Bunny::AuthenticationFailureError is a new auth failure exception
that subclasses Bunny::PossibleAuthenticationFailureError for
backwards compatibility.
As such, Bunny::PossibleAuthenticationFailureError's error message
has changed.
This extension is available in RabbitMQ 3.2+.
Bunny::Session#exchange_exists? is a new predicate that makes it
easier to check if a exchange exists.
It uses a one-off channel and exchange.declare with passive set to true
under the hood.
Bunny::Session#queue_exists? is a new predicate that makes it
easier to check if a queue exists.
It uses a one-off channel and queue.declare with passive set to true
under the hood.
It is now possible to provide inline client certificate and private key (as strings) instead of filesystem paths. The options are the same:
:tlswhich, when set totrue, will set SSL context up and switch to TLS port (5671):tls_certwhich now can be a client certificate (public key) in PEM format:tls_keywhich now can be a client key (private key) in PEM format:tls_ca_certificateswhich is an array of string paths to CA certificates in PEM format
For example:
conn = Bunny.new(:tls => true,
:tls_cert => ENV["TLS_CERTIFICATE"],
:tls_key => ENV["TLS_PRIVATE_KEY"],
:tls_ca_certificates => ["./examples/tls/cacert.pem"])Ruby 1.8.7 compatibility fixes around timeouts.
Minimum amq-protocol version is now 1.8.0 which includes
a bug fix for messages exactly 128 Kb in size.
Bunny::ConsumerWorkPool#join now accepts an optional
timeout argument.
RABBITMQ_URL env variable will now have effect even if
Bunny.new is invoked without arguments.
Bunny is Ruby 1.8-compatible again and no longer references
RUBY_ENGINE.
Bunny::Session.parse_uri is a new method that parses
connection URIs into hashes that Bunny::Session#initialize
accepts.
Bunny::Session.parse_uri("amqp://user:pwd@broker.eng.megacorp.local/myapp_qa")Bunny now uses OpenSSL to detect default TLS/SSL CA's paths, extending this feature to OS'es other than Linux.
Contributed by Jingwen Owen Ou.
Bunny now will use the following TLS/SSL CA's paths on Linux by default:
/etc/ssl/certs/ca-certificates.crton Ubuntu/Debian/etc/ssl/certs/ca-bundle.crton Amazon Linux/etc/ssl/ca-bundle.pemon OpenSUSE/etc/pki/tls/certs/ca-bundle.crton Fedora/RHEL
and will log a warning if no CA files are available via default paths
or :tls_ca_certificates.
Contributed by Carl Hörberg.
It is now possible to re-register a consumer (and use any other synchronous methods)
from Bunny::Consumer#handle_cancellation, which is now invoked in the channel's
thread pool.
Bunny::Session#close with single threaded connections no longer fails
with a nil pointer exception.
This release has breaking API changes.
Previously if a channel was recovered (reopened) by automatic connection recovery before a message was acknowledged or rejected, it would cause any operation on the channel that uses delivery tags to fail and cause the channel to be closed.
To avoid this issue, every channel keeps a counter of how many times it has been reopened and marks delivery tags with them. Using a stale tag to ack or reject a message will produce no method sent to RabbitMQ. Note that unacknowledged messages will be requeued by RabbitMQ when connection goes down anyway.
This involves an API change: Bunny::DeliveryMetadata#delivery_tag is now
and instance of a class that responds to #tag and #to_i and is accepted
by Bunny::Channel#ack and related methods.
Integers are still accepted by the same methods.
When a second consumer is registered for the same queue on different channels,
a reasonable exception (Bunny::AccessRefused) will be raised.
Bunny now allows mutex impl to be configurable, uses reentrant Monitor by default.
Non-reentrant mutexes is a major PITA and may affect code that uses Bunny.
Avg. publishing throughput with Monitor drops slightly from 5.73 Khz to 5.49 Khz (about 4% decrease), which is reasonable for Bunny.
Apps that need these 4% can configure what mutex implementation is used on per-connection basis.
Bunny::Session#close had a race condition that caused (non-deterministic)
exceptions when connection transport was closed before connection
reader loop was guaranteed to have stopped.
