The team was recently asked about whether and how quorum queues can offer the same message ordering guarantees as classic queues given that they will deliver messages from a local queue replica (leader or follower) when possible. Mirrored queues always deliver from the master (the leader), so delivering from any queue replica sounds like it could impact those guarantees. 

That is the subject of this post. Be warned, this post is a technical deep dive for the curious and the distributed systems enthusiast. We’ll take a look at how quorum queues can deliver messages from any queue replica, leader or follower, without additional coordination (extra to Raft) but maintaining message ordering guarantees.

In the last post we started a sizing analysis of our workload using quorum queues. We focused on the happy scenario that consumers are keeping up meaning that there are no queue backlogs and all brokers in the cluster are operating normally. By running a series of benchmarks modelling our workload at different intensities we identified the top 5 cluster size and storage volume combinations in terms of cost per 1000 msg/s per month.

1. Cluster: 7 nodes, 8 vCPUs (c5.2xlarge), gp2 SDD. Cost: $54
2. Cluster: 9 nodes, 8 vCPUs (c5.2xlarge), gp2 SDD. Cost: $69
3. Cluster: 5 nodes, 8 vCPUs (c5.2xlarge), st1 HDD. Cost: $93
4. Cluster: 5 nodes, 16 vCPUs (c5.4xlarge), gp2 SDD. Cost: $98
5. Cluster: 7 nodes, 16 vCPUs (c5.4xlarge), gp2 SDD. Cost: $107

There are more tests to run to ensure these clusters can handle things like brokers failing and large backlogs accumulating during things like outages or system slowdowns.

All quorum queues are declared with the following properties:

• x-quorum-initial-group-size=3
• x-max-in-memory-length=0

The x-max-in-memory-length property forces the quorum queue to remove message bodies from memory as soon as it is safe to do. You can set it to a longer limit, this is the most aggressive - designed to avoid large memory growth at the cost of more disk reads when consumers do not keep up. Without this property message bodies are kept in memory at all times which can place memory growth to the point of memory alarms setting off which severely impacts the publish rate - something we want to avoid in this workload case study.

In a first post in this sizing series we covered the workload, the tests, and the cluster and storage volume configurations on AWS ec2. In this post we’ll run a sizing analysis with quorum queues. We also ran a sizing analysis on mirrored queues.

In this post we'll run the increasing intensity tests that will measure our candidate cluster sizes at varying publish rates, under ideal conditions. In the next post we'll run resiliency tests that measure whether our clusters can handle our target peak load under adverse conditions.

All quorum queues are declared with the following properties:

• x-quorum-initial-group-size=3 (replication factor)
• x-max-in-memory-length=0

The x-max-in-memory-length property forces the quorum queue to remove message bodies from memory as soon as it is safe to do. You can set it to a longer limit, this is the most aggressive - designed to avoid large memory growth at the cost of more disk reads when consumers do not keep up. Without this property message bodies are kept in memory at all times which can place memory growth to the point of memory alarms setting off which severely impacts the publish rate - something we want to avoid in this workload case study.

In the last post we started a sizing analysis of our workload using mirrored queues. We focused on the happy scenario that consumers are keeping up meaning that there are no queue backlogs and all brokers in the cluster are operating normally. By running a series of benchmarks modelling our workload at different intensities we identified the top 5 cluster size and storage volume combinations in terms of cost per 1000 msg/s per month.

1. Cluster: 5 nodes, 8 vCPUs, gp2 SDD. Cost: $58
2. Cluster: 7 nodes, 8 vCPUs, gp2 SDD. Cost: $81
3. Cluster: 5 nodes, 8 vCPUs, st1 HDD. Cost: $93
4. Cluster: 5 nodes, 16 vCPUs, gp2 SDD. Cost: $98
5. Cluster: 9 nodes, 8 vCPUs, gp2 SDD. Cost: $104

There are more tests to run to ensure these clusters can handle things like brokers failing and large backlogs accumulating during things like outages or system slowdowns.

In a first post in this sizing series we covered the workload, cluster and storage volume configurations on AWS ec2. In this post we’ll run a sizing analysis with mirrored queues.

The first phase of our sizing analysis will be assessing what intensities each of our clusters and storage volumes can handle easily and which are too much.

All tests use the following policy:

• ha-mode: exactly
• ha-params: 2
• ha-sync-mode: manual

This is the start of a short series where we look at sizing your RabbitMQ clusters. The actual sizing wholly depends on your hardware and workload, so rather than tell you how many CPUs and how much RAM you should provision, we’ll create some general guidelines and use a case study to show what things you should consider.

There can be many reasons to do benchmarking:

• Sizing and capacity planning
• Product assessment (can RabbitMQ handle my load?)
• Discover best configuration for your workload

In this post we’ll take a look at the various options for running RabbitMQ benchmarks. But before we do, you’ll need a way to see the results and look at system metrics.

A Webinar on Quorum Queues​

Before we start with RabbitMQ project and community updates from April, we have a webinar to announce! Jack Vanlightly, a RabbitMQ core team member, will present on High Availability and Data Safety in Messaging on June 11th, 2020.

In this webinar, Jack Vanlightly will explain quorum queues, a new replicated queue type in RabbitMQ. Quorum queues were introduced in RabbitMQ 3.8 with a focus on data safety and efficient, predictable recovery from node failures. Jack will cover and contrast the design of quorum and classic mirrored queues.

After this webinar, you'll understand:

• Why quorum queues offer better data safety than mirrored queues
• How and why server resource usage changes when switching to quorum queues from mirrored queues
• Some best practices when using quorum queues

In the last post we ran some simple benchmarks on a single queue to see what effect pipelining publisher confirms and consumer acknowledgements had on flow control. 

Specifically we looked at:

• Publishers: Restricting the number of in-flight messages (messages sent but pending a confirm).
• Consumers: Prefetch (the number in-flight messages the broker will allow on the channel)
• Consumers: Ack Interval (multiple flag usage)

Unsurprisingly, we saw when we restricted publishers and the brokers to a small number of in-flight messages at a time, that throughput was low. When we increased that limit, throughput increased, but only to a point, after which we saw no more throughput gains but instead just latency increases. We also saw that allowing consumers to use the multiple flag was beneficial to throughput.

In this post we’re going to look at those same three settings, but with many clients, many queues and different amounts of load, including stress tests. We’ll see that publisher confirms and consumer acknowledgements play a role in flow control to help prevent overload of a broker. 

In the last post we covered what flow control is, both as a general concept and the various flow control mechanisms available in RabbitMQ. We saw that publisher confirms and consumer acknowledgements are not just data safety measures, but also play a role in flow control. 

In this post we’re going to look at how application developers can use publisher confirms and consumer acknowledgements to get a balance of safety and high performance, in the context of a single queue. 

Flow control becomes especially important when a broker is being overloaded. A single queue is unlikely to overload your broker. If you send large messages then sure, you can saturate your network, or if you only have a single CPU core, then one queue could max it out. But most of us are on 8, 16 or 30+ core machines. But it’s interesting to break down the effects of confirms and acks on a single queue. From there we can take our learnings and see if they apply to larger deployments (the next post).