feat(o11y): Inter-process tracing

CHANGELOG.md

@@ -67,6 +67,8 @@
   information: [#7711](https://github.com/near/nearcore/pull/7711).
 * Change exporter of tracing information from `opentelemetry-jaeger` to
   `opentelemetry-otlp`: [#7563](https://github.com/near/nearcore/pull/7563).
+* Tracing of requests across processes:
+  [#8004](https://github.com/near/nearcore/pull/8004).
 
 ## 1.29.0 [2022-08-15]
 

chain/network/src/network_protocol/mod.rs

@@ -19,10 +19,14 @@ mod _proto {
 
 pub use _proto::network as proto;
 
+use crate::network_protocol::proto_conv::trace_context::{
+    extract_span_context, inject_trace_context,
+};
 use crate::time;
 use borsh::{BorshDeserialize as _, BorshSerialize as _};
 use near_crypto::PublicKey;
 use near_crypto::Signature;
+use near_o11y::OpenTelemetrySpanExt;
 use near_primitives::block::{Approval, Block, BlockHeader, GenesisId};
 use near_primitives::challenge::Challenge;
 use near_primitives::hash::CryptoHash;
@@ -40,7 +44,9 @@ use near_primitives::views::FinalExecutionOutcomeView;
 use protobuf::Message as _;
 use std::collections::HashSet;
 use std::fmt;
+use std::fmt::Debug;
 use std::sync::Arc;
+use tracing::Span;
 
 #[derive(PartialEq, Eq, Clone, Debug, Hash)]
 pub struct PeerAddr {
@@ -281,26 +287,38 @@ pub enum ParsePeerMessageError {
 }
 
 impl PeerMessage {
+    /// Serializes a message in the given encoding.
+    /// If the encoding is `Proto`, then also attaches current Span's context to the message.
     pub(crate) fn serialize(&self, enc: Encoding) -> Vec<u8> {
         match enc {
             Encoding::Borsh => borsh_::PeerMessage::from(self).try_to_vec().unwrap(),
-            Encoding::Proto => proto::PeerMessage::from(self).write_to_bytes().unwrap(),
+            Encoding::Proto => {
+                let mut msg = proto::PeerMessage::from(self);
+                let cx = Span::current().context();
+                msg.trace_context = inject_trace_context(&cx);
+                msg.write_to_bytes().unwrap()
+            }
         }
     }
 
     pub(crate) fn deserialize(
         enc: Encoding,
         data: &[u8],
     ) -> Result<PeerMessage, ParsePeerMessageError> {
+        let span = tracing::trace_span!(target: "network", "deserialize").entered();
         Ok(match enc {
             Encoding::Borsh => (&borsh_::PeerMessage::try_from_slice(data)
                 .map_err(ParsePeerMessageError::BorshDecode)?)
                 .try_into()
                 .map_err(ParsePeerMessageError::BorshConv)?,
-            Encoding::Proto => (&proto::PeerMessage::parse_from_bytes(data)
-                .map_err(ParsePeerMessageError::ProtoDecode)?)
-                .try_into()
-                .map_err(ParsePeerMessageError::ProtoConv)?,
+            Encoding::Proto => {
+                let proto_msg: proto::PeerMessage = proto::PeerMessage::parse_from_bytes(data)
+                    .map_err(ParsePeerMessageError::ProtoDecode)?;
+                if let Ok(extracted_span_context) = extract_span_context(&proto_msg.trace_context) {
+                    span.clone().or_current().add_link(extracted_span_context);
+                }
+                (&proto_msg).try_into().map_err(|err| ParsePeerMessageError::ProtoConv(err))?
+            }
         })
     }
 

