[KAFKA-9700] Negative estimatedCompressionRatio leads to misjudgment about if there is no room

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.5.0, 2.6.0
• Component/s: clients
• Flags: Patch
• Assignee: Unassigned
• Reporter: jiamei xie
• Created: 11/Mar/20 02:00
• Updated: 25/Mar/20 03:53
• Resolved: 25/Mar/20 03:53

Description

• When I run the following command
  bin/kafka-producer-perf-test.sh --topic test --num-records 50000000 --throughput -1 --record-size 5000 --producer-props bootstrap.servers=server04:9092 acks=1 buffer.memory=67108864 batch.size 65536 compression.type=zstd
  There was a warning:
  [2020-03-06 17:36:50,216] WARN [Producer clientId=producer-1] Got error produce response in correlation id 3261 on topic-partition test-1, splitting and retrying (2147483647 attempts left). Error: MESSAGE_TOO_LARGE (org.apache.kafka.clients.producer.internals.Sender)

• The batch size(65536) is smaller than max.message.bytes (1048588) . So it's not the root cause.

• I added some logs in CompressionRatioEstimator.updateEstimation and found there were negative currentEstimation values. The following were logs I added
  public static float updateEstimation(String topic, CompressionType type, float observedRatio) {
  float[] compressionRatioForTopic = getAndCreateEstimationIfAbsent(topic);
  float currentEstimation = compressionRatioForTopic[type.id];
  synchronized (compressionRatioForTopic) {
  if (observedRatio > currentEstimation) { compressionRatioForTopic[type.id] = Math.max(currentEstimation + COMPRESSION_RATIO_DETERIORATE_STEP, observedRatio); }

  else if (observedRatio < currentEstimation) {
  compressionRatioForTopic[type.id] = currentEstimation - COMPRESSION_RATIO_IMPROVING_STEP;
  log.warn("####currentEstimation is {} , COMPRESSION_RATIO_IMPROVING_STEP is {} , compressionRatioForTopic[type.id] is {}, type.id is {}", currentEstimation, COMPRESSION_RATIO_IMPROVING_STEP,compressionRatioForTopic[type.id], type.id);
  }
  }
  return compressionRatioForTopic[type.id];
  }

The observedRatio is smaller than COMPRESSION_RATIO_IMPROVING_STEP in some cases. Some I think the else if block should be changed into

else if (observedRatio < currentEstimation)

{ compressionRatioForTopic[type.id] = Math.max(currentEstimation - COMPRESSION_RATIO_IMPROVING_STEP, observedRatio); }

Issue Links

• links to: GitHub Pull Request #8285