Kafka producer record-size-avg metric

[Zack Kobza]

The Kafka documentation states that the producer's record-size-avg metric is "The average record size", but this does not seem to be the case. The client code calculates the average maximum record size across producer batches for record-size-avg.  I am misinterpreting this metric? Should the documentation be updated to reflect this?

Here are code snippets from version 2.6 that show the creation of the sensor and how it is populated:

Sender.java:

    811: this.maxRecordSizeSensor = metrics.sensor("record-size");
    812: this.maxRecordSizeSensor.add(metrics.recordSizeMax, new Max());
    813: this.maxRecordSizeSensor.add(metrics.recordSizeAvg, new Avg());
    889: this.maxRecordSizeSensor.record(batch.maxRecordSize, now);

ProducerBatch.java:

    108 & 134: this.maxRecordSize = Math.max(this.maxRecordSize, AbstractRecords.estimateSizeInBytesUpperBound(magic(), recordsBuilder.compressionType(), key, value, headers));

[Zack Kobza]

To add more flavor, I have been using this value for performance testing and was mislead because the record-size-avg is not the average record size the client was sending. I calculated the average record size separately and it did not equal the value reported by the metric.