As many of you know, OpenStack is a fully distributed system. As such, it keeps its services (nova, glance, cinder, keystone, etc ) as decoupled as possible and tries to stick to most of the distribution paradigms, deployments strategies and architectures. For example, one of the main tenets throughout OpenStack is that every module should be using Shared Nothing Architecture (SNA) which states. that each node should be independent and self-sufficient. In other words, all nodes in a SNA are completely isolated from each other in terms of space and memory.
There are other distribution principles that are part of OpenStack's tenets, however, this post is not about what principles OpenStack as a whole tries to follow, but rather on how OpenStack sticks together such a heavily distributed architecture and makes it work as one. The first thing we need to do is evaluate some of the integration methods that exist out there and how they're being used within OpenStack. Before we get there, let me explain what an integration method is.
What does integration mean?
An integration method is the way 2 or more distributed services – any kind of services – interact with each other. It may refer to how they exchange states, tasks or just data. There are different ways to integrate distributed systems. In this post we'll discuss some (databases, rpc calls, messaging), which are the ones used within OpenStack.
The integration of a distributed system relies on the capability of its components to communicate between one another. This communication can happen in several ways and could go through different protocols. Each one of these methods have some benefits and drawbacks. Some of them are consistent whereas others are more reliable and scalable.
Databases
I dare to say databases are the most common way of integration nowadays. Integrating 2 or more services is about making those nodes communicate with each other, regardless the scope of that communication. The services able to communicate with each other will use that ability to send data from point A to B, regardless what that data is considered to be. Most of the services throughout OpenStack rely on databases as a way of integration. For example, cinder uses the database not just to store data but also to share volumes state with other cinder nodes. Furthermore, cinder-api for example, is responsible for creating the volume when a volume creation is requested, it'll then forward the request to the cinder-scheduler, which will then send a creation request to cinder-volume. The later reads the data stored by cinder-api and will update it as needed. That way both nodes, cinder-api and cinder-volume, shared data by using a database as a way to communicate with each other.
The data traveling between those services could be anything: events, messages, notifications, database records etc. It doesn't matter, at the very end if its data and what matters is how it is being generated, how it's being sent, how it's being serialized and how it's being consumed.
The reason I mentioned all that is because databases conceptually are not message brokers, although they allow 2 or more services to communicate and share data between them. Databases are collections of structured--and don't play the unstructured NoSQL card here--data. This data is usually organized and can be created, read, updated and deleted at any time.
It is pretty common to see different services relying on a database as a way to share data. It is even more common to see that happening for a horizontally scaled service. Some people would argue saying that nodes of the same service that rely on a database are not actually being integrated by it, although they are. If a service relies on a database, it means it uses it to store information that will be useful to other nodes of the same service running in parallel. Otherwise, each node of that service would be a rogue, isolated instance and that would break the consistency of the whole distributed system.
This is the first level of integration that services throughout OpenStack use. Most of the services depend on a database. However, none of them share the database access with services that don’t belong to their cluster. For example, Nova instances won't ever access Cinder's database directly, instead, they'll go through Cinder's API service. The motivation for this is pretty simple, Nova doesn't have any knowledge of how volumes should be handled, nor what state they are in or when their state should be changed. That's something Cinder has to take care of.
Nevertheless, different nodes of the same service do share access to the same database. For example, cinder-volume, cinder-scheduler and cinder-api access the same database concurrently to operate on the volumes data, among other things.
Remote Procedure Calls
As I already mentioned, I believe databases are the most common way to integrate services nowadays, many people do it without even knowing they are. However, as far as OpenStack is concerned, I believe RPC (remote procedure call) is the most used method. Almost 90% of the integrated projects--those that are part of OpenStack's release cycle--rely on RPC calls to communicate with other services.
But, what is RPC exactly?
Remote Procedure Call is a way of inter-process communication that allows clients to trigger the execution of subroutines in remote locations. There are many different implementations of RPC, although there's an--expired - RFC (Request For Comments) for it[1]. Different patterns have been created based on those implementations.
RPC subroutines can do anything. In many cases, they're responsible for storing and retrieving data. In other cases\they just execute code and return the result to the caller. Likewise, the channel through which RPC calls are sent could be anything, message brokers, raw TCP connections, etc, as long as it's possible to pack and send the message through the channel.
RPC has many benefits that databases don't. For example, it's possible to make synchronous calls and wait for responses whereas with databases the data transmission is always asynchronous. RPC also makes it easier for services to isolate operations and communications. However, just like databases, RPC is a very tightly coupled protocol that requires both endpoints to agree on a structure. This brings consistency and 'predictability' to the protocol at the cost of making it less flexible and obviously less decoupled.
Throughout OpenStack, modules using RPC rely on a messaging library - oslo.messaging - that takes care of the message serialization, transport communication and message acknowledgment. As for the time of this writing, the library has support for 3 different drivers - rabbit, qpid and zmq - that handle the underlying message bus this library sits on top of.
Some projects like Nova, Cinder and Neutron use RPC calls heavily. Almost everything that happens in those projects is triggered by RPC calls. For example, a volume creation request in Cinder is first sent from the cinder-api service to the cinder-scheduler which will then pick one of the cinder-volume nodes that are available and forward the request to it. All this happens in an asynchronous fashion.
Throughout OpenStack, asynchronous calls (cast) are used most of the time. One of the benefits behind this is that services can handle more load since they're not blocking on every call, therefore *-api services can return a response back to the client in less time.