chain/network/src/network_protocol/network.proto

@@ -317,6 +317,22 @@ message RoutingSyncV2 {
   bytes borsh = 1;
 }
 
+// Inter-process tracing information.
+message TraceContext {
+  enum SamplingPriority {
+    UNKNOWN = 0;
+    AutoReject = 1;
+    UserReject = 2;
+    AutoKeep = 3;
+    UserKeep = 4;
+  }
+  // 16 bytes representing TraceId: https://docs.rs/opentelemetry/latest/opentelemetry/trace/struct.TraceId.html
+  bytes trace_id = 1;
+  // 8 bytes representing SpanId: https://docs.rs/opentelemetry/latest/opentelemetry/trace/struct.SpanId.html
+  bytes span_id = 2;
+  SamplingPriority sampling_priority = 3;
+}
+
 // PeerMessage is a wrapper of all message types exchanged between NEAR nodes.
 // The wire format of a single message M consists of len(M)+4 bytes:
 // <len(M)> : 4 bytes : little endian uint32
@@ -329,6 +345,9 @@ message PeerMessage {
   // Deprecated fields.
   reserved 9,20,21,22,23,24;
 
+  // Inter-process tracing information.
+  TraceContext trace_context = 26;
+
   oneof message_type {
     Handshake handshake = 4;
     HandshakeFailure handshake_failure = 5;

chain/network/src/network_protocol/proto_conv/mod.rs

@@ -4,6 +4,7 @@ mod handshake;
 mod net;
 mod peer_message;
 mod time;
+pub mod trace_context;
 /// Contains protobuf <-> network_protocol conversions.
 mod util;
 

chain/network/src/network_protocol/proto_conv/trace_context.rs

@@ -0,0 +1,78 @@
+use crate::network_protocol::proto::trace_context::SamplingPriority;
+use crate::network_protocol::proto::TraceContext;
+use opentelemetry::trace::{SpanContext, SpanId, TraceContextExt, TraceFlags, TraceId, TraceState};
+use opentelemetry::Context;
+use protobuf::{EnumOrUnknown, MessageField};
+
+/// Lowest available value.
+/// 0x01 is reserved for `SAMPLED`: https://docs.rs/opentelemetry/latest/opentelemetry/trace/struct.TraceFlags.html#associatedconstant.SAMPLED
+const TRACE_FLAG_DEFERRED: TraceFlags = TraceFlags::new(0x02);
+
+#[derive(Debug)]
+pub(crate) enum ExtractError {
+    // Malformed or invalid TraceId.
+    TraceId,
+    // Malformed or invalid SpanId.
+    SpanId,
+    // Missing trace_id or span_id.
+    Empty,
+}
+
+/// Extracts a `SpanContext` from a potentially empty `TraceContext`.
+///
+pub(crate) fn extract_span_context(
+    trace_context: &MessageField<TraceContext>,
+) -> Result<SpanContext, ExtractError> {
+    if trace_context.is_some() {
+        let trace_id = extract_trace_id(&trace_context.trace_id)?;
+        // If we have a trace_id but can't get the parent span, we default it to invalid instead of completely erroring
+        // out so that the rest of the spans aren't completely lost.
+        let span_id = extract_span_id(&trace_context.span_id).unwrap_or(SpanId::INVALID);
+        let sampled = match trace_context.sampling_priority.enum_value() {
+            Ok(SamplingPriority::UserReject) | Ok(SamplingPriority::AutoReject) => {
+                TraceFlags::default()
+            }
+            Ok(SamplingPriority::UserKeep) | Ok(SamplingPriority::AutoKeep) => TraceFlags::SAMPLED,
+            // Treat the sampling as DEFERRED instead of erring on extracting the span context
+            Ok(SamplingPriority::UNKNOWN) | Err(_) => TRACE_FLAG_DEFERRED,
+        };
+        let trace_state = TraceState::default();
+        Ok(SpanContext::new(trace_id, span_id, sampled, true, trace_state))
+    } else {
+        Err(ExtractError::Empty)
+    }
+}
+
+/// Populates `TraceContext` representing the current span.
+/// Returns `None` if no current span is available.
+pub(crate) fn inject_trace_context(cx: &Context) -> MessageField<TraceContext> {
+    let span = cx.span();
+    let span_context = span.span_context();
+    if span_context.is_valid() {
+        let mut trace_context = TraceContext::new();
+
+        // Uses `u128::to_be_bytes()` internally.
+        trace_context.trace_id = span_context.trace_id().to_bytes().to_vec();
+        // Uses `u64::to_be_bytes()` internally.
+        trace_context.span_id = span_context.span_id().to_bytes().to_vec();
+
+        if span_context.trace_flags() & TRACE_FLAG_DEFERRED != TRACE_FLAG_DEFERRED {
+            let sampling_priority = if span_context.is_sampled() {
+                SamplingPriority::AutoKeep
+            } else {
+                SamplingPriority::AutoReject
+            };
+            trace_context.sampling_priority = EnumOrUnknown::new(sampling_priority);
+        }
+        MessageField::some(trace_context)
+    } else {
+        MessageField::none()
+    }
+}
+fn extract_trace_id(trace_id: &[u8]) -> Result<TraceId, ExtractError> {
+    Ok(TraceId::from_bytes(trace_id.try_into().map_err(|_| ExtractError::TraceId)?))
+}
+
+fn extract_span_id(span_id: &[u8]) -> Result<SpanId, ExtractError> {
+    Ok(SpanId::from_bytes(span_id.try_into().map_err(|_| ExtractError::SpanId)?))
+}

