training-guides/doc/training-guides/module001-ch008-queues-mess...

438 lines
19 KiB
XML
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

<?xml version="1.0" encoding="utf-8"?>
<chapter xmlns="http://docbook.org/ns/docbook"
xmlns:xi="http://www.w3.org/2001/XInclude"
xmlns:xlink="http://www.w3.org/1999/xlink" version="5.0"
xml:id="module001-ch008-queues-messaging">
<title>OpenStack Messaging and Queues</title>
<figure>
<title>Messaging in OpenStack</title>
<mediaobject>
<imageobject>
<imagedata fileref="figures/image04.png" contentwidth="7in"/>
</imageobject>
</mediaobject>
</figure>
<para>AMQP is the messaging technology chosen by the OpenStack
cloud. The AMQP broker, either RabbitMQ or Qpid, sits between any
two Nova components and allows them to communicate in a loosely
coupled fashion. More precisely, Nova components (the compute
fabric of OpenStack) use Remote Procedure Calls (RPC hereinafter)
to communicate to one another; however such a paradigm is built
atop the publish/subscribe paradigm so that the following benefits
can be achieved:</para>
<itemizedlist>
<listitem>
<para>Decoupling between client and servant (such as the client
does not need to know where the servant reference
is).</para>
</listitem>
<listitem>
<para>Full a-synchronism between client and servant (such as the
client does not need the servant to run at the same time of
the remote call).</para>
</listitem>
<listitem>
<para>Random balancing of remote calls (such as if more servants
are up and running, one-way calls are transparently dispatched
to the first available servant).</para>
</listitem>
</itemizedlist>
<para>Nova uses direct, fanout, and topic-based exchanges. The
architecture looks like the one depicted in the figure
below:</para>
<figure>
<title>AMQP</title>
<mediaobject>
<imageobject>
<imagedata fileref="figures/image24.png" contentwidth="7in"/>
</imageobject>
</mediaobject>
</figure>
<para>Nova implements RPC (both request+response, and one-way,
respectively nicknamed rpc.call and rpc.cast) over AMQP by
providing an adapter class which take cares of marshaling and
un-marshaling of messages into function calls. Each Nova service,
such as Compute, Scheduler, and so on, creates two queues at the
initialization time, one which accepts messages with routing keys
NODE-TYPE.NODE-ID, for example, compute.hostname, and another,
which accepts messages with routing keys as generic NODE-TYPE, for example compute. The former is used specifically when
Nova-API needs to redirect commands to a specific node like
euca-terminate instance. In this case, only the compute node
whose hosts hypervisor is running the virtual machine can kill
the instance. The API acts as a consumer when RPC calls are
request/response, otherwise is acts as publisher only.</para>
<para><guilabel>Nova RPC Mappings</guilabel></para>
<para>The figure below shows the internals of a message broker node
(referred to as a RabbitMQ node in the diagrams) when a single
instance is deployed and shared in an OpenStack cloud. Every component
within Nova connects to the message broker and, depending on its
personality, such as a compute node or a network node, may
use the queue either as an Invoker (such as API or Scheduler) or a
Worker (such as Compute or Network). Invokers and Workers do not
actually exist in the Nova object model, but in this example they are used
as an abstraction for the sake of clarity. An Invoker is a
component that sends messages in the queuing system using <command>rpc.call
</command> and <command>rpc.cast</command>. A worker is a component
that receives messages from the queuing system and replies
accordingly to rcp.call operations.</para>
<para>Figure 2 shows the following internal elements:</para>
<itemizedlist>
<listitem>
<para><emphasis role="bold">Topic Publisher:</emphasis> A Topic
Publisher comes to life when an rpc.call or an rpc.cast
operation is executed; this object is instantiated and used to
push a message to the queuing system. Every publisher connects
always to the same topic-based exchange; its life-cycle is
limited to the message delivery.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Direct Consumer:</emphasis> A Direct
Consumer comes to life if (an only if) a rpc.call operation is
executed; this object is instantiated and used to receive a
response message from the queuing system; Every consumer
connects to a unique direct-based exchange via a unique
exclusive queue; its life-cycle is limited to the message
delivery; the exchange and queue identifiers are determined by
a UUID generator, and are marshaled in the message sent by the
Topic Publisher (only rpc.call operations).</para>
</listitem>
<listitem>
<para><emphasis role="bold">Topic Consumer:</emphasis> A Topic
Consumer comes to life as soon as a Worker is instantiated and
exists throughout its life-cycle; this object is used to
receive messages from the queue and it invokes the appropriate
action as defined by the Worker role. A Topic Consumer
connects to the same topic-based exchange either via a shared
queue or via a unique exclusive queue. Every Worker has two
topic consumers, one that is addressed only during rpc.cast
operations (and it connects to a shared queue whose exchange
key is topic) and the other that is addressed only during
rpc.call operations (and it connects to a unique queue whose
exchange key is topic.host).</para>
</listitem>
<listitem>
<para><emphasis role="bold">Direct Publisher:</emphasis> A Direct
Publisher comes to life only during rpc.call operations and it
is instantiated to return the message required by the
request/response operation. The object connects to a
direct-based exchange whose identity is dictated by the
incoming message.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Topic Exchange:</emphasis> The
Exchange is a routing table that exists in the context of a
virtual host (the multi-tenancy mechanism provided by Qpid or
RabbitMQ); its type (such as topic vs. direct) determines the
routing policy; a message broker node will have only one
topic-based exchange for every topic in Nova.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Direct Exchange:</emphasis> This is a
routing table that is created during rpc.call operations;
there are many instances of this kind of exchange throughout
the life-cycle of a message broker node, one for each rpc.call
invoked.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Queue Element:</emphasis> A Queue is
a message bucket. Messages are kept in the queue until a
Consumer (either Topic or Direct Consumer) connects to the
queue and fetch it. Queues can be shared or can be exclusive.
