Brian Byrne created KAFKA-8904:
----------------------------------
Summary: Reduce metadata lookups when producting to a large number
of topics
Key: KAFKA-8904
URL: https://issues.apache.org/jira/browse/KAFKA-8904
Project: Kafka
Issue Type: Improvement
Components: controller, producer
Reporter: Brian Byrne
Per [~lbradstreet]:
"The problem was that the producer starts with no knowledge of topic metadata.
So they start the producer up, and then they start sending messages to any of
the thousands of topics that exist. Each time a message is sent to a new topic,
it'll trigger a metadata request if the producer doesn't know about it. These
metadata requests are done in serial such that if you send 2000 messages to
2000 topics, it will trigger 2000 new metadata requests.
Each successive metadata request will include every topic seen so far, so the
first metadata request will include 1 topic, the second will include 2 topics,
etc.
An additional problem is that this can take a while, and metadata expiry (for
metadata that has not been recently used) is hard coded to 5 mins, so if this
the initial fetches take long enough you can end up evicting the metadata
before you send another message to a topic.
So the approaches above are:
1. We can linger for a bit before making a metadata request, allow more sends
to go through, and then batch the metadata request for topics we we need in a
single metadata request.
2. We can allow pre-seeding the producer with metadata for a list of topics you
care about.
I prefer 1 if we can make it work."
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