chain/network/src/peer/peer_actor.rs

@@ -261,7 +261,6 @@ impl PeerActor {
     }
 
     fn parse_message(&mut self, msg: &[u8]) -> Result<PeerMessage, ParsePeerMessageError> {
-        let _span = tracing::trace_span!(target: "network", "parse_message").entered();
         if let Some(e) = self.encoding() {
             return PeerMessage::deserialize(e, msg);
         }
@@ -1206,7 +1205,7 @@ impl actix::Handler<stream::Frame> for PeerActor {
     type Result = ();
     #[perf]
     fn handle(&mut self, stream::Frame(msg): stream::Frame, ctx: &mut Self::Context) {
-        let _span = tracing::trace_span!(
+        let _span = tracing::debug_span!(
             target: "network",
             "handle",
             handler = "bytes",

chain/network/src/peer_manager/peer_manager_actor.rs

@@ -24,7 +24,10 @@ use crate::types::{
 use actix::fut::future::wrap_future;
 use actix::{Actor, AsyncContext, Context, Handler, Running};
 use anyhow::Context as _;
-use near_o11y::{handler_trace_span, OpenTelemetrySpanExt, WithSpanContext, WithSpanContextExt};
+use near_o11y::{
+    handler_debug_span, handler_trace_span, OpenTelemetrySpanExt, WithSpanContext,
+    WithSpanContextExt,
+};
 use near_performance_metrics_macros::perf;
 use near_primitives::block::GenesisId;
 use near_primitives::network::{AnnounceAccount, PeerId};
@@ -1068,7 +1071,7 @@ impl Handler<WithSpanContext<PeerManagerMessageRequest>> for PeerManagerActor {
         ctx: &mut Self::Context,
     ) -> Self::Result {
         let msg_type: &str = (&msg.msg).into();
-        let (_span, msg) = handler_trace_span!(target: "network", msg, msg_type);
+        let (_span, msg) = handler_debug_span!(target: "network", msg, msg_type);
         let _timer =
             metrics::PEER_MANAGER_MESSAGES_TIME.with_label_values(&[(&msg).into()]).start_timer();
         self.handle_peer_manager_message(msg, ctx)

core/o11y/src/lib.rs

@@ -258,13 +258,18 @@ where
     let (filter, handle) = reload::Layer::<LevelFilter, S>::new(filter);
 
     let mut resource = vec![
-        KeyValue::new(SERVICE_NAME, "neard"),
         KeyValue::new("chain_id", chain_id),
         KeyValue::new("node_id", node_public_key.to_string()),
     ];
-    if let Some(account_id) = account_id {
+    // Prefer account name as the node name.
+    // Fallback to a node public key if a validator key is unavailable.
+    let service_name = if let Some(account_id) = account_id {
         resource.push(KeyValue::new("account_id", account_id.to_string()));
-    }
+        format!("neard:{}", account_id)
+    } else {
+        format!("neard:{}", node_public_key)
+    };
+    resource.push(KeyValue::new(SERVICE_NAME, service_name));
 
     let tracer = opentelemetry_otlp::new_pipeline()
         .tracing()