diff --git a/sentry_streams/src/batch_step.rs b/sentry_streams/src/batch_step.rs
index c2a4d3a8..e84a003c 100644
--- a/sentry_streams/src/batch_step.rs
+++ b/sentry_streams/src/batch_step.rs
@@ -21,6 +21,9 @@ use sentry_arroyo::utils::timing::Deadline;
 use std::collections::{BTreeMap, VecDeque};
 use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 
+const METRIC_BATCH_SIZE: &str = "streams.pipeline.batch.size";
+const METRIC_BATCH_TIME_MS: &str = "streams.pipeline.batch.time_ms";
+
 fn first_element_schema(py: Python<'_>, first: &PyStreamingMessage) -> Option<String> {
     match first {
         PyStreamingMessage::PyAnyMessage { content } => content.bind(py).borrow().schema.clone(),
@@ -54,6 +57,8 @@ pub(crate) struct Batch {
     max_batch_size: Option<usize>,
     /// Set when the window is time-bounded; elapsed means flush by time.
     batch_deadline: Option<Deadline>,
+    /// Wall time when the first element opened this batch window.
+    created_at: Instant,
     elements: Vec<PyStreamingMessage>,
     batch_offsets: BTreeMap<Partition, u64>,
 }
@@ -82,6 +87,7 @@ impl Batch {
             route,
             max_batch_size,
             batch_deadline,
+            created_at: Instant::now(),
             elements: vec![first],
             batch_offsets,
         }
@@ -258,9 +264,14 @@ impl BatchStep {
         // We create a synthetic watermark to avoid waiting for the next batch to complete before
         // allowing the consumer to commit.
         let committable_for_synthetic = b.current_offsets_snapshot();
+        let batch_elements = b.len() as f64;
+        let batch_open_ms = b.created_at.elapsed().as_millis() as f64;
         let flush_start = Instant::now();
         let batch_msg = b.flush()?;
         get_stats().step_timing(&self.step_name, flush_start.elapsed().as_secs_f64());
+        let step_labels = vec![("step".to_string(), self.step_name.clone())];
+        metrics::histogram!(METRIC_BATCH_SIZE, &step_labels).record(batch_elements);
+        metrics::histogram!(METRIC_BATCH_TIME_MS, &step_labels).record(batch_open_ms);
         self.batch = None;
         let wm_after_batch: Vec<_> = std::mem::take(&mut self.watermark_buffer);
 
diff --git a/sentry_streams/src/gcs_writer.rs b/sentry_streams/src/gcs_writer.rs
index e6782797..94e00f8f 100644
--- a/sentry_streams/src/gcs_writer.rs
+++ b/sentry_streams/src/gcs_writer.rs
@@ -21,6 +21,8 @@ use gcp_auth::{provider, TokenProvider};
 use std::sync::Arc;
 use tokio::sync::OnceCell;
 
+const METRIC_SINK_GCS_WRITER_BYTES: &str = "streams.pipeline.sink.gcs_writer.bytes";
+
 pub struct GCSWriter {
     client: Client,
     bucket: String,
@@ -102,6 +104,7 @@ impl TaskRunner<RoutedValue, RoutedValue, anyhow::Error> for GCSWriter {
         let pybytes_ms = pybytes_start.elapsed().as_millis();
 
         let bytes_len = bytes.len();
+        let route_source = self.route.source.clone();
 
         let auth_provider_cell = self.auth_provider.clone();
 
@@ -176,6 +179,9 @@ impl TaskRunner<RoutedValue, RoutedValue, anyhow::Error> for GCSWriter {
                     token_ms,
                     request_ms
                 );
+                let gcs_labels = vec![("source".to_string(), route_source.clone())];
+                metrics::histogram!(METRIC_SINK_GCS_WRITER_BYTES, &gcs_labels)
+                    .record(bytes_len as f64);
                 Ok(message)
             }
         })