...

The produce latency is an important metric to monitor for the cluster performance.   As we know, latency is the sum from those parts, RequestQueueTime, LocalTime, RemoteTime, ResponseSend/QueueTime, ThrottleTime, MessageConversionsTime.  RemoteTime means the time spent in the purgatory. 

...

However, the produce remoteTime is hard to troubleshoot.   That is the time spent by the leader waiting for the follower to fetch the data and send back the confirmed watermark. It can be affected by 1. Network is congested,  2. Fetch processing itself is slow.  3. Something else.

Currently we don't have the correct metrics to tell which part is contributing to the high Produce RemoteTime.    That That is because,   the reported fetch latency didn't reflect the true fetch latency because it sometimes needs to stay in purgatory and wait for replica.fetch.wait.max.ms when data is more than fetch.min.bytes.   This greatly skewed the real latency number. 

...

Add waitTimeMs in FetchResponse, and bump the FetchResponse and FetchRequest protocol version. 

FetchResponse => TransactionalId RequiredAcks Timeout [FetchableTopicResponse]
   ThrottleTimeMs => INT32
   WaitTimeMs => INT32  # Add a new field to record the request wait time in purgatory
   Timeout => INT32
   ErrorCode => INT16
   SessionId => INT32
   FetchableTopicResponse => [Name [FetchablePartitionResponse] PreferredReadReplica Records]
   ......

Proposed Changes

We like to propose to track the real end to end fetch latency with those changes:

...

Additionally,  at the leader side, we will add a new metric called TotalLocalTime, which is TotalLocalTime = TotalTime - RemoteTime. This metric measures the time spent to process the fetch request on leader side, excluding the time spent in the purgatory.

...

To better illustrate how the proposed fetch latency metric can be used for monitoring the latency between each pair of brokers, we simulated network latency by introducing artificial delays to the incoming packets on one broker(broker 0) of a 3-node cluster. 

As the following graph shows, the broker fetch latency increased on all three brokers, pegging at 500ms(replica.fetch.wait.max.ms). This gave us very little information on where the slowness was introduced.

...