }
     }
@@ -56,7 +63,7 @@ impl RandomBatcher {
             // disable file sync until catched up
             if !catched_up.load(Ordering::Relaxed) {
                 trace!("Cannot sync file in catch-up phase");
-                sleep(self.batcher.config.auto_sync_idle_interval).await;
+                sleep(self.config.auto_sync_idle_interval).await;
                 continue;
             }
 
@@ -73,11 +80,11 @@ impl RandomBatcher {
                         "File sync still in progress or idle, state = {:?}",
                         self.get_state().await
                     );
-                    sleep(self.batcher.config.auto_sync_idle_interval).await;
+                    sleep(self.config.auto_sync_idle_interval).await;
                 }
                 Err(err) => {
                     warn!(%err, "Failed to sync file once, state = {:?}", self.get_state().await);
-                    sleep(self.batcher.config.auto_sync_error_interval).await;
+                    sleep(self.config.auto_sync_error_interval).await;
                 }
             }
         }
@@ -96,7 +103,7 @@ impl RandomBatcher {
 
         debug!(%tx_seq, ?sync_result, "Completed to sync file, state = {:?}", self.get_state().await);
         match sync_result {
-            SyncResult::Completed => metrics::RANDOM_SYNC_RESULT_COMPLETED.inc(1),
+            SyncResult::Completed => metrics::RANDOM_SYNC_RESULT_COMPLETED.mark(1),
             SyncResult::Failed => metrics::RANDOM_SYNC_RESULT_FAILED.inc(1),
             SyncResult::Timeout => metrics::RANDOM_SYNC_RESULT_TIMEOUT.inc(1),
         }

node/sync/src/auto_sync/batcher_serial.rs

@@ -23,6 +23,7 @@ use tokio::{
 /// Supports to sync files in sequence concurrently.
 #[derive(Clone)]
 pub struct SerialBatcher {
+    config: Config,
     batcher: Batcher,
 
     /// Next tx seq to sync.
@@ -80,13 +81,17 @@ impl SerialBatcher {
         sync_send: SyncSender,
         sync_store: Arc<SyncStore>,
     ) -> Result<Self> {
-        let capacity = config.max_sequential_workers;
-
         // continue file sync from break point in db
         let (next_tx_seq, max_tx_seq) = sync_store.get_tx_seq_range().await?;
 
         Ok(Self {
-            batcher: Batcher::new(config, capacity, store, sync_send),
+            config,
+            batcher: Batcher::new(
+                config.max_sequential_workers,
+                config.sequential_find_peer_timeout,
+                store,
+                sync_send,
+            ),
             next_tx_seq: Arc::new(AtomicU64::new(next_tx_seq.unwrap_or(0))),
             max_tx_seq: Arc::new(AtomicU64::new(max_tx_seq.unwrap_or(u64::MAX))),
             pending_completed_txs: Default::default(),
@@ -136,7 +141,7 @@ impl SerialBatcher {
             // disable file sync until catched up
             if !catched_up.load(Ordering::Relaxed) {
                 trace!("Cannot sync file in catch-up phase");
-                sleep(self.batcher.config.auto_sync_idle_interval).await;
+                sleep(self.config.auto_sync_idle_interval).await;
                 continue;
             }
 
@@ -157,11 +162,11 @@ impl SerialBatcher {
                         "File sync still in progress or idle, state = {:?}",
                         self.get_state().await
                     );
-                    sleep(self.batcher.config.auto_sync_idle_interval).await;
+                    sleep(self.config.auto_sync_idle_interval).await;
                 }
                 Err(err) => {
                     warn!(%err, "Failed to sync file once, state = {:?}", self.get_state().await);
-                    sleep(self.batcher.config.auto_sync_error_interval).await;
+                    sleep(self.config.auto_sync_error_interval).await;
                 }
             }
         }
@@ -257,7 +262,7 @@ impl SerialBatcher {
 
