Changes between Bunny 2.6.3 and 2.6.4 (unreleased)
Recovery Attempt Counting Strategy Changed
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
Changes between Bunny 2.6.2 and 2.6.3 (January 19th, 2016)
Retry on new Ruby 2.1+ variations of EAGAIN
, EWOULDBLOCK
GitHub issue: #456
Changes between Bunny 2.6.1 and 2.6.2 (December 17th, 2016)
Do Not Modify Host Arrays
Bunny now can work with frozen host arrays.
GitHub issue: #446
Changes between Bunny 2.6.0 and 2.6.1 (October 23rd, 2016)
Bunny::ConsumerWorkPool#shutdown Terminates Early When It's Safe to Do So
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.
Changes between Bunny 2.5.0 and 2.6.0 (October 15th, 2016)
Graceful Shutdown of Consumers
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.
Bunny::Channel#wait_for_confirms
Now Throws When Used on a Closed Channel
GitHub issue: #428
Contributed by Dimitar Dimitrov.
Race Condition Eliminated in Bunny::Channel#wait_for_confirms
GitHub issue: #424
Contributed by Dimitar Dimitrov.
More Defensive Consumer Work Pool
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.
Changes between Bunny 2.4.0 and 2.5.0 (July 20th, 2016)
Exchange Bindings are Now Correctly Recovered
GitHub issue: #410
Contributed by Andrew Bruce.
Bunny::Channel#wait_for_confirms
Awaits While There're Outstanding Unconfirmed Messages
GitHub issue: #424
Contributed by Dimitar Dimitrov.
Queue Recovery Respects the :no_declare
Option
Queue recovery now respects the :no_declare
option.
Bunny::Channel#wait_for_confirms
Throws Early
Bunny::Channel#wait_for_confirms
now throws an exception
early when invoked on a closed channel.
GitHub issue: #428.
Contributed by Dimitar Dimitrov.
Changes between Bunny 2.3.0 and 2.4.0 (June 11th, 2016)
This release includes minor breaking API changes.
Unconfirmed Delivery Tag Set Reset on Network Recovery
Channels will now reset their unconfirmed delivery tag set after recovery.
GitHub issue: #406
Contributed by Bill Ruddock.
Support (Quoted) IPv6 Addresses in Address Lists
GitHub issue: #383.
Contributed by Jeremy Heiler.
Transport#read_fully Doesn't Try to Recover
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
Merged with Defaults
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.
More Predictable RABBITMQ_URL Handling
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.
Changes between Bunny 2.2.0 and 2.3.0 (Feb 26th, 2016)
Thread#abort_on_exception Setting for Consumer Work Pool Threads
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.
Explicit Transport Closure on Recovery
Bunny now will explicitly close previosly used transport before starting connection recovery.
GitHub issue: #377.
Contributed by bkanhoopla.
No TLS Socket Double-init
Makes sure that TLS sockets are not double-initialized.
GH issue: #345.
Contributed by Carl Hörberg.
Lazily Evaluated Debug Log Strings
GH issue: #375
Contributed by Omer Katz.
Changes between Bunny 2.1.0 and 2.2.0 (Sep 6th, 2015)
Add :addresses to connect options
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).
Recover from connection.close by default
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).
Changes between Bunny 2.0.0 and 2.1.0
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 Value Serialisation in 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.
Explicit Ruby 2.0 Requirement
Bunny now requires Ruby 2.0 in the gemspec.
Contributed by Carl Hörberg.
JRuby Fix
Bunny runs again on JRuby. Note that JRuby users are strongly advised to use March Hare instead.
Contributed by Teodor Pripoae.
Changes between Bunny 1.7.0 and 2.0.0
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.
Minimum Required Ruby Version is 2.0
Bunny 2.0
requires Ruby 2.0 or later.
Non-Blocking Writes
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.
Reduced Timeout Use
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.
Capped Number of Connection Recovery Attempts
: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 and Related Methods Improvements
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.
Logger Output Remains Consistent
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.
prefetch_count is Limited to 65535
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.
Per-Consumer and Per-Channel Prefetch
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.
Changes between Bunny 1.6.0 and 1.7.0
TLS Peer Verification Enabled by Default
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).
Higher Default Connection Timeout
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.
Socket Read Timeout No Longer Set to 0 With Disabled Heartbeats
GH issue: #267.
JRuby Writes Fixes
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 Synchornisation Improvements
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.
