|Reported by:||dustin||Owned by:|
Description (last modified by dustin)
When Buildbot adopts an MQ framework, there are a number of choices. Among them:
- pure-python, internal implementation (no network, only a single master)
- pure-python, internal implementation that uses TCP or UDP between masters
The first is already implemented in 'nine', but we need more.
The more difficult part of this project is to ensure that messages are only delivered after the corresponding database changes are visible.
The issue is a race condition between message passing and database replication. Imagine you have a large Buildbot installation with several replicated MySQL servers and a redundant RabbitMQ cluster. One buildbot master writes changes to a build to the database (an UPDATE operation), then sends a message describing the change. On another master, some service gets the message and queries a different MySQL server to see the build's new status. If the message arrives before the database replication occurs, then this master will see stale data. That will lead to a lot of subtle, rare bugs.
This project would involve separate tasks:
- implement one or more of the above MQ plugins (mostly just coding)
- solve the data-ordering problem (requiring some CS theory)