         info!(%tx_seq, ?sync_result, "Completed to sync file, state = {:?}", self.get_state().await);
         match sync_result {
-            SyncResult::Completed => metrics::SEQUENTIAL_SYNC_RESULT_COMPLETED.inc(1),
+            SyncResult::Completed => metrics::SEQUENTIAL_SYNC_RESULT_COMPLETED.mark(1),
             SyncResult::Failed => metrics::SEQUENTIAL_SYNC_RESULT_FAILED.inc(1),
             SyncResult::Timeout => metrics::SEQUENTIAL_SYNC_RESULT_TIMEOUT.inc(1),
         }

node/sync/src/auto_sync/metrics.rs

@@ -1,6 +1,6 @@
 use std::sync::Arc;
 
-use metrics::{Counter, CounterUsize, Gauge, GaugeUsize, Histogram, Sample};
+use metrics::{register_meter, Counter, CounterUsize, Gauge, GaugeUsize, Histogram, Meter, Sample};
 
 lazy_static::lazy_static! {
     // sequential auto sync
@@ -9,8 +9,7 @@ lazy_static::lazy_static! {
     pub static ref SEQUENTIAL_STATE_TXS_PENDING: Arc<dyn Histogram> = Sample::ExpDecay(0.015).register("sync_auto_sequential_state_txs_pending", 1024);
     pub static ref SEQUENTIAL_STATE_GAP_NEXT_DB: Arc<dyn Gauge<usize>> = GaugeUsize::register("sync_auto_sequential_state_gap_next_db");
 
-    pub static ref SEQUENTIAL_SYNC_RESULT_TOTAL: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_sequential_sync_result_total");
+    pub static ref SEQUENTIAL_SYNC_RESULT_COMPLETED: Arc<dyn Meter> = register_meter("sync_auto_sequential_sync_result_completed");
-    pub static ref SEQUENTIAL_SYNC_RESULT_COMPLETED: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_sequential_sync_result_completed");
     pub static ref SEQUENTIAL_SYNC_RESULT_FAILED: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_sequential_sync_result_failed");
     pub static ref SEQUENTIAL_SYNC_RESULT_TIMEOUT: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_sequential_sync_result_timeout");
 
@@ -19,8 +18,7 @@ lazy_static::lazy_static! {
     pub static ref RANDOM_STATE_TXS_READY: Arc<dyn Histogram> = Sample::ExpDecay(0.015).register("sync_auto_random_state_txs_ready", 1024);
     pub static ref RANDOM_STATE_TXS_PENDING: Arc<dyn Histogram> = Sample::ExpDecay(0.015).register("sync_auto_random_state_txs_pending", 1024);
 
-    pub static ref RANDOM_SYNC_RESULT_TOTAL: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_random_sync_result_total");
+    pub static ref RANDOM_SYNC_RESULT_COMPLETED: Arc<dyn Meter> = register_meter("sync_auto_random_sync_result_completed");
-    pub static ref RANDOM_SYNC_RESULT_COMPLETED: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_random_sync_result_completed");
     pub static ref RANDOM_SYNC_RESULT_FAILED: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_random_sync_result_failed");
     pub static ref RANDOM_SYNC_RESULT_TIMEOUT: Arc<dyn Counter<usize>> = CounterUsize::register("sync_auto_random_sync_result_timeout");
 }

node/sync/src/controllers/metrics.rs

@@ -0,0 +1,10 @@
+use std::sync::Arc;
+
+use metrics::{register_timer, Counter, CounterUsize, Histogram, Sample, Timer};
+
+lazy_static::lazy_static! {
+    pub static ref SERIAL_SYNC_FILE_COMPLETED: Arc<dyn Timer> = register_timer("sync_controllers_serial_sync_file_completed");
+    pub static ref SERIAL_SYNC_CHUNKS_COMPLETED: Arc<dyn Timer> = register_timer("sync_controllers_serial_sync_chunks_completed");
+    pub static ref SERIAL_SYNC_SEGMENT_LATENCY: Arc<dyn Histogram> = Sample::ExpDecay(0.015).register("sync_controllers_serial_sync_segment_latency", 1024);
+    pub static ref SERIAL_SYNC_UNEXPECTED_ERRORS: Arc<dyn Counter<usize>> = CounterUsize::register("sync_controllers_serial_sync_unexpected_errors");
+}

