What changed

• A replace directive switches Prometheus to a Grafana fork at commit 8cf60eef5463.
• The fork is based on upstream v0.309.1 and carries a targeted fix: correct sent_batch_duration_seconds to measure time starting when the HTTP request is actually sent (not earlier).
• Context: “Fix sent_batch_duration_seconds measuring before the request was sent.” (Upstream PR: remote write: fix sent_batch_duration_seconds measuring before the request is sent prometheus/prometheus#18214, fork branch note: https://github.com/grafana/prometheus/tree/fix-sent-batch-duration-v0.309.1)

Scope and impact

• The change is localized to Prometheus remote-write internals (storage/remote). It does not modify exported APIs that consumers (such as the OpenTelemetry Collector’s prometheusreceiver) import.
• No breaking changes to package structure, public types, or function signatures were introduced between upstream v0.309.1 and this forked commit.
• Behavioral change is limited to the semantics of the sent_batch_duration_seconds metric:
  • Previously included time before the request was sent.
  • Now only measures the duration after the request is sent.
  • If you consume/alert on this metric, expect values to decrease; otherwise, no code changes are required.

Evidence

• Upstream issue/PR describing the problem and the fix intent:
  • remote write: fix sent_batch_duration_seconds measuring before the request is sent prometheus/prometheus#18214
• Fork branch with the single-purpose fix:
  • https://github.com/grafana/prometheus/tree/fix-sent-batch-duration-v0.309.1
• The change is a timing/observation point adjustment in remote-write’s queue/HTTP path (storage/remote); API surface is unchanged.

Code changes required in this repository

• None. This is a compatible, forked patch-level change with no public API changes.
• If you have dashboards/alerts tied to sent_batch_duration_seconds, you may want to recalibrate thresholds due to the corrected measurement scope.

Illustrative patch from the fork (conceptual)

• The fix delays starting the histogram timer until right before the HTTP request is sent:

  - start := time.Now()
  - defer func() { sentBatchDuration.Observe(time.Since(start).Seconds()) }()
  + var start time.Time
  + defer func() {
  +   if !start.IsZero() {
  +     sentBatchDuration.Observe(time.Since(start).Seconds())
  +   }
  + }()
    ...
  + // right before sending the HTTP request
  + start = time.Now()
    resp, err := httpClient.Do(req)

• No function signatures or imports that downstream users rely on are altered.

Recommended validation

• Run existing tests for components depending on github.com/prometheus/prometheus (e.g., prometheusreceiver).
• If you monitor sent_batch_duration_seconds, verify expected drop in observed duration due to corrected timing.