loki.source.kubernetes: restart tailers less aggressively

CHANGELOG.md

@@ -25,6 +25,9 @@ Main (unreleased)
 - In `prometheus.exporter.kafka`, the interpolation table used to compute estimated lag metrics is now pruned
   on `metadata_refresh_interval` instead of `prune_interval_seconds`. (@wildum)
 
+- Don't restart tailers in `loki.source.kubernetes` component by above-average
+  time deltas if K8s version is >= 1.29.1 (@hainenber)
+
 ### Bugfixes
 
 - Fixed issue with defaults for Beyla component not being applied correctly. (marctc)  

internal/component/loki/source/kubernetes/kubetail/tailer.go

@@ -9,6 +9,7 @@ import (
 	"strings"
 	"time"
 
+	"github.com/blang/semver/v4"
 	"github.com/go-kit/log"
 	"github.com/grafana/alloy/internal/alloy/logging/level"
 	"github.com/grafana/alloy/internal/component/common/loki"
@@ -170,6 +171,15 @@ func (t *tailer) tail(ctx context.Context, handler loki.EntryHandler) error {
 		return err
 	}
 
+	k8sServerVersion, err := t.opts.Client.Discovery().ServerVersion()
+	if err != nil {
+		return err
+	}
+	k8sComparableServerVersion, err := semver.ParseTolerant(k8sServerVersion.GitVersion)
+	if err != nil {
+		return err
+	}
+
 	// Create a new rolling average calculator to determine the average delta
 	// time between log entries.
 	//
@@ -180,41 +190,48 @@ func (t *tailer) tail(ctx context.Context, handler loki.EntryHandler) error {
 	// The computed average will never be less than the minimum of 2s.
 	calc := newRollingAverageCalculator(10000, 100, 2*time.Second, maxTailerLifetime)
 
-	go func() {
-		rolledFileTicker := time.NewTicker(1 * time.Second)
-		defer func() {
-			rolledFileTicker.Stop()
-			_ = stream.Close()
-		}()
-		for {
-			select {
-			case <-ctx.Done():
-				return
-			case <-rolledFileTicker.C:
-				// Versions of Kubernetes which do not contain
-				// kubernetes/kubernetes#115702 will fail to detect rolled log files
-				// and stop sending logs to us.
-				//
-				// To work around this, we use a rolling average to determine how
-				// frequent we usually expect to see entries. If 3x the normal delta has
-				// elapsed, we'll restart the tailer.
-				//
-				// False positives here are acceptable, but false negatives mean that
-				// we'll have a larger spike of missing logs until we detect a rolled
-				// file.
-				avg := calc.GetAverage()
-				last := calc.GetLast()
-				if last.IsZero() {
-					continue
-				}
-				s := time.Since(last)
-				if s > avg*3 {
-					level.Info(t.log).Log("msg", "have not seen a log line in 3x average time between lines, closing and re-opening tailer", "rolling_average", avg, "time_since_last", s)
+	// Versions of Kubernetes which do not contain
+	// kubernetes/kubernetes#115702 (<= v1.29.1) will fail to detect rotated log files
+	// and stop sending logs to us.
+	//
+	// To work around this, we use a rolling average to determine how
+	// frequent we usually expect to see entries. If 3x the normal delta has
+	// elapsed, we'll restart the tailer.
+	//
+	// False positives here are acceptable, but false negatives mean that
+	// we'll have a larger spike of missing logs until we detect a rolled
+	// file.
+	if k8sComparableServerVersion.LT(semver.Version{Major: 1, Minor: 29, Patch: 0}) {
+		go func() {
+			rolledFileTicker := time.NewTicker(1 * time.Second)
+			defer func() {
+				rolledFileTicker.Stop()
+				_ = stream.Close()
+			}()
+			for {
+				select {
+				case <-ctx.Done():
 					return
+				case <-rolledFileTicker.C:
+					avg := calc.GetAverage()
+					last := calc.GetLast()
+					if last.IsZero() {
+						continue
+					}
+					s := time.Since(last)
+					if s > avg*3 {
+						level.Debug(t.log).Log("msg", "have not seen a log line in 3x average time between lines, closing and re-opening tailer", "rolling_average", avg, "time_since_last", s)
+						return
+					}
 				}
 			}
-		}
-	}()
+		}()
+	} else {
+		go func() {
+			<-ctx.Done()
+			_ = stream.Close()
+		}()
+	}
 
 	level.Info(t.log).Log("msg", "opened log stream", "start time", lastReadTime)