No changes yet.
Bunny::Channel#queue will now throw an ArgumentError if a nil
is passed for queue name.
GitHub issue: #570
During connection recovery, if recover_attempts is not set (is nil)
connection could produce confusing log messages.
GitHub issue: #569
Responses for queue.declare are now checked against a memoized
queue name (but only if the queue is not server-named). This helps
avoids scenarios with overlapping/concurrent requests due to high
network latency as demonstrated in #558.
"Mismatched" responses will be ignored: Bunny channel API would throw an exception for such declarations and there would be no way to "return to" even if a matching response arrived and was matched with one of the pending requests in a reasonable period of time.
As part of this work a new Toxiproxy-based test suite was introduced to Bunny.
GitHub issue: #558
Reproduction steps contributed by Brian Morton and Scott Bonebraker.
Heartbeat sender's purpose is to notify the peer, not so much to detect local connectivity failures; those will be detected by the I/O loop and transport.
For single threaded connection users that prefer to roll their own recovery strategies getting exceptions from the heartbeat sender was counterproductive and painful to deal with.
As part of this work a new Toxiproxy-based test suite was introduced to Bunny.
GitHub issue: #559
Contributed by Scott Bonebraker.
Connections that experienced connection closure did not always correctly transition to the closed state.
Bunny::ConnectionClosedError will now be thrown when an operation is attempted on such
connections.
GitHub issue: #561
Contributed by Scott Bonebraker.
GitHub issue: #549
Contributed by Arlandis Word.
Many warnings have been eliminated.
GitHub issue: #563
Contributed by @dacto.
GitHub issue: #557
Contributed by Bruno Costa.
Bunny no longer tries to compile a list of trusted CA certificates on its own. Instead it uses an OpenSSL API method that makes OpenSSL set the path(s), which should cover more platforms and be forward- and backward-compatible.
GitHub issue: #555.
Contributed by Ana María Martínez Gómez.
2.10.0 is a maintenance release that introduces a couple of
minor potentially breaking changes.
Disabling heartbeats will now disable TCP socket read timeouts.
They go hand in hand and users who prefer TCP keepalives via kernel configuration previously had to also explicitly configure a zero read timeout.
GitHub issue: #551.
Contributed by Carl Hörberg.
Make sure verify_peer: false has the expected effect again.
Default value of connection's :verify_peer option to true only when
all of :verify_ssl, :verify_peer, and :verify are nil.
GitHub issue: #541.
Contributed by Howard Ding.
Default maximum number of channels is limited to 2047 to reduce the probability of severe channel leaks. See rabbitmq/rabbitmq-server#1593 for details.
Applications that want to use more channels per connection can still configure a higher value
using the channel_max setting (for both Bunny and RabbitMQ server).
GitHub issue: #553.
GitHub issue: #552.
Contributed by @utilum.
Disabling heartbeats will also disable TCP socket read timeouts, since the two are effectively interconnected. In this case a mechanism such as TCP keepalives is assumed to be used.
See RabbitMQ heartbeats guide for a more detailed overview of the options.
GH issue: #519.
Contributed by Carl Hörberg.
Bunny now requires Ruby 2.2.
Connection recovery now will retry on TCP connection timeouts.
GitHub issue: #537.
Bunny now supports more URI query parameters plus aliases that are identical to those of the server.
Contributed by Andrew Babichev.
GitHub issue: #534
This release has minor breaking public API changes.
Bunny::Channel#close on an already closed channel will now raise a sensible exception.
If the channel was closed due to a channel-level protocol exception, that exception will
be mentioned.
GitHub issue: #528, see 9df7cb for details.
A JRuby 9K compatibility issue was corrected by Marian Posăceanu. Note that JRuby users are recommended to use March Hare, a JRuby-oriented client, instead of Bunny.
GitHub issue: #529
When automatic recovery is enabled, connection errors are now logged as warnings and not errors.
Contributed by Merten Falk.
GitHub issue: #531
It is now possible to specify a server-defined heartbeat value as a string ("server"), not just
a symbol. This makes it easier to load settings from YAML files.
Contributed by Tyrone Wilson.
GitHub issue: #524
Bunny now correctly handles scenarios where server is configured to disable heartbeats (which is a terrible idea, don't do it!)
GitHub issue: #519.
Bunny::Channel#basic_get invoked with a non-existent queue now
throws a channel exception instead of a generic operation timeout.
GitHub issue: #518.
BUNNY_CERTIFICATE_DIR environment variable now can be used
to override local CA and client certificate/key pair directory.
The directory is expected to be the result directory generated
by the basic tls-gen profile.