node/sync/src/controllers/mod.rs

@@ -1,3 +1,4 @@
+mod metrics;
 mod peers;
 mod serial;
 

node/sync/src/controllers/serial.rs

@@ -1,6 +1,6 @@
 use crate::context::SyncNetworkContext;
 use crate::controllers::peers::{PeerState, SyncPeers};
-use crate::controllers::{FileSyncGoal, FileSyncInfo};
+use crate::controllers::{metrics, FileSyncGoal, FileSyncInfo};
 use crate::{Config, InstantWrapper};
 use file_location_cache::FileLocationCache;
 use libp2p::swarm::DialError;
@@ -311,15 +311,15 @@ impl SerialSyncController {
                 .peers
                 .add_new_peer_with_config(peer_id, addr.clone(), shard_config)
             {
-                info!(%self.tx_seq, %peer_id, %addr, "Found new peer");
+                debug!(%self.tx_seq, %peer_id, %addr, "Found new peer");
                 true
             } else {
                 // e.g. multiple `AnnounceFile` messages propagated
-                debug!(%self.tx_seq, %peer_id, %addr, "Found an existing peer");
+                trace!(%self.tx_seq, %peer_id, %addr, "Found an existing peer");
                 false
             }
         } else {
-            debug!(%self.tx_seq, %peer_id, %addr, "Found peer without shard config");
+            info!(%self.tx_seq, %peer_id, %addr, "Found peer without shard config");
             false
         }
     }
@@ -406,7 +406,6 @@ impl SerialSyncController {
     }
 
     pub async fn on_response(&mut self, from_peer_id: PeerId, response: ChunkArrayWithProof) {
-        debug!(%self.tx_seq, %from_peer_id, "Received RPC response");
         if self.handle_on_response_mismatch(from_peer_id) {
             return;
         }
@@ -429,6 +428,7 @@ impl SerialSyncController {
         let data_len = response.chunks.data.len();
         if data_len == 0 || data_len % CHUNK_SIZE > 0 {
             warn!(%from_peer_id, %self.tx_seq, %data_len, "Invalid chunk response data length");
+            metrics::SERIAL_SYNC_UNEXPECTED_ERRORS.inc(1);
             self.ban_peer(from_peer_id, "Invalid chunk response data length");
             self.state = SyncState::Idle;
             return;
@@ -466,6 +466,7 @@ impl SerialSyncController {
             }
             Err(err) => {
                 warn!(%err, %self.tx_seq, "Failed to validate chunks response");
+                metrics::SERIAL_SYNC_UNEXPECTED_ERRORS.inc(1);
                 self.ban_peer(from_peer_id, "Chunk array validation failed");
                 self.state = SyncState::Idle;
                 return;
@@ -474,6 +475,8 @@ impl SerialSyncController {
 
         self.failures = 0;
 
+        metrics::SERIAL_SYNC_SEGMENT_LATENCY.update_since(since.0);
+
         let shard_config = self.store.get_store().flow().get_shard_config();
         let next_chunk = shard_config.next_segment_index(
             (from_chunk / PORA_CHUNK_SIZE as u64) as usize,
@@ -488,6 +491,7 @@ impl SerialSyncController {
             Ok(true) => self.next_chunk = next_chunk as u64,
             Ok(false) => {
                 warn!(%self.tx_seq, ?self.tx_id, "Transaction reverted while storing chunks");
+                metrics::SERIAL_SYNC_UNEXPECTED_ERRORS.inc(1);
                 self.state = SyncState::Failed {
                     reason: FailureReason::TxReverted(self.tx_id),
                 };
@@ -495,6 +499,7 @@ impl SerialSyncController {
             }
             Err(err) => {
                 error!(%err, %self.tx_seq, "Unexpected DB error while storing chunks");
+                metrics::SERIAL_SYNC_UNEXPECTED_ERRORS.inc(1);
                 self.state = SyncState::Failed {
                     reason: FailureReason::DBError(err.to_string()),
                 };
@@ -511,6 +516,7 @@ impl SerialSyncController {
         // completed to download chunks
         if !self.goal.is_all_chunks() {
             self.state = SyncState::Completed;
+            metrics::SERIAL_SYNC_CHUNKS_COMPLETED.update_since(self.since.0);
             return;
         }
 