Connection-level exceptions (including when a connection is closed via
management UI or rabbitmqctl) will now be raised on the connection
thread so they
- can be handled by applications
- do not start connection recovery, which may be uncalled for
Bunny will no longer require client TLS certificates. Note that CA certificate list is still necessary.
If RabbitMQ TLS configuration requires peer verification, client certificate and private key are mandatory.
Publishing a message over a closed connection (during a network outage, before the connection is open) will now correctly result in an exception.
Contributed by Matt Campbell.
Bunny now ensures a new connection transport (socket) is initialized before any recovery is attempted.
Bunny::Session#create_channel now uses two separate mutexes to avoid
a (very rare) issue when the previous implementation would try to
re-acquire the same mutex and fail (Ruby mutexes are non-reentrant).
In a case when all consumers are cancelled, Bunny::Channel
will shut down its consumer delivery thread pool.
It will also now mark the pool as not running so that it can be started again successfully if new consumers are registered later.
GH issue: #133.
A little bit of background: on MRI, the method raised ThreadErrors
reliably. On JRuby, we used a different [internal] queue implementation
from JDK so it wasn't an issue.
Timeout.timeout uses Thread#kill and Thread#join, both of which
eventually attempt to acquire a mutex used by Queue#pop, which Bunny
currently uses for continuations. The mutex is already has an owner
and so a ThreadError is raised.
This is not a problem on JRuby because there we don't use Ruby's Timeout and Queue and instead rely on a JDK concurrency primitive which provides "poll with a timeout".
The issue with Thread#kill and Thread#raise
has been first investigated and blogged about by Ruby implementers
in 2008.
Finding a workaround will probably take a bit of time and may involve reimplementing standard library and core classes.
We don't want this issue to block Bunny 0.9 release. Neither we want to ship a broken feature. So as a result, we will drop Bunny::Queue#pop_waiting since it cannot be reliably implemented in a reasonable amount of time on MRI.
Per issue #131.
Bunny will now upgrade connection to SSL in Bunny::Session#start,
so it is possible to fine tune SSLContext and socket settings
before that:
require "bunny"
conn = Bunny.new(:tls => true,
:tls_cert => "examples/tls/client_cert.pem",
:tls_key => "examples/tls/client_key.pem",
:tls_ca_certificates => ["./examples/tls/cacert.pem"])
puts conn.transport.socket.inspect
puts conn.transport.tls_context.inspectThis also means that Bunny.new will now open the socket. Previously
it was only done when Bunny::Session#start was invoked.
Bunny 0.9 finally supports TLS. There are 3 new options Bunny.new takes:
:tlswhich, when set totrue, will set SSL context up and switch to TLS port (5671):tls_certwhich is a string path to the client certificate (public key) in PEM format:tls_keywhich is a string path to the client key (private key) in PEM format:tls_ca_certificateswhich is an array of string paths to CA certificates in PEM format
An example:
conn = Bunny.new(:tls => true,
:tls_cert => "examples/tls/client_cert.pem",
:tls_key => "examples/tls/client_key.pem",
:tls_ca_certificates => ["./examples/tls/cacert.pem"])This function was removed in v0.9.0.rc2
Bunny::Queue#pop_waiting is a new function that mimics Bunny::Queue#pop
but will wait until a message is available. It uses a :timeout option and will
raise an exception if the timeout is hit:
# given 1 message in the queue,
# works exactly as Bunny::Queue#get
q.pop_waiting
# given no messages in the queue, will wait for up to 0.5 seconds
# for a message to become available. Raises an exception if the timeout
# is hit
q.pop_waiting(:timeout => 0.5)This method only makes sense for collecting Request/Reply ("RPC") replies.
Bunny::InvalidCommand is now Bunny::CommandInvalid (follows
the exception class naming convention based on response status
name).
Channels without consumers left (when all consumers were cancelled)
will now tear down their consumer work thread pools, thus making
HotBunnies::Queue#subscribe(:block => true) calls unblock.
This is typically the desired behavior.