TLSv1.0 is no longer used in tests because it's being disabled by default by more and more installations as it has known vulnerabilities and is no longer considered to be acceptable by several compliance standards (e.g. PCI DSS).
channel.close handler will now acquire a lock . This avoids concurrency
hazards in some rare scenarios when a channel is closed due a protocol
exception by the server and concurrently opened by user code
at the same time.
Sometimes users attempt to open a channel on a connection that
isn't connected yet because Bunny::Session#start was never invoked.
Bunny::Session#create_channel will now provide a more sensible exception message
in those cases.
Minimum amq-protocol version is now 2.2.0 which includes
a change in how timestamps are encoded.
Bunny::ContinuationQueue#poll was reworked with feedback from Joseph Wong.
GitHub issue: #462
Previous behehavior is not unreasonable but is not what many users and even RabbitMQ team members come to expect. Therefore it can be considered a bug.
Previously a reconnection counter was preserved between successful recoveries. This made the integration test that uses server-sent connection.close possible.
With this change, the counter is reset after successful reconnection but there's an option to go back to the original behavior. We also do a hell of a lot more logging.
GitHub issue: #408
Contributed by Jared Smartt.
GitHub issue: #492.
Contributed by Alessandro Verlato.
GitHub issue: #465
Bunny::ConsumerWorkPool#shutdown(true) waited for consumer shutdown
even if the pool wasn't active (there were no consumers on its
channel).
GitHub issue: #438.
GitHub issue: #456
Bunny now can work with frozen host arrays.
GitHub issue: #446
Consumer work pool will now allow for a grace period before stopping pool threads so that delivery processing in progress can have a chance to finish.
GitHub issue: #437
Contributed by Stefan Sedich.
GitHub issue: #428
Contributed by Dimitar Dimitrov.
GitHub issue: #424
Contributed by Dimitar Dimitrov.
Bunny::ConsumerWorkPool#join and Bunny::ConsumerWorkPool#pause
no longer fails with a NoMethodError on nil when executed
on a work pool that doesn't have active threads (consumers).
This change is largely cosmetic and won't affect the majority of of projects in any way.
GitHub issue: #410
Contributed by Andrew Bruce.
GitHub issue: #424
Contributed by Dimitar Dimitrov.
Queue recovery now respects the :no_declare option.
Bunny::Channel#wait_for_confirms now throws an exception
early when invoked on a closed channel.
GitHub issue: #428.
Contributed by Dimitar Dimitrov.
This release includes minor breaking API changes.
Channels will now reset their unconfirmed delivery tag set after recovery.
GitHub issue: #406
Contributed by Bill Ruddock.
GitHub issue: #383.
Contributed by Jeremy Heiler.
Since transport is replaced by a recovering connection anyway, and this produces confusing errors up the stack.
GitHub issue: #359
Contributed by Donal McBreen.
Client-Provided Session :properties will now be merged with defaults
instead of replacing them. This makes it much more convenient to
override a single key.
This is a breaking API change.
RABBITMQ_URL no longer will be used if any other
connection options are provided. This makes it possible
to use RABBITMQ_URL for some connections and options
for others in a single OS process.
GitHub issue: #403
Contributed by Jimmy Petersen.
Bunny::Session#create_channel now supports a 3rd argument that,
when set to true, makes consumer work pool threads to have
Thread#abort_on_exception set on them.
GH issue: #382
Contributed by Seamus Abshere.
Bunny now will explicitly close previosly used transport before starting connection recovery.
GitHub issue: #377.
Contributed by bkanhoopla.
Makes sure that TLS sockets are not double-initialized.
GH issue: #345.
Contributed by Carl Hörberg.
GH issue: #375
Contributed by Omer Katz.
Before this the connection options only allowed multiple hosts, an address is a combination of a host and a port. This makes it possible to specify different hosts with different ports.
Contributed by Bart van Zon (Tele2).
Bunny will now try to reconnect also when server sent connection.close is
received, e.g. when a server is restarting (but also when the connection is
force closed by the server). This is in-line with how many other clients behave.
The old default was recover_from_connection_close: false.
Contributed by Carl Hörberg (CloudAMQP).
Bunny 2.1.0 has an important breaking change. It is highly advised that 2.1.0 is not mixed with earlier versions of Bunny in case your applications include integers in message headers.
Integer values in headers are now serialised as signed 64-bit integers. Previously they were serialised as 32-bit unsigned integers, causing both underflows and overflows: incorrect values were observed by consumers.
It is highly advised that 2.1.0 is not mixed with earlier versions of Bunny in case your applications include integers in message headers.
If that's not the case, Bunny 2.1 will integeroperate with any earlier version starting with 0.9.0 just fine. Popular clients in other languages (e.g. Java and .NET) will interoperate with Bunny 2.1.0 without issues.
Bunny now requires Ruby 2.0 in the gemspec.
Contributed by Carl Hörberg.