@@ -523,15 +529,18 @@ impl SerialSyncController {
             Ok(true) => {
                 info!(%self.tx_seq, "Succeeded to finalize file");
                 self.state = SyncState::Completed;
+                metrics::SERIAL_SYNC_FILE_COMPLETED.update_since(self.since.0);
             }
             Ok(false) => {
                 warn!(?self.tx_id, %self.tx_seq, "Transaction reverted during finalize_tx");
+                metrics::SERIAL_SYNC_UNEXPECTED_ERRORS.inc(1);
                 self.state = SyncState::Failed {
                     reason: FailureReason::TxReverted(self.tx_id),
                 };
             }
             Err(err) => {
                 error!(%err, %self.tx_seq, "Unexpected error during finalize_tx");
+                metrics::SERIAL_SYNC_UNEXPECTED_ERRORS.inc(1);
                 self.state = SyncState::Failed {
                     reason: FailureReason::DBError(err.to_string()),
                 };

node/sync/src/lib.rs

@@ -52,7 +52,9 @@ pub struct Config {
     pub max_sequential_workers: usize,
     pub max_random_workers: usize,
     #[serde(deserialize_with = "deserialize_duration")]
-    pub find_peer_timeout: Duration,
+    pub sequential_find_peer_timeout: Duration,
+    #[serde(deserialize_with = "deserialize_duration")]
+    pub random_find_peer_timeout: Duration,
 }
 
 impl Default for Config {
@@ -61,26 +63,27 @@ impl Default for Config {
             // sync service config
             heartbeat_interval: Duration::from_secs(5),
             auto_sync_enabled: false,
-            max_sync_files: 16,
+            max_sync_files: 32,
             sync_file_by_rpc_enabled: true,
             sync_file_on_announcement_enabled: false,
 
             // serial sync config
             max_chunks_to_request: 2 * 1024,
             max_request_failures: 5,
-            peer_connect_timeout: Duration::from_secs(5),
+            peer_connect_timeout: Duration::from_secs(15),
-            peer_disconnect_timeout: Duration::from_secs(5),
+            peer_disconnect_timeout: Duration::from_secs(15),
-            peer_find_timeout: Duration::from_secs(5),
+            peer_find_timeout: Duration::from_secs(30),
-            peer_chunks_download_timeout: Duration::from_secs(5),
+            peer_chunks_download_timeout: Duration::from_secs(15),
             peer_wait_outgoing_connection_timeout: Duration::from_secs(10),
             peer_next_chunks_request_wait_timeout: Duration::from_secs(3),
 
             // auto sync config
             auto_sync_idle_interval: Duration::from_secs(3),
             auto_sync_error_interval: Duration::from_secs(10),
-            max_sequential_workers: 8,
+            max_sequential_workers: 24,
-            max_random_workers: 4,
+            max_random_workers: 8,
-            find_peer_timeout: Duration::from_secs(10),
+            sequential_find_peer_timeout: Duration::from_secs(60),
+            random_find_peer_timeout: Duration::from_secs(500),
         }
     }
 }

node/sync/src/service.rs

@@ -157,7 +157,7 @@ impl SyncService {
         event_recv: broadcast::Receiver<LogSyncEvent>,
         catch_up_end_recv: oneshot::Receiver<()>,
     ) -> Result<SyncSender> {
-        let (sync_send, sync_recv) = channel::Channel::unbounded();
+        let (sync_send, sync_recv) = channel::Channel::unbounded("sync");
         let store = Store::new(store, executor.clone());
 
