...
Upgrade the brokers once with the inter-broker protocol set to the previous deployed version.
Upgrade the brokers again with an updated inter-broker protocol, but leaving the message format unchanged.
Notify clients that they can upgrade, BUT should not start using the idempotent / transactional message APIs yet.
[When observed that most of the clients have upgraded] Restart the brokers, with the message format version set to the latest.
Notify upgraded clients that they can now start using the idempotent / transactional message APIs.
The reason for step 3 is to avoid the performance cost for down-converting messages to an older format, which effectively loses the “zero-copy” optimization. Ideally, all consumers are upgraded before the producers even begin writing to the new message format.
Note: Since the old producer has long since been deprecated and the old consumer will be deprecated in 0.11.0, these clients will not support the new format. In order to avoid the conversion hit, users will have to upgrade to the new clients. It is possible to selectively enable the message format on topics which are already using the new clients.
Test Plan
Correctness
The new features will be tested through unit, integration, and system tests.
The integration tests will focus on ensuring that the basic guarantees (outlined in the Summary of Guarantees section) are satisfied across components.
The system tests will focus on ensuring that the guarantees are satisfied even with failing components, ie. that the system works even when consumers, producers, brokers are killed in various states.
We will also add to existing compatibility system tests to ensure that old clients can still talk to the new brokers with the new message format.
Performance
This KIP introduces significant changes to the message format along with the new features.
We plan on introducing changes in a staged fashion, with the first change being to the message format. We will run our performance test suite on these message format changes and ensure that there is a minimal performance impact thanks to these changes at worst. Note that the message format changes are the only ones which can affect users who don't enable the idempotent producer and don't use transactions.
Then, we will benchmark the performance of the idempotent producer and the transactional producer separately. Finally, we will benchmark the consumer and broker performance when transactions are in use and read_committed mode is enabled. We will publish the results of all these benchmarks so that users can make informed decisions about when and how to use these features.
Rejected Alternatives
As mentioned earlier, we have a separate design document which explores the design space --including rejected alternatives-- as well as all the implementation details. The latter also includes the specifics of message format changes, new RPCs, error handling, etc.
The design document is available here.
...