Error handling and recovery

This Documentation Has Moved to rubyamqp.info

amqp gem documentation guides are now hosted on rubyamqp.info.

About this guide

Development of a robust application, be it message publisher or message consumer, involves dealing with multiple kinds of failures: protocol exceptions, network failures, broker failures and so on. Correct error handling and recovery is not easy. This guide explains how amqp gem helps you in dealing with issues like

  • Initial broker connection failures
  • Network connection interruption
  • AMQP connection-level exceptions
  • AMQP channel-level exceptions
  • Broker failure
  • TLS (SSL) related issues

as well as

  • How to recover after a network failure
  • What is automatic recovery mode, when you should and should not use it

This work is licensed under a Creative Commons Attribution 3.0 Unported License (including images & stylesheets). The source is available on Github.

Covered versions

This guide covers Ruby amqp gem v0.8.0.RC14 and later.

Code examples

There are several examples in the git repository dedicated to the topic of error handling and recovery. Feel free to contribute new examples.

Initial broker connection failures

When applications connect to the broker, they need to handle connection failures. Networks are not 100% reliable, even with modern system configuration tools like Chef or Puppet misconfigurations happen and broker might be down, too. Error detection should happen as early as possible. There are two ways of detecting TCP connection failure, the first one is to catch an exception:


begin
  AMQP.start(connection_settings) do |connection, open_ok|
    raise "This should not be reachable"
  end
rescue AMQP::TCPConnectionFailed => e
  puts "Caught AMQP::TCPConnectionFailed => TCP connection failed, as expected."
end

Full example:

AMQP.connect (and AMQP.start) will raise AMQP::TCPConnectionFailed if connection fails. Code that catches it can write to log about the issue or use retry to execute begin block one more time. Because initial connection failures are due to misconfiguration or network outage, reconnection to the same endpoint (hostname, port, vhost combination) will result in the same issue over and over. TBD: failover, connection to the cluster.

Alternative way of handling connection failure is with an errback (a callback for specific kind of error):


handler             = Proc.new { |settings| puts "Failed to connect, as expected"; EventMachine.stop }
connection_settings = {
  :port     => 9689,
  :vhost    => "/amq_client_testbed",
  :user     => "amq_client_gem",
  :password => "amq_client_gem_password",
  :timeout        => 0.3,
  :on_tcp_connection_failure => handler
}

Full example:

:on_tcp_connection_failure option accepts any object that responds to #call.

If you connect to the broker from a code in a class (as opposed to top-level scope in a script), Object#method can be used to pass object method as a handler instead of a Proc.

TBD: provide an example

Authentication failures

Another reason why connection may fail is authentication failure. Handling authentication failure is very similar to handling initial TCP connection failure:

Default handler

default handler raises AMQP::PossibleAuthenticationFailureError:

In case you wonder why callback name has “possible” in it: AMQP 0.9.1 spec requires broker implementations to simply close TCP connection without sending any more data when an exception (such as authentication failure) occurs before AMQP connection is open. In practice, however, when broker closes TCP connection between successful TCP connection and before AMQP connection is open, it means that authentication has failed.

Handling network connection interruptions

Network connectivity issues are sad fact of life in modern software systems. Event small products and projects these days consist of multiple applications, often running on more than one machine. Ruby amqp gem detects TCP connection failures and lets you handle them by defining a callback using AMQP::Session#on_tcp_connection_loss. That callback will be run when TCP connection fails, and will be passed two parameters: connection object and settings of the last successful connection.


connection.on_tcp_connection_loss do |connection, settings|
  # reconnect in 10 seconds, without enforcement
  connection.reconnect(false, 10)
end

Sometimes it is necessary for other entities in an application to react to network failures. amqp gem 0.8.0 and later provides a number event handlers to make this task easier for developers. This set of features is know as the “shutdown protocol” (the word “protocol” here means “API interface” or “behavior”, not network protocol).

AMQP::Session, AMQP::Channel, AMQP::Exchange, AMQP::Queue and all implement shutdown protocol and thus error handling API is consistent for all classes, with {AMQP::Session and #AMQP::Channel having a few methods other entities do not have.

The Shutdown protocol revolves around two events:

  • Network connection fails
  • Broker closes AMQP connection (or channel)

In this section, we concentrate only on the former. When network connection fails, the underlying networking library detects it and runs a piece of code on AMQP::Session to handle it. That, in turn, propagates this event to channels, channels propagate it to exchanges and queues, queues propagate it to their consumers (if any). Each of these entities in the object graph can react to network interruption by executing application-defined callbacks.