Bunny runs again on JRuby. Note that JRuby users are strongly advised to use March Hare instead.
Contributed by Teodor Pripoae.
Bunny 2.0 doesn't have any breaking API changes
but drops Ruby 1.8 and 1.9 (both EOL'ed) support,
hence the version.
Bunny 2.0 requires Ruby 2.0 or later.
Bunny now uses non-blocking socket writes, uses a reduced number of writes for message publishing (frames are batched into a single write), and handles TCP back pressure from RabbitMQ better.
Contributed by Irina Bednova and Michael Klishin.
Bunny::ContinuationQueue#poll no longer relies on Ruby's Timeout which has
numerous issues, including starting a new "interruptor" thread per operation,
which is far from efficient.
Contributed by Joe Eli McIlvain and Carl Hörberg.
:recovery_attempts is a new option that limits the number of
connection recovery attempts performed by Bunny. nil means
"no limit".
Contributed by Irina Bednova.
Bunny::Channel#basic_ack, Bunny::Channel#basic_nack, and Bunny::Channel#basic_reject
now adjust delivery tags between connection recoveries, as well as have a default value for
the second argument.
Contributed by Wayne Conrad.
Setting the @logger.progname attribute changes the output of the logger.
This is not expected behaviour when the client provides a custom logger.
Behaviour remains unchainged when the internally initialized logger is used.
Contributed by Justin Carter.
Since basic.qos's prefetch_count field is of type short in the protocol,
Bunny must enforce its maximum allowed value to 2^16 - 1 to avoid
confusing issues due to overflow.
Recent RabbitMQ versions support basic.qos global flag, controlling whether
prefetch applies per-consumer or per-channel. Bunny Channel#prefetch now
allows flag to be set as optional parameter, with the same default behaviour as
before (per-consumer).
Contributed by tiredpixel.
When using TLS, peer verification is now enabled by default. It is still possible to disable verification, e.g. for convenient development locally.
Peer verification is a means of protection against man-in-the-middle attacks and is highly recommended in production settings. However, it can be an inconvenience during local development. We believe it's time to have the default to be more secure.
Contributed by Michael Klishin (Pivotal) and Andre Foeken (Nedap).
Default connection timeout has been increased to 25 seconds. The older default of 5 seconds wasn't sufficient in some edge cases with DNS resolution (e.g. when primary DNS server is down).
The value can be overriden at connection time.
Contributed by Yury Batenko.
GH issue: #267.
On JRuby, Bunny reverts back to using plain old write(2) for writes. The CRuby implementation
on JRuby suffers from I/O incompatibilities. Until JRuby
Bunny users who run on JRuby are highly recommended to switch to March Hare, which has nearly identical API and is significantly more efficient.
Bunny::Session#with_channel is now fully synchronised and won't run into COMMAND_INVALID errors
when used from multiple threads that share a connection.
TLS connections now prefer TLSv1 (or later, if available) due to the recently discovered POODLE attack on SSLv3.
Contributed by Michael Klishin (Pivotal) and Justin Powers (Desk.com).
GH issues:
Bunny now sets a read timeout on the sockets it opens, and uses
IO.select timeouts as the most reliable option available
on Ruby 1.9 and later.
GH issue: #254.
Contributed by Andre Foeken (Nedap).
TLS certificate options now accept inline certificates as well as file paths.
Contributed by Will Barrett (Sqwiggle).
Uncaught exception handler now provides more information about the exception, including its caller (one more stack trace line).
Contributed by Carl Hörberg (CloudAMQP).
Bunny::Channel#temporary_queue is a convenience method that declares a new
server-named exclusive queue:
q = ch.temporary_queueContributed by Daniel Schierbeck (Zendesk).
Automatic connection recovery robustness improvements. Contributed by Andre Foeken (Nedap).
It is now possible to pass the :hosts option to Bunny.new/Bunny::Session#initialize.
When connection to RabbitMQ (including during connection recovery), a random host
will be chosen from the list.
Connection shuffling and robustness improvements.
Contributed by Andre Foeken (Nedap).
Breaks compatibility with Bunny 0.8.x.
Bunny:Session#default_channel was removed. Please open channels explicitly now,
as all the examples in the docs do.
Contributed by Matt Campbell.
When a connection is recovered, the sequence counter resets on the broker, but not the client. To keep things in sync the client must store a confirmation offset after a recovery.
Contributed by Devin Christensen.
During abnormal termination, Bunny::Session#close no longer tries
to call the non-existent terminate_with method on its origin
thread.
TLS now can be explicitly disabled even when connecting (without TLS) to the default RabbitMQ TLS/amqps port (5671):
conn = Bunny.new(:port => 5671, :tls => false)Contributed by Muhan Zou.
