Dear RabbitMQ community,
We have an update on Erlang/OTP 20 and RabbitMQ compatibility.
First, let's clarify when OTP 20 support will be available: RabbitMQ versions before 3.6.11 will not work correctly with OTP-20.
What's the risk of upgrading from 19.x to 20 on an unsupported version? When upgrading from OTP-19.x(or earlier) to OTP-20 all the persistent data will be permanently lost!
What about newly provisioned nodes or clusters? Although it's possible to run RabbitMQ with OTP-20 from scratch, there will be crashes in queue mirroring and management API.
Key breaking change for us is the new external (binary) terms format, which was introduced in R16 but now is used by default. This term format will encode atoms as unicode with a different type tag. Unfortunately there is no way to generate a value in the old format in OTP itself, so it's not easy to perform migrations.
The format change will affect PID decoding and encoding (as in rabbit_misc:decompose_pid/1)
, which will break mirroring, node renaming and parts of management API.
The issue was addressed in https://github.com/rabbitmq/rabbitmq-common/commit/4aa1bf4c1450fa8ba60b381c8b295e03fbd17c44
We also use term_to_binary to generate hash IDs for various entities in the system.
While some entities are transient, like manegement binadings properties and
federation links IDs, some of them are persistent and should not change during upgrade.
An umbrella issue for that is rabbitmq/rabbitmq-server#1243
It's been decided to not use term_to_binary to generate hashes in future versions
(3.7 and above).
Management and federation hash strategies were updated in:
rabbitmq/rabbitmq-management-agent#47 ,
rabbitmq/rabbitmq-management#415
and
rabbitmq/rabbitmq-federation#58
Hashing that involves term_to_binary/1 is used when calculating queue index directory names
and queue names in the STOMP plugin. They should be addressed in 3.6.11.
Since we don't have upgrade steps for patch releases and cannot rename
directories and queue names, we decided to generate the old term_to_binary
format for specific types we use in our own code.
The generation functions are in the term_to_binary_compat module.
In case of STOMP queue generation, a tuple of strings or binaries was used, so the function
is called string_and_binary_tuple_2_to_binary and accepts a 2-element tuple of
strings or binaries.
The issue is rabbitmq/rabbitmq-stomp#115
This is addressed in
rabbitmq/rabbitmq-server#1262
and
rabbitmq/rabbitmq-stomp#116
Queue index directories is the most serious change, since RabbtiMQ will delete all "unknown" queue index directories on boot. When queue index names change, all the directories become "unknown", so all the persistent data will be lost. GitHub issue for this specific problem is rabbitmq/rabbitmq-server#1243.
In 3.6 we use compat function queue_name_to_binary to generate the same
format of the directory name as in pre-20.
This is addressed in rabbitmq/rabbitmq-server#1246.
In 3.7 we decided to generate a queue index name with different algorithm, so
we won't be affected with further changes in term_to_binary in future versions.
The directories should be renamed during migration step.
This will require to generate the old name first using compat function.
This is addressed in rabbitmq/rabbitmq-server#1250.
To summarise, we are trying to make 3.6.11 and 3.7.0 (by the time it reaches RC stage) support OTP 20 and introducing a CI pipeline that will test Erlang upgrades to 20. Any updates that are worth mentioning will be posted in this thread.
Cheers.