Shutdown Protocol methods on AMQP::Session

The difference between these methods is that AMQP::Session#on_tcp_connection_loss is used to define a callback that will be executed once when TCP connection fails. It is possible that reconnection attempts will not succeed immediately, so there will be subsequent failures. To react to those, AMQP::Session#on_connection_interruption method is used.

First argument that both of these methods yield to the handler your application defines is the connection itself. This is done to make sure you can register Ruby objects as handlers, and they do not have to keep any state around (for example, connection instances):


connection.on_connection_interruption do |conn|
  puts "Connection detected connection interruption"
end

# or

class ConnectionInterruptionHandler

  #
  # API
  #

  def handle(connection)
    # handling logic
  end

end

handler = ConnectionInterruptionHandler.new
connection.on_connection_interruption(&handler.method(:handle))

Note that AMQP::Session#on_connection_interruption callback is called before this event is propagated to channels, queues and so on.

Different applications handle connection failures differently. It is very common to use AMQP::Session#reconnect method to schedule a reconnection to the same host, or use AMQP::Session#reconnect_to to connect to a different one.

For some applications it is OK to simply exit and wait to be restarted at a later point in time, for example, by a process monitoring system like Nagios or Monit.

Shutdown Protocol methods on AMQP::Channel

AMQP::Channel provides only one method: AMQP::Channel#on_connection_interruption, that registers a callback similar to the one seen in the previous section:


channel.on_connection_interruption do |ch|
  puts "Channel #{ch.id} detected connection interruption"
end

Note that AMQP::Channel#on_connection_interruption callback is called after this event is propagated to exchanges, queues and so on. Right after that channel state is reset, except for error handling/recovery-related callbacks.

Many applications do not need per-channel network failure handling.

Shutdown Protocol methods on AMQP::Exchange

AMQP::Exchange provides only one method: AMQP::Exchange#on_connection_interruption, that registers a callback similar to the one seen in the previous section:


exchange.on_connection_interruption do |ex|
  puts "Exchange #{ex.name} detected connection interruption"
end

Many applications do not need per-exchange network failure handling.

Shutdown Protocol methods on AMQP::Queue

AMQP::Queue provides only one method: AMQP::Queue#on_connection_interruption, that registers a callback similar to the one seen in the previous section:


queue.on_connection_interruption do |q|
  puts "Queue #{q.name} detected connection interruption"
end

Note that AMQP::Queue#on_connection_interruption callback is called after this event is propagated to consumers.

Many applications do not need per-queue network failure handling.

Shutdown Protocol methods on AMQP::Consumer

AMQP::Consumer provides only one method: AMQP::Consumer#on_connection_interruption, that registers a callback similar to the one seen in the previous section:


consumer.on_connection_interruption do |c|
  puts "Consumer with consumer tag #{c.consumer_tag} detected connection interruption"
end

Many applications do not need per-consumer network failure handling.

Recovering from network connection failures

Detecting network connections is nearly useless if AMQP-based application cannot recover from them. Recovery is the hard part in “error handling and recovery”. Fortunately, recovery process for many applications follows one simple scheme that amqp gem can perform automatically for you.

Recovery process, both manual and automatic, always begins with re-opening AMQP connection and then all the channels on that connection.

Manual recovery

Similarly to the Shutdown Protocol, amqp gem entities implement Recovery Protocol. Recovery Protocol consists of 3 methods connections, channels, queues, consumers and exchanges implement:

AMQP::Session#before_recovery lets application developers register a callback that will be executed after TCP connection is re-established but before AMQP connection is reopened. AMQP::Session#after_recovery is similar except that the callback is run after AMQP connection is reopened.

AMQP::Channel, AMQP::Queue, AMQP::Consumer and AMQP::Exchange methods behavior is identical.