Single threaded Bunny connections will now raise exceptions that occur during shutdown as is (instead of trying to shut down I/O loop which only threaded ones have).
Contributed by Carl Hörberg.
Bunny::Session#close now better synchronizes state transitions,
eliminating a few race condition scenarios with I/O reader thread.
Bunny::Exchange.default no longer raises an exception.
Note that it is a legacy compatibility method. Please use
Bunny::Channel#default_exchange instead.
Contributed by Justin Litchfield.
GH issue #211.
Bunny::Queue#pop_as_hash, which was added to ease migration
to Bunny 0.9, was removed.
Bunny::Queue#pop now wraps basic.get-ok and message properties
into Bunny::GetResponse and Bunny::MessageProperties, just like
basic.consume deliveries.
GH issue: #212.
Publisher confirms implementation now synchronizes unconfirmed set better.
Contributed by Nicolas Viennot.
Channel id allocator is no longer reset after recovery if there are channels open. Makes it possible to open channels on a recovered connection (in addition to the channels it already had).
It is now possible to use :key (which Bunny versions prior to 0.9 used)
as well as :routing_key as an argument to Bunny::Queue#bind.
Bunny now rescues StandardError instead of Exception where
it automatically does so (e.g. when dispatching deliveries to consumers).
Contributed by Alex Young.
Initial socket connection timeout again raises Bunny::TCPConnectionFailed
on the connection origin thread.
Bunny::Session#close on connections that have experienced a network failure
will correctly clean up I/O and heartbeat sender threads.
Contributed by m-o-e.
Bunny::Concurrent::ContinuationQueue#poll no longer floors the argument
to the nearest second.
Contributed by Brian Abreu.
Per AMQP 0-9-1 spec, routing keys cannot be longer than 255 characters.
Bunny::Channel#basic_publish and Bunny::Exchange#publish now enforces
this limit.
Bunny now properly disables Nagle's algorithm on the sockets it opens. This likely means significantly lower latency for workloads that involve sending a lot of small messages very frequently.
Contributed by Nelson Gauthier (AirBnB).
Exchanges now can be declared as internal:
ch = conn.create_channel
x = ch.fanout("bunny.tests.exchanges.internal", :internal => true)Internal exchanges cannot be published to by clients and are solely used for Exchange-to-Exchange bindings and various plugins but apps may still need to bind them. Now it is possible to do so with Bunny.
Uncaught consumer exceptions are now handled by uncaught exceptions handler that can be defined per channel:
ch.on_uncaught_exception do |e, consumer|
# ...
endFull bodies of Bunny::Session#create_channel and Bunny::Session#close_channel
are now synchronized, which makes sure concurrent channel.open and subsequent
operations (e.g. exchange.declare) do not result in connection-level exceptions
(incorrect connection state transitions).
Bunny will now use actual recovery interval in the log.
Contributed by Chad Fowler.
Channel recovery now involves recovery of publisher confirms and transaction modes.
Bunny now successfully performs TLS upgrade when peer verification is disabled.
Contributed by Jordan Curzon.
Bunny::Session#with_channel now makes sure the channel is closed
even if provided block raises an exception
Contributed by Carl Hoerberg.
Bunny::Session#create_channel will now reject channel number 0.
Single threaded mode no longer fails with
undefined method `event_loop'
connection.tune.channel_max could previously be configured to values
greater than 2^16 - 1 (65535). This would result in a silent overflow
during serialization. The issue was harmless in practice but is still
a bug that can be quite confusing.
Bunny now caps max number of channels to 65535. This allows it to be forward compatible with future RabbitMQ versions that may allow limiting total # of open channels via server configuration.
Minimum amq-protocol version is now 1.9.0 which includes
bug fixes and performance improvements for channel ID allocator.
Bunny will now correctly release heartbeat sender when allocating a new one (usually happens only when connection recovers from a network failure).
Versioned delivery tag now ensures all the arguments it operates
(original delivery tag, atomic fixnum instances, etc) are coerced to Integer
before comparison.
GitHub issues: #171.
Bunny now can use any logger that provides the same API as Ruby standard library's Logger:
require "logger"
require "stringio"
io = StringIO.new
# will log to `io`
Bunny.new(:logger => Logger.new(io))Bunny uses OpenSSL provided CA certificate paths. This caused problems on some platforms on JRuby (see jruby/jruby#155).
To avoid these issues, Bunny no longer uses default CA certificate paths on JRuby (there are no changes for other Rubies), so it's necessary to provide CA certificate explicitly.
Bunny now uses slightly different ways of continuously reading from the socket
on CRuby and JRuby, to prevent abnormally high CPU usage on JRuby after a
certain period of time (the frequency of EWOULDBLOCK being raised spiked
sharply).
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:[email protected]/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