Always fall back to string payload if we can't handle structured logs

exporter/collector/README.md

@@ -186,12 +186,17 @@ Additional configuration for the metric exporter:
 
 Addition configuration for the logging exporter:
 
-- `log.default_log_name` (optional): Defines a default name for log entries. If left unset, and a log entry does not have the `gcp.log_name` 
-attribute set, the exporter will return an error processing that entry.
-- `log.error_reporting_type` (option, default = false): If `true`, log records with a severity of `error` or higher will be converted to
-JSON payloads with the `@type` field set for [GCP Error Reporting](https://cloud.google.com/error-reporting/docs/formatting-error-messages#log-text).
-If the body is currently a string, it will be converted to a `message` field in the new JSON payload. If the body is already a map, the `@type`
-field will be added to the map. Other body types (such as byte) are undefined for this behavior.
+- `log.default_log_name` (optional): Defines a default name for log entries. If
+left unset, and a log entry does not have the `gcp.log_name` attribute set, the
+exporter will return an error processing that entry.
+- `log.error_reporting_type` (option, default = false): If `true`, log records
+with a severity of `error` or higher will be converted to JSON payloads with the
+`@type` field set for [GCP Error
+Reporting](https://cloud.google.com/error-reporting/docs/formatting-error-messages#log-text).
+If the body is currently a string, it will be converted to a `message` field in
+the new JSON payload. If the body is already a map, the `@type` field will be
+added to the map. Other body types (such as byte) are undefined for this
+behavior.
 
 Example:
 

exporter/collector/logs.go

@@ -455,66 +455,69 @@
 		}
 	}
 
-	if len(logRecord.Body().AsString()) == 0 {
-		return []*logpb.LogEntry{entry}, nil
-	}
-
+	// Handle map and bytes as JSON-structured logs if they are successfully converted.
 	switch logRecord.Body().Type() {
-	case pcommon.ValueTypeBytes:
-		s, err := toProtoStruct(logRecord.Body().Bytes().AsRaw())
-		if err != nil {
-			return nil, err
-		}
-		entry.Payload = &logpb.LogEntry_JsonPayload{JsonPayload: s}
 	case pcommon.ValueTypeMap:
 		s, err := structpb.NewStruct(logRecord.Body().Map().AsRaw())
-		if err != nil {
-			return nil, err
+		if err == nil {
+			entry.Payload = &logpb.LogEntry_JsonPayload{JsonPayload: s}
+			return []*logpb.LogEntry{entry}, nil
 		}
-		entry.Payload = &logpb.LogEntry_JsonPayload{JsonPayload: s}
-	case pcommon.ValueTypeStr:
-		// Calculate the size of the internal log entry so this overhead can be accounted
-		// for when determining the need to split based on payload size
-		// TODO(damemi): Find an appropriate estimated buffer to account for the LogSplit struct as well
-		overheadBytes := proto.Size(entry)
-		// Split log entries with a string payload into fewer entries
-		payloadString := logRecord.Body().Str()
-		splits := int(math.Ceil(float64(len([]byte(payloadString))) / float64(l.maxEntrySize-overheadBytes)))
-		if splits <= 1 {
-			entry.Payload = &logpb.LogEntry_TextPayload{TextPayload: payloadString}
+		l.obs.log.Warn(fmt.Sprintf("map body cannot be converted to a json payload, exporting as raw string: %+v", err))
+	case pcommon.ValueTypeBytes:
+		s, err := toProtoStruct(logRecord.Body().Bytes().AsRaw())
+		if err == nil {
+			entry.Payload = &logpb.LogEntry_JsonPayload{JsonPayload: s}
 			return []*logpb.LogEntry{entry}, nil
 		}
-		entries := make([]*logpb.LogEntry, splits)
-		// Start by assuming all splits will be even (this may not be the case)
-		startIndex := 0
-		endIndex := int(math.Floor((1.0 / float64(splits)) * float64(len(payloadString))))
-		for i := 0; i < splits; i++ {
-			newEntry := proto.Clone(entry).(*logpb.LogEntry)
-			currentSplit := payloadString[startIndex:endIndex]
-
-			// If the current split is larger than the entry size, iterate until it is within the max
-			// (This may happen since not all characters are exactly 1 byte)
-			for len([]byte(currentSplit)) > l.maxEntrySize {
-				endIndex--
-				currentSplit = payloadString[startIndex:endIndex]
-			}
-			newEntry.Payload = &logpb.LogEntry_TextPayload{TextPayload: currentSplit}
-			newEntry.Split = &logpb.LogSplit{
-				Uid:         fmt.Sprintf("%s-%s", logName, entry.Timestamp.AsTime().String()),
-				Index:       int32(i),
-				TotalSplits: int32(splits),
-			}
-			entries[i] = newEntry
+		l.obs.log.Debug(fmt.Sprintf("bytes body cannot be converted to a json payload, exporting as base64 string: %+v", err))
+	}
+	// For all other ValueTypes, export as a string payload.
 