Queues whose routing key is topic are shared amongst Workers
of the same personality.</para>
</listitem>
</itemizedlist>
<figure>
<title>RabbitMQ</title>
<mediaobject>
<imageobject>
<imagedata fileref="figures/image20.png" contentwidth="7in"/>
</imageobject>
</mediaobject>
</figure>
<para><guilabel>RPC Calls</guilabel></para>
<para>The diagram below shows the message flow during an rp.call
operation:</para>
<orderedlist>
<listitem>
<para>A Topic Publisher is instantiated to send the message
request to the queuing system; immediately before the
publishing operation. A Direct Consumer is instantiated to
wait for the response message.</para>
</listitem>
<listitem>
<para>Once the message is dispatched by the exchange, it is
fetched by the Topic Consumer dictated by the routing key
(such as topic.host) and passed to the Worker in charge of
the task.</para>
</listitem>
<listitem>
<para>Once the task is completed, a Direct Publisher is
allocated to send the response message to the queuing
system.</para>
</listitem>
<listitem>
<para>Once the message is dispatched by the exchange, it is
fetched by the Direct Consumer dictated by the routing key
(such as msg_id) and passed to the Invoker.</para>
</listitem>
</orderedlist>
<figure>
<title>RabbitMQ</title>
<mediaobject>
<imageobject>
<imagedata fileref="figures/image28.png" contentwidth="7in"/>
</imageobject>
</mediaobject>
</figure>
<para><guilabel>RPC Casts</guilabel></para>
<para>The diagram below the message flow during an rp.cast
operation:</para>
<orderedlist>
<listitem>
<para>A Topic Publisher is instantiated to send the message
request to the queuing system.</para>
</listitem>
<listitem>
<para>Once the message is dispatched by the exchange, it is
fetched by the Topic Consumer dictated by the routing key
(such as topic) and passed to the Worker in charge of the
task.</para>
</listitem>
</orderedlist>
<figure>
<title>RabbitMQ</title>
<mediaobject>
<imageobject>
<imagedata fileref="figures/image20.png" contentwidth="7in"/>
</imageobject>
</mediaobject>
</figure>
<para><guilabel>AMQP Broker Load</guilabel></para>
<para>At any given time the load of a message broker node running
either Qpid or RabbitMQ is a function of the following
parameters:</para>
<itemizedlist>
<listitem>
<para>Throughput of API calls: the number of API calls (more
precisely rpc.call ops) being served by the OpenStack cloud
dictates the number of direct-based exchanges, related queues
and direct consumers connected to them.</para>
</listitem>
<listitem>
<para>Number of Workers: there is one queue shared amongst
workers with the same personality; however there are as many
exclusive queues as the number of workers; the number of
workers dictates also the number of routing keys within the
topic-based exchange, which is shared amongst all
workers.</para>
</listitem>
</itemizedlist>
<para>The figure below shows the status of a RabbitMQ node after
Nova components bootstrap in a test environment. Exchanges and
queues being created by Nova components are:</para>
<itemizedlist>
<listitem>
<para>Exchanges</para>
</listitem>
</itemizedlist>
<orderedlist>
<listitem>
<para>nova (topic exchange)</para>
</listitem>
</orderedlist>
<itemizedlist>
<listitem>
<para>Queues</para>
</listitem>
</itemizedlist>
<orderedlist>
<listitem>
<para>compute.phantom (phantom is the hostname)</para>
</listitem>
<listitem>
<para>compute</para>
</listitem>
<listitem>
<para>network.phantom (phantom is the hostname)</para>
</listitem>
<listitem>
<para>network</para>
</listitem>
<listitem>
<para>scheduler.phantom (phantom is the hostname)</para>
</listitem>
<listitem>
<para>scheduler</para>
</listitem>
</orderedlist>
<para><guilabel>RabbitMQ Gotchas</guilabel></para>
<para>Nova uses Kombu to connect to the RabbitMQ environment. Kombu
is a Python library that in turn uses AMQPLib, a library that
implements the standard AMQP 0.8 at the time of writing. When
using Kombu, Invokers and Workers need the following parameters in
order to instantiate a Connection object that connects to the
RabbitMQ server (please note that most of the following material
can be also found in the Kombu documentation; it has been
summarized and revised here for the sake of clarity):</para>
<itemizedlist>
<listitem>
<para><emphasis role="bold">Hostname:</emphasis> The hostname to
the AMQP server.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Userid:</emphasis> A valid username
used to authenticate to the server.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Password:</emphasis> The password
used to authenticate to the server.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Virtual_host:</emphasis> The name of
the virtual host to work with. This virtual host must exist on
the server, and the user must have access to it. Default is
“/”.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Port:</emphasis> The port of the
AMQP server. Default is 5672 (amqp).</para>
</listitem>
</itemizedlist>
<para>The following parameters are default:</para>
<itemizedlist>
<listitem>
<para><emphasis role="bold">Insist:</emphasis> Insist on