Delivery handlers registered via Bunny::Queue#subscribe now will have
access to the consumer and channel they are associated with via the
delivery_info argument:
q.subscribe do |delivery_info, properties, payload|
delivery_info.consumer # => the consumer this delivery is for
delivery_info.consumer # => the channel this delivery is on
endThis allows using Bunny::Queue#subscribe for one-off consumers
much easier, including when used with the :block option.
Bunny::Exchange#wait_for_confirms is a convenience method on Bunny::Exchange that
delegates to the method with the same name on exchange's channel.
Bunny::Socket no longer uses Ruby 1.9-specific constants.
Bunny::Channel#wait_for_confirms returns true or false again.
Bunny::Session#create_channel now accepts consumer work pool size as
the second argument:
# nil means channel id will be allocated by Bunny.
# 8 is the number of threads in the consumer work pool this channel will use.
ch = conn.create_channel(nil, 8)Long running consumers that don't send any data will no longer suffer from connections closed by RabbitMQ because of skipped heartbeats.
Activity tracking now takes sent frames into account.
If a network loop exception causes "main" session thread to never
receive a response, methods such as Bunny::Channel#queue will simply time out
and raise Timeout::Error now, which can be handled.
It will not start automatic recovery for two reasons:
- It will be started in the network activity loop anyway
- It may do more damage than good
Kicking off network recovery manually is a matter of calling
Bunny::Session#handle_network_failure.
The main benefit of this implementation is that it will never block the main app/session thread forever, and it is really efficient on JRuby thanks to a j.u.c. blocking queue.
Fixes #112.
Every Bunny connection now has a logger. By default, Bunny will use STDOUT
as logging device. This is configurable using the :log_file option:
require "bunny"
conn = Bunny.new(:log_level => :warn)or the BUNNY_LOG_LEVEL environment variable that can take one of the following
values:
debug(very verbose)infowarnerrorfatal(least verbose)
Severity is set to warn by default. To disable logging completely, set the level
to fatal.
To redirect logging to a file or any other object that can act as an I/O entity,
pass it to the :log_file option.
This release contains a breaking API change.
On JRuby, Bunny now will use java.util.concurrent-backed implementations
of some of the concurrency primitives. This both improves client stability
(JDK concurrency primitives has been around for 9 years and have
well-defined, documented semantics) and opens the door to solving
some tricky failure handling problems in the future.
Bunny now will correctly close the socket previous connection had when recovering from network issues.
Bunny::Exception now inherits from StandardError and not Exception.
Naked rescue like this
begin
# ...
rescue => e
# ...
endcatches only descendents of StandardError. Most people don't
know this and this is a very counter-intuitive practice, but
apparently there is code out there that can't be changed that
depends on this behavior.
This is a breaking API change.
Bunny::Session#start now returns a session instead of the default channel
(which wasn't intentional, default channel is a backwards-compatibility implementation
detail).
Bunny::Session#start also no longer leaves dead threads behind if called multiple
times on the same connection.
Heartbeat sender no longer slips into an infinite loop if it encounters an exception. Instead, it will just stop (and presumably re-started when the network error recovery kicks in or the app reconnects manually).
Network reconnection now kicks in after a delay to avoid aggressive reconnections in situations when we don't want to endlessly reconnect (e.g. when the connection was closed via the Management UI).
The :network_recovery_interval option passed to Bunny::Session#initialize and Bunny.new
controls the interval. Default is 5 seconds.
Bunny will now use heartbeat value provided by RabbitMQ by default.
Several stability improvements in the network layer, connection error handling, and concurrency hazards.
Automatic connection recovery now can be disabled by passing
the :automatically_recover => false option to Bunny#initialize).
When the recovery is disabled, network I/O-related exceptions will cause an exception to be raised in thee thread the connection was started on.
Bunny::Exchange#publish and Bunny::Channel#basic_publish no
longer perform timeout control (using the timeout module) which
roughly increases throughput for flood publishing by 350%.
Apps that need delivery guarantees should use publisher confirms.
Bunny::Channel#on_error is a new method that lets you define
handlers for channel errors that are caused by methods that have no
responses in the protocol (basic.ack, basic.reject, and basic.nack).
