fix(rustproxy-http,rustproxy-metrics): reduce per-frame metrics overhead by batching body byte accounting

This commit is contained in:
2026-03-17 12:22:51 +00:00
parent 8f6bb30367
commit 92d7113c6c
4 changed files with 92 additions and 49 deletions

View File

@@ -9,10 +9,17 @@ use bytes::Bytes;
use http_body::Frame;
use rustproxy_metrics::MetricsCollector;
/// Flush accumulated bytes to the metrics collector every 64 KB.
/// This reduces per-frame DashMap shard-locked reads from ~15 to ~1 per 4 frames
/// (assuming typical 16 KB upload frames). The 1 Hz throughput sampler still sees
/// data within one sampling period even at low transfer rates.
const BYTE_FLUSH_THRESHOLD: u64 = 65_536;
/// Wraps any `http_body::Body` and counts data bytes passing through.
///
/// Each chunk is reported to the `MetricsCollector` immediately so that
/// the throughput tracker (sampled at 1 Hz) reflects real-time data flow.
/// Bytes are accumulated and flushed to the `MetricsCollector` every
/// [`BYTE_FLUSH_THRESHOLD`] bytes (and on Drop) so the throughput tracker
/// (sampled at 1 Hz) reflects real-time data flow without per-frame overhead.
///
/// The inner body is pinned on the heap to support `!Unpin` types like `hyper::body::Incoming`.
pub struct CountingBody<B> {
@@ -22,6 +29,8 @@ pub struct CountingBody<B> {
source_ip: Option<String>,
/// Whether we count bytes as "in" (request body) or "out" (response body).
direction: Direction,
/// Accumulated bytes not yet flushed to the metrics collector.
pending_bytes: u64,
/// Optional connection-level activity tracker. When set, poll_frame updates this
/// to keep the idle watchdog alive during active body streaming (uploads/downloads).
connection_activity: Option<Arc<AtomicU64>>,
@@ -57,6 +66,7 @@ impl<B> CountingBody<B> {
route_id,
source_ip,
direction,
pending_bytes: 0,
connection_activity: None,
activity_start: None,
active_requests: None,
@@ -81,14 +91,19 @@ impl<B> CountingBody<B> {
self
}
/// Report a chunk of bytes immediately to the metrics collector.
/// Flush accumulated bytes to the metrics collector.
#[inline]
fn report_chunk(&self, len: u64) {
fn flush_pending(&mut self) {
if self.pending_bytes == 0 {
return;
}
let bytes = self.pending_bytes;
self.pending_bytes = 0;
let route_id = self.route_id.as_deref();
let source_ip = self.source_ip.as_deref();
match self.direction {
Direction::In => self.metrics.record_bytes(len, 0, route_id, source_ip),
Direction::Out => self.metrics.record_bytes(0, len, route_id, source_ip),
Direction::In => self.metrics.record_bytes(bytes, 0, route_id, source_ip),
Direction::Out => self.metrics.record_bytes(0, bytes, route_id, source_ip),
}
}
}
@@ -113,9 +128,12 @@ where
Poll::Ready(Some(Ok(frame))) => {
if let Some(data) = frame.data_ref() {
let len = data.len() as u64;
// Report bytes immediately so the 1 Hz throughput sampler sees them
this.report_chunk(len);
// Keep the connection-level idle watchdog alive during body streaming
this.pending_bytes += len;
if this.pending_bytes >= BYTE_FLUSH_THRESHOLD {
this.flush_pending();
}
// Keep the connection-level idle watchdog alive on every frame
// (this is just one atomic store — cheap enough per-frame)
if let (Some(activity), Some(start)) = (&this.connection_activity, &this.activity_start) {
activity.store(start.elapsed().as_millis() as u64, Ordering::Relaxed);
}
@@ -123,7 +141,11 @@ where
Poll::Ready(Some(Ok(frame)))
}
Poll::Ready(Some(Err(e))) => Poll::Ready(Some(Err(e))),
Poll::Ready(None) => Poll::Ready(None),
Poll::Ready(None) => {
// End of stream — flush any remaining bytes
this.flush_pending();
Poll::Ready(None)
}
Poll::Pending => Poll::Pending,
}
}
@@ -139,6 +161,8 @@ where
impl<B> Drop for CountingBody<B> {
fn drop(&mut self) {
// Flush any remaining accumulated bytes so totals stay accurate
self.flush_pending();
// Decrement the active-request counter so the HTTP idle watchdog
// knows this response body is no longer streaming.
if let Some(ref counter) = self.active_requests {