Architecture

Apollo started as an experiment to see what it would take to make ActiveMQ work better on machines with higher core counts. It has resulted in broker that is much more deterministic, stable, and scaleable.

Architectural Changes

The major fundamental architectural changes it brings are:

Reactor Based Thread Model

Apollo uses HawtDispatch to implement the sever using a multi-threaded non-blocking variation of the reactor design pattern. HawtDispatch is a Java clone of libdispatch (aka Grand Central Dispatch). It uses a fixed sized thread pool and only executes non-blocking tasks on those threads.

The thread model allows Apollo to reach very high levels of scaleability and efficiency, but it places a huge restriction on the developer: all the tasks it executes must be non-blocking and ideally lock-free and wait free. This means that the previous ActiveMQ broker architecture had to go through a major overhaul. All synchronous broker interfaces had to be changed so that they would instead return results via asynchronous callbacks.

Scala 2.9 Implementation

Even though Apollo started as a fork of ActiveMQ 5.x, the new reactor design restrictions required major changes from the network IO handling, to the flow control design, all the way to the Store interfaces. Scala provided a much more concise way to express callbacks, namely by it's support for partial functions and closures.

Protocol Agnostic

ActiveMQ has supported multiple protocols for many years, but under the covers what it's doing is converting all the protocols to use the OpenWire messaging protocol. Naturally, this approach is not optimal if you want to efficiently support other protocols and not be limited by what OpenWire can support.

The Apollo server is much more modular and protocol agnostic. All protocols are equal and are built as a plugin to the the broker which just make use of exposed broker services for routing, flow control, queueing services etc. For example, this means that messages will be persisted in a Store in the original protocol's encoding. There is no protocol conversion occurring under the covers unless it is required.

REST Based Management

ActiveMQ choose to exposed it's management interfaces via JMX. JMX was a natural choice since ActiveMQ is Java based, but JMX has a couple of serious limitations:

Apollo exposes a rich and detailed state of the sever using REST based JSON services.

See the Management API documentation for details.

Message Swapping

Apollo can very efficiently work with both large and small queues due to the way it implements message swapping. If you have a large queue with millions of messages, and are slowly processing them, then it makes no sense to keep all those messages in memory. They just need to be loaded when the consumers are ready to receive them.

A queue in apollo has a configuration entry called consumer_buffer which is the amount of memory dedicated to that consumer for prefetching into memory the next set of messages that consumer will need. The queue will asynchronously load messages from the message store so that they will be in memory by the time the consumer is ready to receive the the message.

The rate of consumption/position of the consumers in the queue will also affect how newly enqueued messages are handled. If no consumers are near the tail of the queue where new messages are placed, then the message gets swapped out of memory asap. If they consumers are near the tail of the queue, then the message is retained in memory for as long as possible in hopes that you can avoid a swap out and then back in.

When a message is swapped out of memory, it can be in one of 2 swapped out states: 'swapped' or 'swapped range'. A message in 'swapped' state still has a small reference node in the list of messages the queue maintains. This small reference holds onto some accounting information about the message and how to quickly retrieve the message from the message store. Once a queue builds up many adjacent messages (defaults to 10,000) that are in the 'swapped' state, it will replace all those individual reference node entires in memory with a single range reference node. Once that happens, the message is in a 'swapped range'.