Notice we're not taking reliability, scalability under consideration here but just the ability to of the message channel to take the message from the sender to the receiver
Messaging
NOTE: This section is not about message brokers. It refers to the use of messages as an atomic unit for sharing data between services, regardless if they're homogeneous or not.
Messaging is perhaps the most decoupled method of integration. It's based on messages that travel through a channel which gets them from the producer to the consumer. Services sending and consuming messages don't necessarily need to agreed on a structure to use. Furthermore, consumers are not suppose to consume all messages, it all depends on the messaging pattern being used.
Messaging--loose coupling to be more precise gives flexibility and scalability to the cost of being more complex in terms of implementations. Although services being integrated are not required to agreed on a structure, they do expect to get a message they can read the data from. Furthermore, the message channel--whatever it is--may need to have support for message routing, message transformation and message filtering among other things. All this is mostly coupled to the architecture but not to the services that are part of it.
Something interesting about messaging is that it can be the base for other integration methods. For example, it is possible to send messages containing RPC requests. In order to do this, though, it is necessary to apply all RPC requirements to messaging, for instance, both parts-producer and consumer--will need to agreed on the message structure and type.
Within OpenStack, messages are mostly used by the RPC library--(oslo.messaging) itself. The library packages all RPC requests into messages that are sent to a message broker which then forwards them to the consumers. There are cases, though, where messaging is used in an almost fully decoupled and asynchronous way. Notifications, for instance, are messages sent by OpenStack services hoping they’ll be a consumer ready to process them. These notifications contain information about events that had happened at a very specific moment.
In Glance, for example, when an image is downloaded an image.download notification is sent and it'll hopefully be consumed. Whatever it is that happens to that notification, Glance doesn't necessarily care about it. However, Ceilometer is a good example - perhaps the only one at the time of this writing - of a service interested in those notifications. It consumes all these events to meter resources usage and to allow users to bill based on that information.
It is now clear that messaging is heavily used in OpenStack and that most of the messages sent between homogeneous services are RPC calls. These calls travel through the message broker picked in the form of atomic messages and they are processed - at least most of them – asynchronously. The asynchronous nature of OpenStack's interoperability helps keep developers focused on making distributed nodes faster, more scalable and more reliable.
Wrapping up
This post covered 3 different integration methods—databases, RPC and messaging--. how they're used throughout OpenStack, and how they're mixed together to reach great levels of integration. The post also covered some of the benefits of each one of the methods over the others and touched on some of the drawbacks of each method as well.
One thing to keep in mind, though, is that OpenStack couldn't have been implemented in any other way. The fact that it relies on such a heavily distributed architecture has helped the project to succeed. The mixed integration styles this architecture allow the project to have logically - as opposed to functionally - distributed services in several nodes. Furthermore, it allows OpenStack to scale massively and dynamically.
However, it's true that not everything is perfect and that there's definitely room for improvement. For instance, the fact that most of the operations throughout OpenStack rely on a message broker is not funny. We all know message brokers are hard –and can be impossible to scale. It is easy for message brokers to become a SPOF (Single Point of Failure) in the architecture, which means that a big part of your deployment will be in tremendous trouble if one of the message brokers would go down. I believe this is something that peer-to-peer protocols - amqp 1.0, could alleviate.
[1] https://tools.ietf.org/html/rfc5531
Sobre o autor
Navegue por canal
Automação
Últimas novidades em automação de TI para empresas de tecnologia, equipes e ambientes
Inteligência artificial
Descubra as atualizações nas plataformas que proporcionam aos clientes executar suas cargas de trabalho de IA em qualquer ambiente
Nuvem híbrida aberta
Veja como construímos um futuro mais flexível com a nuvem híbrida
Segurança
Veja as últimas novidades sobre como reduzimos riscos em ambientes e tecnologias
Edge computing
Saiba quais são as atualizações nas plataformas que simplificam as operações na borda
Infraestrutura
Saiba o que há de mais recente na plataforma Linux empresarial líder mundial
Aplicações
Conheça nossas soluções desenvolvidas para ajudar você a superar os desafios mais complexos de aplicações
Programas originais
Veja as histórias divertidas de criadores e líderes em tecnologia empresarial
Produtos
- Red Hat Enterprise Linux
- Red Hat OpenShift
- Red Hat Ansible Automation Platform
- Red Hat Cloud Services
- Veja todos os produtos
Ferramentas
- Treinamento e certificação
- Minha conta
- Suporte ao cliente
- Recursos para desenvolvedores
- Encontre um parceiro
- Red Hat Ecosystem Catalog
- Calculadora de valor Red Hat
- Documentação
Experimente, compre, venda
Comunicação
- Contate o setor de vendas
- Fale com o Atendimento ao Cliente
- Contate o setor de treinamento
- Redes sociais
Sobre a Red Hat
A Red Hat é a líder mundial em soluções empresariais open source como Linux, nuvem, containers e Kubernetes. Fornecemos soluções robustas que facilitam o trabalho em diversas plataformas e ambientes, do datacenter principal até a borda da rede.
Selecione um idioma
Red Hat legal and privacy links
- Sobre a Red Hat
- Oportunidades de emprego
- Eventos
- Escritórios
- Fale com a Red Hat
- Blog da Red Hat
- Diversidade, equidade e inclusão
- Cool Stuff Store
- Red Hat Summit