connecting to a server. In a configuration with multiple
load-sharing servers, the Insist option tells the server that
the client is insisting on a connection to the specified
server. Default is False.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Connect_timeout:</emphasis> The
timeout in seconds before the client gives up connecting to
the server. The default is no timeout.</para>
</listitem>
<listitem>
<para><emphasis role="bold">SSL:</emphasis> Use SSL to connect
to the server. The default is False.</para>
</listitem>
</itemizedlist>
<para>More precisely consumers need the following parameters:</para>
<itemizedlist>
<listitem>
<para><emphasis role="bold">Connection:</emphasis> The above
mentioned Connection object.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Queue:</emphasis> Name of the
queue.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Exchange:</emphasis> Name of the
exchange the queue binds to.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Routing_key:</emphasis> The
interpretation of the routing key depends on the value of the
exchange_type attribute.</para>
</listitem>
</itemizedlist>
<itemizedlist>
<listitem>
<para><emphasis role="bold">Direct exchange:</emphasis> If the
routing key property of the message and the routing_key
attribute of the queue are identical, then the message is
forwarded to the queue.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Fanout exchange:</emphasis> Messages
are forwarded to the queues bound the exchange, even if the
binding does not have a key.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Topic exchange:</emphasis> If the
routing key property of the message matches the routing key of
the key according to a primitive pattern matching scheme, then
the message is forwarded to the queue. The message routing key
then consists of words separated by dots (”.”, like domain
names), and two special characters are available; star (“”)
and hash (“#”). The star matches any word, and the hash
matches zero or more words. For example ”.stock.#” matches the
routing keys “usd.stock” and “eur.stock.db” but not
“stock.nasdaq”.</para>
</listitem>
</itemizedlist>
<itemizedlist>
<listitem>
<para><emphasis role="bold">Durable:</emphasis> This flag
determines the durability of both exchanges and queues;
durable exchanges and queues remain active when a RabbitMQ
server restarts. Non-durable exchanges/queues (transient
exchanges/queues) are purged when a server restarts. It is
worth noting that AMQP specifies that durable queues cannot
bind to transient exchanges. Default is True.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Auto_delete:</emphasis> If set, the
exchange is deleted when all queues have finished using it.
Default is False.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Exclusive:</emphasis> Exclusive
queues (such as non-shared) may only be consumed from by the
current connection. When exclusive is on, this also implies
auto_delete. Default is False.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Exchange_type:</emphasis> AMQP
defines several default exchange types (routing algorithms)
that covers most of the common messaging use cases.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Auto_ack:</emphasis> Acknowledgement
is handled automatically once messages are received. By
default auto_ack is set to False, and the receiver is required
to manually handle acknowledgment.</para>
</listitem>
<listitem>
<para><emphasis role="bold">No_ack:</emphasis> It disables
acknowledgement on the server-side. This is different from
auto_ack in that acknowledgement is turned off altogether.
This functionality increases performance but at the cost of
reliability. Messages can get lost if a client dies before it
can deliver them to the application.</para>
</listitem>
<listitem>
<para><emphasis role="bold">Auto_declare:</emphasis> If this is
True and the exchange name is set, the exchange will be
automatically declared at instantiation. Auto declare is on by
default. Publishers specify most the parameters of consumers
(they do not specify a queue name), but they can also
specify the following:</para>
</listitem>
<listitem>
<para><emphasis role="bold">Delivery_mode:</emphasis> The default
delivery mode used for messages. The value is an integer. The
following delivery modes are supported by RabbitMQ:</para>
</listitem>
</itemizedlist>
<itemizedlist>
<listitem>
<para><emphasis role="bold">1 or “transient”:</emphasis> The
message is transient. Which means it is stored in memory only,
and is lost if the server dies or restarts.</para>
</listitem>
<listitem>
<para><emphasis role="bold">2 or “persistent”:</emphasis> The
message is persistent. Which means the message is stored both
in-memory, and on disk, and therefore preserved if the server
dies or restarts.</para>
</listitem>
</itemizedlist>
<para>The default value is 2 (persistent). During a send operation,
Publishers can override the delivery mode of messages so that, for
example, transient messages can be sent over a durable
queue.</para>
</chapter>