This is rarely necessary but helps make sure no error goes unnoticed.
Example:
channel.on_error |ch, channel_close|
puts channel_close.inspect
endLarger (over 128K) messages with non-ASCII characters are now always encoded
correctly with amq-protocol 1.2.0.
Publishing of large messages is now done more efficiently.
Contributed by Greg Brockman.
Bunny API reference is now up online.
Bunny::Channel#basic_publish now supports both
:delivery_mode and :persistent options.
Bunny::Channel#nacked_set is a counter-part to Bunny::Channel#unacked_set
that contains basic.nack-ed (rejected) delivery tags.
Passing :threaded => false to Bunny.new now will use the same
thread for publisher confirmations (may be useful for retry logic
implementation).
Contributed by Greg Brockman.
Automatic Network Failure Recovery is a new Bunny feature that was earlier impemented and vetted out in amqp gem. What it does is, when a network activity loop detects an issue, it will try to periodically recover [first TCP, then] AMQP 0.9.1 connection, reopen all channels, recover all exchanges, queues, bindings and consumers on those channels (to be clear: this only includes entities and consumers added via Bunny).
Publishers and consumers will continue operating shortly after the network connection recovers.
Learn more in the Error Handling and Recovery documentation guide.
Bunny now supports listeners (callbacks) on
ch.confirm_select do |delivery_tag, multiple, nack|
# handle confirms (e.g. perform retries) here
endContributed by Greg Brockman.
Publisher confirms implementation now uses non-strict equality (<=) for
cases when multiple messages are confirmed by RabbitMQ at once.
Bunny::Channel#unconfirmed_set is now part of the public API that lets
developers access unconfirmed delivery tags to perform retries and such.
Contributed by Greg Brockman.
Bunny::Channel#wait_for_confirms will now correctly block the calling
thread until all pending confirms are received.
Channel error information is now properly reset when a channel is (re)opened.
GH issue: #83.
the default value of Bunny::Consumer noack argument changed from false to true
for consistency.
Global prefetch is not implemented in RabbitMQ, so Bunny::Session#prefetch
is gone from the API.
Fixed a problem when a queue was not declared after being deleted and redeclared
GH issue: #80
Channel queue and exchange caches are now properly invalidated when queues and exchanges are deleted.
Heartbeats are now correctly sent at safe intervals (half of the configured
interval). In addition, setting :heartbeat => 0 (or nil) will disable
heartbeats, just like in Bunny 0.8 and amqp gem.
Default :heartbeat value is now 600 (seconds), the same as RabbitMQ 3.0
default.
Fixes a potential race condition between basic.consume-ok handler and
delivery handler when a consumer is registered for a queue that has
messages in it.
GH issue: #78.
Bunny now supports two authentication mechanisms and can be extended
to support more. The supported methods are "PLAIN" (username
and password) and "EXTERNAL" (typically uses TLS, UNIX sockets or
another mechanism that does not rely on username/challenge pairs).
To use the "EXTERNAL" method, pass :auth_mechanism => "EXTERNAL" to
Bunny.new:
# uses the EXTERNAL authentication mechanism
conn = Bunny.new(:auth_method => "EXTERNAL")
conn.startA new high-level API method: Bunny::Consumer#cancel, can be used to
cancel a consumer. Bunny::Queue#subscribe will now return consumer
instances when the :block option is passed in as false.
Bunny::Exchange#delete will no longer delete pre-declared exchanges
that cannot be declared by Bunny (amq.* and the default exchange).
Bunny::DeliveryInfo#redelivered? is a new method that is an alias
to Bunny::DeliveryInfo#redelivered but follows the Ruby community convention
about predicate method names.
Bunny::DeliveryInfo#delivery_tag had a typo which is now fixed.
Bunny now correctly lists RabbitMQ extensions it currently supports in client capabilities:
basic.nack- exchange-to-exchange bindings
- consumer cancellation notifications
- publisher confirms
Lightweight Publisher Confirms is a RabbitMQ feature that lets publishers keep track of message routing without adding noticeable throughput degradation as it is the case with AMQP 0.9.1 transactions.