-			// Update slice indices to the next chunk
-			startIndex = endIndex
-			endIndex = int(math.Floor((float64(i+2) / float64(splits)) * float64(len(payloadString))))
+	// log.Body().AsString() can be expensive, and we use it several times below this, so
+	// do it once and save that as a variable.
+	logBodyString := logRecord.Body().AsString()
+	if len(logBodyString) == 0 {
+		return []*logpb.LogEntry{entry}, nil
+	}
+
+	// Calculate the size of the internal log entry so this overhead can be accounted
+	// for when determining the need to split based on payload size
+	// TODO(damemi): Find an appropriate estimated buffer to account for the LogSplit struct as well
+	overheadBytes := proto.Size(entry)
+	// Split log entries with a string payload into fewer entries
+	splits := int(math.Ceil(float64(len([]byte(logBodyString))) / float64(l.maxEntrySize-overheadBytes)))
+	if splits <= 1 {
+		entry.Payload = &logpb.LogEntry_TextPayload{TextPayload: logBodyString}
+		return []*logpb.LogEntry{entry}, nil
+	}
+	entries := make([]*logpb.LogEntry, splits)
+	// Start by assuming all splits will be even (this may not be the case)
+	startIndex := 0
+	endIndex := int(math.Floor((1.0 / float64(splits)) * float64(len(logBodyString))))
+	for i := 0; i < splits; i++ {
+		newEntry := proto.Clone(entry).(*logpb.LogEntry)
+		currentSplit := logBodyString[startIndex:endIndex]
+
+		// If the current split is larger than the entry size, iterate until it is within the max
+		// (This may happen since not all characters are exactly 1 byte)
+		for len([]byte(currentSplit)) > l.maxEntrySize {
+			endIndex--
+			currentSplit = logBodyString[startIndex:endIndex]
 		}
-		return entries, nil
-	default:
-		return nil, fmt.Errorf("unknown log body value %v", logRecord.Body().Type().String())
+		newEntry.Payload = &logpb.LogEntry_TextPayload{TextPayload: currentSplit}
+		newEntry.Split = &logpb.LogSplit{
+			Uid:         fmt.Sprintf("%s-%s", logName, entry.Timestamp.AsTime().String()),
+			Index:       int32(i),
+			TotalSplits: int32(splits),
+		}
+		entries[i] = newEntry
+
+		// Update slice indices to the next chunk
+		startIndex = endIndex
+		endIndex = int(math.Floor((float64(i+2) / float64(splits)) * float64(len(logBodyString))))
 	}
-	return []*logpb.LogEntry{entry}, nil
+	return entries, nil
 }
 
 // JSON keys derived from:

exporter/collector/logs_test.go

@@ -139,6 +139,24 @@ func TestLogMapping(t *testing.T) {
 				},
 			},
 			maxEntrySize: defaultMaxEntrySize,
+		}, {
+			name: "log with invalid json byte body returns raw byte string",
+			log: func() plog.LogRecord {
+				log := plog.NewLogRecord()
+				log.Body().SetEmptyBytes().FromRaw([]byte(`"this is not json"`))
+				return log
+			},
+			mr: func() *monitoredrespb.MonitoredResource {
+				return nil
+			},
+			expectedEntries: []*logpb.LogEntry{
+				{
+					LogName:   logName,
+					Payload:   &logpb.LogEntry_TextPayload{TextPayload: "InRoaXMgaXMgbm90IGpzb24i"},
+					Timestamp: timestamppb.New(testObservedTime),
+				},
+			},
+			maxEntrySize: defaultMaxEntrySize,
 		},
 		{
 			name: "log with json and httpRequest, empty monitoredresource",