Recovery process for AMQP applications usually involves the following steps:

  1. Re-open AMQP connection.
  2. Once connection is open again, re-open all AMQP channels on that connection.
  3. For each channel, re-declare all exchanges
  4. For each channel, re-declare all queues
  5. Once queue is declared, for each queue, re-register all bindings
  6. Once queue is declared, for each queue, re-register all consumers

Automatic recovery

Many applications use the same recovery strategy, that consists of the following steps:

  • Re-open channels
  • For each channel, re-declare exchanges
  • For each channel, re-declare queues
  • For each queue, recover all bindings
  • For each queue, recover all consumers

amqp gem provides a feature known as “automatic recovery” that is opt-in (not opt-out, not used by default) and lets application developers get aforementioned recovery strategy by setting one additional attribute on AMQP::Channel instance:


ch = AMQP::Channel.new(connection)
ch.auto_recovery = true

A more verbose way to do the same thing:


ch = AMQP::Channel.new(connection, AMQP::Channel.next_channel_id, :auto_recovery => true)

Note that if you do not want to pass any options, 2nd argument can be left out as well, then it will default to AMQP::Channel.next_channel_id.

To find out whether channel uses automatic recovery mode, use AMQP::Channel#auto_recovering?.

Auto recovery mode can be turned on and off any number of times during channel life cycle, although very small percentage of applications really does this. Typically you decide what channels should be using automatic recovery at application design stage.

Full example (run it, then shut down AMQP broker running on localhost, then bring it back up and use management tools such as `rabbitmqctl` to see that queues & bindings & consumer have all recovered):

Server-named queues, when recovered automatically, will get new server-generated names to guarantee there are no name collisions.

When in doubt, try using automatic recovery first. If it is not sufficient for you application, switch to manual recovery using events and callbacks introduced in the “Manual recovery” section.

Detecting broker failures

AMQP applications see broker failure as TCP connection loss. There is no reliable way to know whether there is a network split or network peer is down.

AMQP connection-level exceptions

Handling connection-level exceptions

Connection-level exceptions are rare and may indicate a serious issue with client library or in-flight data corruption. They mandate that connection cannot be used any more and must be closed. In any case, your application should be prepared to handle this kind of errors. To define a handler, use AMQP::Session#on_error method that takes a callback and yields two arguments to it when connection-level exception happens:


connection.on_error do |conn, connection_close|
  puts "Handling a connection-level exception."
  puts
  puts "AMQP class id : #{connection_close.class_id}"
  puts "AMQP method id: #{connection_close.method_id}"
  puts "Status code   : #{connection_close.reply_code}"
  puts "Error message : #{connection_close.reply_text}"
end

Status codes are similar to those of HTTP. For the full list of error codes and their meaning, consult AMQP 0.9.1 constants reference.

Only one connection-level exception handler can be defined per connection instance (the one added last replaces previously added ones).

Full example:

Handling graceful broker shutdown

When AMQP broker is shut down, it properly closes connection first. To do so, it uses connection.close AMQP method. AMQP clients then need to check if the reply code is equal to 320 (CONNECTION_FORCED) to distinguish graceful shutdown. With RabbitMQ, when broker is stopped using

rabbitmqctl stop

reply_text will be set to

CONNECTION_FORCED - broker forced connection closure with reason 'shutdown'

Each application choose how to handle graceful broker shutdowns individually, so amqp gem’s automatic reconnection does not cover graceful broker shutdowns. Applications that want to reconnect when broker is stopped can use AMQP::Session#periodically_reconnect like so:


connection.on_error do |conn, connection_close|
  puts "[connection.close] Reply code = #{connection_close.reply_code}, reply text = #{connection_close.reply_text}"
  if connection_close.reply_code == 320
    puts "[connection.close] Setting up a periodic reconnection timer..."
    # every 30 seconds
    conn.periodically_reconnect(30)
  end
end

Once AMQP connection is re-opened, channels in automatic recovery mode will recover just like they do after network outages.

Integrating channel-level exceptions handling with object-oriented Ruby code

Error handling can be easily integrated into object-oriented Ruby code (in fact, this is highly encouraged). A common technique is to combine Object#method and Method#to_proc and use object methods as error handlers:


class ConnectionManager

  #
  # API
  #

  def connect(*args, &block)
    @connection = AMQP.connect(*args, &block)

    # combines Object#method and Method#to_proc to use object
    # method as a callback
    @connection.on_error(&method(:on_error))
  end # connect(*args, &block)


  def on_error(connection, connection_close)
    puts "Handling a connection-level exception."
    puts
    puts "AMQP class id : #{connection_close.class_id}"
    puts "AMQP method id: #{connection_close.method_id}"
    puts "Status code   : #{connection_close.reply_code}"
    puts "Error message : #{connection_close.reply_text}"
  end # on_error(connection, connection_close)
end

Full example that uses objects:

TBD

AMQP channel-level exceptions

Handling channel-level exceptions

Channel-level exceptions are more common than connection-level ones. They are handled in a similar manner, by defining a callback with AMQP::Channel#on_error method that takes a callback and yields two arguments to it when channel-level exception happens:


channel.on_error do |ch, channel_close|
  puts "Handling a channel-level exception."
  puts
  puts "AMQP class id : #{channel_close.class_id}"
  puts "AMQP method id: #{channel_close.method_id}"
  puts "Status code   : #{channel_close.reply_code}"
  puts "Error message : #{channel_close.reply_text}"
end

Status codes are similar to those of HTTP. For the full list of error codes and their meaning, consult AMQP 0.9.1 constants reference.