Bunny 0.9.0.pre3 supports publisher confirms. Publisher confirms are enabled per channel,
using the Bunny::Channel#confirm_select method. Bunny::Channel#wait_for_confirms is a method
that blocks current thread until the client gets confirmations for all unconfirmed published
messages:
ch = connection.create_channel
ch.confirm_select
ch.using_publisher_confirmations? # => true
q = ch.queue("", :exclusive => true)
x = ch.default_exchange
5000.times do
x.publish("xyzzy", :routing_key => q.name)
end
ch.next_publish_seq_no.should == 5001
ch.wait_for_confirms # waits until all 5000 published messages are acknowledged by RabbitMQIt is now possible to register a consumer as an object instead of a block. Consumers that are class instances support cancellation notifications (e.g. when a queue they're registered with is deleted).
To support this, Bunny introduces two new methods: Bunny::Channel#basic_consume_with
and Bunny::Queue#subscribe_with, that operate on consumer objects. Objects are
supposed to respond to three selectors:
:handle_deliverywith 3 arguments:handle_cancellationwith 1 argument:consumer_tag=with 1 argument
An example:
class ExampleConsumer < Bunny::Consumer
def cancelled?
@cancelled
end
def handle_cancellation(_)
@cancelled = true
end
end
# "high-level" API
ch1 = connection.create_channel
q1 = ch1.queue("", :auto_delete => true)
consumer = ExampleConsumer.new(ch1, q)
q1.subscribe_with(consumer)
# "low-level" API
ch2 = connection.create_channel
q1 = ch2.queue("", :auto_delete => true)
consumer = ExampleConsumer.new(ch2, q)
ch2.basic_consume_with.(consumer)If RABBITMQ_URL environment variable is set, Bunny will assume
it contains a valid amqp URI string and will use it. This is convenient
with some PaaS technologies such as Heroku.
It makes more sense for beginners that way.
Bunny::Queue#subscribe support the new :block option
(a boolean).
It controls whether the current thread will be blocked
by Bunny::Queue#subscribe.
Bunny::Exchange#publish now supports :key as an alias for
:routing_key.
Bunny::Session#queue, Bunny::Session#direct, Bunny::Session#fanout, Bunny::Session#topic,
and Bunny::Session#headers were added to simplify migration. They all delegate to their respective
Bunny::Channel methods on the default channel every connection has.
Bunny::Channel#exchange and Bunny::Session#exchange were added to simplify
migration:
b = Bunny.new
b.start
# uses default connection channel
x = b.exchange("logs.events", :topic)q.subscribe(:exclusive => false, :ack => false) do |delivery_info, properties, payload|
# ...
endBunny::Channel#fanout, Bunny::Channel#topic, Bunny::Channel#direct, Bunny::Channel#headers,
andBunny::Channel#default_exchange are new convenience methods to instantiate exchanges:
conn = Bunny.new
conn.start
ch = conn.create_channel
x = ch.fanout("logging.events", :durable => true)Bunny < 0.9.x example:
h = queue.pop
puts h[:delivery_info], h[:header], h[:payload]Bunny >= 0.9.x example:
delivery_info, properties, payload = queue.popThe improve is both in that Ruby has positional destructuring, e.g.
delivery_info, _, content = q.popbut not hash destructuring, like, say, Clojure does.
In addition we return nil for content when it should be nil (basic.get-empty) and unify these arguments betwee
-
Bunny::Queue#pop
-
Consumer (Bunny::Queue#subscribe, etc) handlers
-
Returned message handlers
The unification moment was the driving factor.
Bunny::Client#write now raises Bunny::ConnectionError instead of Bunny::ServerDownError when network
I/O operations fail.
Instead of reusing channel instances, Bunny::Client.create_channel now opens new channels and
uses bitset-based allocator to keep track of used channel ids. This avoids situations when
channels are reused or shared without developer's explicit intent but also work well for
long running applications that aggressively open and release channels.
This is also how amqp gem and RabbitMQ Java client manage channel ids.
Bunny::ServerDownError is now an alias for Bunny::TCPConnectionFailed