Changes between Bunny 1.5.0 and 1.6.0
TLSv1 by Default
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:
Socket Read and Write Timeout Improvements
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).
Inline TLS Certificates Support
TLS certificate options now accept inline certificates as well as file paths.
Contributed by Will Barrett (Sqwiggle).
Changes between Bunny 1.4.0 and 1.5.0
Improved Uncaught Exception Handler
Uncaught exception handler now provides more information about the exception, including its caller (one more stack trace line).
Contributed by Carl Hörberg (CloudAMQP).
Convenience Method for Temporary (Server-named, Exclusive) Queue Declaration
Bunny::Channel#temporary_queue
is a convenience method that declares a new
server-named exclusive queue:
q = ch.temporary_queue
Contributed by Daniel Schierbeck (Zendesk).
Recovery Reliability Improvements
Automatic connection recovery robustness improvements. Contributed by Andre Foeken (Nedap).
Host Lists
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).
Default Channel Removed
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.
Changes between Bunny 1.3.0 and 1.4.0
Channel#wait_for_confirms Returns Immediately If All Publishes Confirmed
Contributed by Matt Campbell.
Publisher Confirms is In Sync After Recovery
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.
NoMethodError on Thread During Shutdown
During abnormal termination, Bunny::Session#close
no longer tries
to call the non-existent terminate_with
method on its origin
thread.
Changes between Bunny 1.2.0 and 1.3.0
TLS Can Be Explicitly Disabled
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 Connections Raise Shutdown Exceptions
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.
Synchronization Improvements for Session#close
Bunny::Session#close
now better synchronizes state transitions,
eliminating a few race condition scenarios with I/O reader thread.
Bunny::Exchange.default Fix
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 Removed
Bunny::Queue#pop_as_hash
, which was added to ease migration
to Bunny 0.9, was removed.
Bunny::Queue#pop Wraps Metadata
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.
Better Synchronization for Publisher Confirms
Publisher confirms implementation now synchronizes unconfirmed set better.
Contributed by Nicolas Viennot.
Channel Allocation After Recovery
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).
Changes between Bunny 1.1.0 and 1.2.0
:key Supported in Bunny::Channel#queue_bind
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
.
System Exceptions Not Rescued by the Library
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
Initial socket connection timeout again raises Bunny::TCPConnectionFailed
on the connection origin thread.
Thread Leaks Plugged
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 Rounding Fix
Bunny::Concurrent::ContinuationQueue#poll
no longer floors the argument
to the nearest second.
Contributed by Brian Abreu.
Routing Key Limit
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.
Nagle's Algorithm Disabled Correctly
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).
Internal Exchanges
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
Uncaught consumer exceptions are now handled by uncaught exceptions handler that can be defined per channel:
ch.on_uncaught_exception do |e, consumer|
# ...
end
Changes between Bunny 1.1.0.rc1 and 1.1.0
Synchronized Session#create_channel and Session#close_channel
Full 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).
Corrected Recovery Log Message
Bunny will now use actual recovery interval in the log.
Contributed by Chad Fowler.
Changes between Bunny 1.1.0.pre2 and 1.1.0.rc1
Full Channel State Recovery
Channel recovery now involves recovery of publisher confirms and transaction modes.
TLS Without Peer Verification
Bunny now successfully performs TLS upgrade when peer verification is disabled.
Contributed by Jordan Curzon.
Bunny::Session#with_channel Ensures the Channel is Closed
Bunny::Session#with_channel
now makes sure the channel is closed
even if provided block raises an exception
Contributed by Carl Hoerberg.
Channel Number = 0 is Rejected
Bunny::Session#create_channel
will now reject channel number 0.
Single Threaded Mode Fixes
Single threaded mode no longer fails with
undefined method `event_loop'
Changes between Bunny 1.1.0.pre1 and 1.1.0.pre2
connection.tune.channel_max No Longer Overflows
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.
amq-protocol Update
Minimum amq-protocol
version is now 1.9.0
which includes
bug fixes and performance improvements for channel ID allocator.
Thread Leaks Fixes
Bunny will now correctly release heartbeat sender when allocating a new one (usually happens only when connection recovers from a network failure).
Changes between Bunny 1.0.0 and 1.1.0.pre1
Versioned Delivery Tag Fix
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.
User-Provided Loggers
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))
Default CA's Paths Are Disabled on JRuby
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.