Only one channel-level exception handler can be defined per channel instance (the one added last replaces previously added ones).

Full example:

Integrating channel-level exceptions handling with object-oriented Ruby code

Error handling can be easily integrated into object-oriented Ruby code (in fact, this is highly encouraged). A common technique is to combine Object#method and Method#to_proc and use object methods as error handlers. For example of this, see section on connection-level exceptions above.

Because channel-level exceptions may be raised because of multiple unrelated reasons and often indicate misconfigurations, how they are handled is very specific to particular applications. A common strategy is to log an error and then open and use another channel.

Common channel-level exceptions and what they mean

A few channel-level exceptions are common and deserve more attention.

406 Precondition Failed

Description
The client requested a method that was not allowed because some precondition failed.
What might cause it
  • AMQP entity (a queue or exchange) was re-declared with attributes different from original declaration. Maybe two applications or pieces of code declare the same entity with different attributes. Note that different AMQP client libraries historically use slightly different defaults for entities and this may cause attribute mismatches.
  • AMQP::Channel#tx_commit or AMQP::Channel#tx_rollback might be run on a channel that wasn’t previously made transactional with AMQP::Channel#tx_select
Example RabbitMQ error message
  • PRECONDITION_FAILED – parameters for queue ‘amqpgem.examples.channel_exception’ in vhost ‘/’ not equivalent
  • PRECONDITION_FAILED – channel is not transactional

405 Resource Locked

Description
The client attempted to work with a server entity to which it has no access because another client is working with it.
What might cause it
  • Multiple applications (or different pieces of code/threads/processes/routines within a single application) might try to declare queues with the same name as exclusive.
  • Multiple consumer across multiple or single app might be registered as exclusive for the same queue.
Example RabbitMQ error message
RESOURCE_LOCKED – cannot obtain exclusive access to locked queue ‘amqpgem.examples.queue’ in vhost ‘/’

403 Access Refused

Description
The client attempted to work with a server entity to which it has no access due to security settings.
What might cause it
Application tries to access a queue or exchange it has no permissions for (or right kind of permissions, for example, write permissions)
Example RabbitMQ error message
ACCESS_REFUSED – access to queue ‘amqpgem.examples.channel_exception’ in vhost ‘amqp_gem_testbed’ refused for user ‘amqp_gem_reader’

TLS (SSL) related issues

TBD

Conclusion

Distributed applications introduce a whole new class of failures developers need to be aware of. Many of them come from unreliability of the network. The famous Fallacies of Distributed Computing list common assumptions software engineers must not make:

  • The network is reliable.
  • Latency is zero.
  • Bandwidth is infinite.
  • The network is secure.
  • Topology doesn’t change.
  • There is one administrator.
  • Transport cost is zero.
  • The network is homogeneous.

Unfortunately, applications that use Ruby and AMQP are not immune to these problems and developers need to always keep that in mind. This list is just as relevant in 2011 as it was in the 90s.

Ruby amqp gem 0.8.x and later lets applications to define handlers that handle connection-level exceptions, channel-level exceptions and TCP connection failures. Handling AMQP exceptions and network connection failures is relatively easy. Re-declaring AMQP instances application works with is where the most of complexity comes from. By using Ruby objects as error handlers, declaration of AMQP entities can be done in one place, making it much easier to understand and maintain.

amqp gem error handling and interruption is not a copy of RabbitMQ Java client’s Shutdown Protocol, but they turn out to be similar with respect to network failures and connection-level exceptions.

TBD

Authors

This guide was written by Michael Klishin and edited by Chris Duncan.

Tell us what you think!

Please take a moment and tell us what you think about this guide on Twitter or Ruby AMQP mailing list: What was unclear? What wasn’t covered? Maybe you don’t like guide style or grammar and spelling are incorrect? Readers feedback is key to making documentation better.

If mailing list communication is not an option for you for some reason, you can contact guides author directly