kafka 5 years feedback

src: 5 ans de bien et de moins bien avec Kafka (Nelson DIONISI et Matthieu MOUMINOUX) - Invidious - 2023-04-26

Abstract

• one topic for all:
  • 👎 one message slow down the other messages
  • 👎 complex monitoring
  • 👎 non specialized consumer
• functional topics
  • 👍 similar messages in processing time
  • 👍 simpler debugging and monitoring
• dead letter
  • 👍 no message loss
  • 👍 limited impact from one tenant to another
  • 👎 order not kept
• transactional messages
  • 👍 no message loss
  • 👍 no distributed transaction
  • 👎 queue based implementation in database
  • 👎 small loss of real time

Anti-patterns

• batch producer
  • let Kafka do the batch
• Kafka message size
  • do not send binaries
  • store the binaries in a file storage
  • send the Kafka message containing the file URL
• message processing time
  • have deterministic and constant processing time in the same topic
  • set the max.poll.interval.ms and max.poll.records depending on the expected consumption
• topic definition
  • no one topic for all
  • one topic per use case
  • define naming convention
    • e.g. <business domain>.<data description>.<classification>
    • catalog.product.cmd containing the message UpdateProduct, DeleteProduct

Number of partitions

• repartition ⇒ order loss
• it’s best to over-partition than the inverse
• choose a number of partition that is easy to divide (e.g. 12, 24 or 36)

Partition key

• partition by homogenous business value

Monitoring

• add a correlation id to the messages
• monitor the message production throughput
• monitor the consumer group lags

Access control

• no manual action
• kafka prod == database prod

Checklist

• producers
  • acks
  • ma.in.flight.requests.per.connection
  • partition keys
• consumers
  • max.poll.interval.ms
  • max.poll.records
  • monitor consumer group lags

Bonus: Kafka Properties Documentation