Cherry-pick #21258 to 7.x: o365input: Restart after fatal error (#21383)

Update the o365input to restart the input after a fatal error is
encountered, for example an authentication token refresh error or a parsing
error.

This enables the input to be more resilient against transient errors.

Before this patch, the input would index an error document and terminate.
Now it will index an error and restart after a fixed timeout of 5 minutes.

(cherry picked from commit 8716d98)

CHANGELOG.next.asciidoc

@@ -193,6 +193,7 @@ field. You can revert this change by configuring tags for the module and omittin
 - [Autodiscover] Handle input-not-finished errors in config reload. {pull}20915[20915]
 - Explicitly detect missing variables in autodiscover configuration, log them at the debug level. {issue}20568[20568] {pull}20898[20898]
 - Fix `libbeat.output.write.bytes` and `libbeat.output.read.bytes` metrics of the Elasticsearch output. {issue}20752[20752] {pull}21197[21197]
+- The `o365input` and `o365` module now recover from an authentication problem or other fatal errors, instead of terminating. {pull}21258[21258]
 
 *Auditbeat*
 

x-pack/filebeat/input/o365audit/input.go

@@ -26,6 +26,9 @@ import (
 const (
 	pluginName   = "o365audit"
 	fieldsPrefix = pluginName
+
+	// How long to retry when a fatal error is encountered in the input.
+	failureRetryInterval = time.Minute * 5
 )
 
 type o365input struct {
@@ -107,6 +110,34 @@ func (inp *o365input) Run(
 	src cursor.Source,
 	cursor cursor.Cursor,
 	publisher cursor.Publisher,
+) error {
+	for ctx.Cancelation.Err() == nil {
+		err := inp.runOnce(ctx, src, cursor, publisher)
+		if err == nil {
+			break
+		}
+		if ctx.Cancelation.Err() != err && err != context.Canceled {
+			msg := common.MapStr{}
+			msg.Put("error.message", err.Error())
+			msg.Put("event.kind", "pipeline_error")
+			event := beat.Event{
+				Timestamp: time.Now(),
+				Fields:    msg,
+			}
+			publisher.Publish(event, nil)
+			ctx.Logger.Errorf("Input failed: %v", err)
+			ctx.Logger.Infof("Restarting in %v", failureRetryInterval)
+			time.Sleep(failureRetryInterval)
+		}
+	}
+	return nil
+}
+
+func (inp *o365input) runOnce(
+	ctx v2.Context,
+	src cursor.Source,
+	cursor cursor.Cursor,
+	publisher cursor.Publisher,
 ) error {
 	stream := src.(*stream)
 	tenantID, contentType := stream.tenantID, stream.contentType
@@ -156,18 +187,7 @@ func (inp *o365input) Run(
 	}
 
 	log.Infow("Start fetching events", "cursor", start)
-	err = poller.Run(action)
-	if err != nil && ctx.Cancelation.Err() != err && err != context.Canceled {
-		msg := common.MapStr{}
-		msg.Put("error.message", err.Error())
-		msg.Put("event.kind", "pipeline_error")
-		event := beat.Event{
-			Timestamp: time.Now(),
-			Fields:    msg,
-		}
-		publisher.Publish(event, nil)
-	}
-	return err
+	return poller.Run(action)
 }
 
 func initCheckpoint(log *logp.Logger, c cursor.Cursor, maxRetention time.Duration) checkpoint {