         // init auto sync
@@ -912,7 +912,7 @@ mod tests {
             create_file_location_cache(init_peer_id, vec![txs[0].id()]);
 
         let (network_send, mut network_recv) = mpsc::unbounded_channel::<NetworkMessage>();
-        let (_, sync_recv) = channel::Channel::unbounded();
+        let (_, sync_recv) = channel::Channel::unbounded("test");
 
         let mut sync = SyncService {
             config: Config::default(),
@@ -941,7 +941,7 @@ mod tests {
             create_file_location_cache(init_peer_id, vec![txs[0].id()]);
 
         let (network_send, mut network_recv) = mpsc::unbounded_channel::<NetworkMessage>();
-        let (_, sync_recv) = channel::Channel::unbounded();
+        let (_, sync_recv) = channel::Channel::unbounded("test");
 
         let mut sync = SyncService {
             config: Config::default(),

run/config-testnet-standard.toml

@@ -228,11 +228,7 @@ reward_contract_address = "0x0496D0817BD8519e0de4894Dc379D35c35275609"
 auto_sync_enabled = true
 
 #  Maximum number of files in sync from other peers simultaneously.
-max_sync_files = 32
+# max_sync_files = 32
-
-# Timeout to terminate a file sync when automatically sync from other peers.
-# If timeout, terminated file sync will be triggered later.
-# find_peer_timeout = "10s"
 
 # Enable to start a file sync via RPC (e.g. `admin_startSyncFile`).
 # sync_file_by_rpc_enabled = true
@@ -241,10 +237,10 @@ max_sync_files = 32
 # sync_file_on_announcement_enabled = false
 
 # Maximum threads to sync files in sequence.
-max_sequential_workers = 24
+# max_sequential_workers = 24
 
 # Maximum threads to sync files randomly.
-# max_random_workers = 4
+# max_random_workers = 8
 
 #######################################################################
 ###                File Location Cache Options                      ###

run/config-testnet-turbo.toml

@@ -228,11 +228,7 @@ reward_contract_address = "0x51998C4d486F406a788B766d93510980ae1f9360"
 auto_sync_enabled = true
 
 #  Maximum number of files in sync from other peers simultaneously.
-max_sync_files = 32
+# max_sync_files = 32
-
-# Timeout to terminate a file sync when automatically sync from other peers.
-# If timeout, terminated file sync will be triggered later.
-# find_peer_timeout = "10s"
 
 # Enable to start a file sync via RPC (e.g. `admin_startSyncFile`).
 # sync_file_by_rpc_enabled = true
@@ -241,10 +237,10 @@ max_sync_files = 32
 # sync_file_on_announcement_enabled = false
 
 # Maximum threads to sync files in sequence.
-max_sequential_workers = 24
+# max_sequential_workers = 24
 
 # Maximum threads to sync files randomly.
-# max_random_workers = 4
+# max_random_workers = 8
 
 #######################################################################
 ###                File Location Cache Options                      ###

run/config.toml

@@ -227,11 +227,7 @@
 # auto_sync_enabled = false
 
 #  Maximum number of files in sync from other peers simultaneously.
-# max_sync_files = 16
+# max_sync_files = 32
-
-# Timeout to terminate a file sync when automatically sync from other peers.
-# If timeout, terminated file sync will be triggered later.
-# find_peer_timeout = "10s"
 
 # Enable to start a file sync via RPC (e.g. `admin_startSyncFile`).
 # sync_file_by_rpc_enabled = true
@@ -240,10 +236,10 @@
 # sync_file_on_announcement_enabled = false
 
 # Maximum threads to sync files in sequence.
-# max_sequential_workers = 8
+# max_sequential_workers = 24
 
 # Maximum threads to sync files randomly.
-# max_random_workers = 4
+# max_random_workers = 8
 
 #######################################################################
 ###                File Location Cache Options                      ###