Fixes CPU Burn on JRuby
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).
Changes between Bunny 1.0.0.rc2 and 1.0.0.rc3
Authentication Failure Notification Support
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?
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?
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.
Inline TLS Certificates and Keys
It is now possible to provide inline client certificate and private key (as strings) instead of filesystem paths. The options are the same:
:tls
which, when set totrue
, will set SSL context up and switch to TLS port (5671):tls_cert
which now can be a client certificate (public key) in PEM format:tls_key
which now can be a client key (private key) in PEM format:tls_ca_certificates
which 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"])
Changes between Bunny 1.0.0.rc1 and 1.0.0.rc2
Ruby 1.8.7 Compatibility Fixes
Ruby 1.8.7 compatibility fixes around timeouts.
Changes between Bunny 1.0.0.pre6 and 1.0.0.rc1
amq-protocol Update
Minimum amq-protocol
version is now 1.8.0
which includes
a bug fix for messages exactly 128 Kb in size.
Add timeout Bunny::ConsumerWorkPool#join
Bunny::ConsumerWorkPool#join
now accepts an optional
timeout argument.
Changes between Bunny 1.0.0.pre5 and 1.0.0.pre6
Respect RABBITMQ_URL value
RABBITMQ_URL
env variable will now have effect even if
Bunny.new is invoked without arguments.
Changes between Bunny 1.0.0.pre4 and 1.0.0.pre5
Ruby 1.8 Compatibility
Bunny is Ruby 1.8-compatible again and no longer references
RUBY_ENGINE
.
Bunny::Session.parse_uri
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")
Default Paths for TLS/SSL CA's on All OS'es
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.
Changes between Bunny 1.0.0.pre3 and 1.0.0.pre4
Default Paths for TLS/SSL CA's on Linux
Bunny now will use the following TLS/SSL CA's paths on Linux by default:
/etc/ssl/certs/ca-certificates.crt
on Ubuntu/Debian/etc/ssl/certs/ca-bundle.crt
on Amazon Linux/etc/ssl/ca-bundle.pem
on OpenSUSE/etc/pki/tls/certs/ca-bundle.crt
on Fedora/RHEL
and will log a warning if no CA files are available via default paths
or :tls_ca_certificates
.
Contributed by Carl Hörberg.
Consumers Can Be Re-Registered From Bunny::Consumer#handle_cancellation
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 Fixed for Single Threaded Connections
Bunny::Session#close
with single threaded connections no longer fails
with a nil pointer exception.
Changes between Bunny 1.0.0.pre2 and 1.0.0.pre3
This release has breaking API changes.
Safe[r] basic.ack, basic.nack and basic.reject implementation
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.
Changes between Bunny 1.0.0.pre1 and 1.0.0.pre2
Exclusivity Violation for Consumers Now Raises a Reasonable Exception
When a second consumer is registered for the same queue on different channels,
a reasonable exception (Bunny::AccessRefused
) will be raised.
Reentrant Mutex Implementation
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.
Eliminated Race Condition in Bunny::Session#close
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.close Raises Exceptions on Connection Thread
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
Client TLS Certificates are Optional
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.
Changes between Bunny 0.9.0 and 1.0.0.pre1
Publishing Over Closed Connections
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.
Reliability Improvement in Automatic Network Failure Recovery
Bunny now ensures a new connection transport (socket) is initialized before any recovery is attempted.
Reliability Improvement in Bunny::Session#create_channel
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).
Changes between Bunny 0.9.0.rc1 and 0.9.0.rc2
Channel Now Properly Restarts Consumer Pool
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.
Bunny::Queue#pop_waiting is Removed
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.
More Flexible SSLContext Configuration
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.inspect
This also means that Bunny.new
will now open the socket. Previously
it was only done when Bunny::Session#start
was invoked.
Changes between Bunny 0.9.0.pre13 and 0.9.0.rc1
TLS Support
Bunny 0.9 finally supports TLS. There are 3 new options Bunny.new
takes:
:tls
which, when set totrue
, will set SSL context up and switch to TLS port (5671):tls_cert
which is a string path to the client certificate (public key) in PEM format:tls_key
which is a string path to the client key (private key) in PEM format:tls_ca_certificates
which 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"])
Bunny::Queue#pop_waiting
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
Bunny::InvalidCommand
is now Bunny::CommandInvalid
(follows
the exception class naming convention based on response status
name).
Changes between Bunny 0.9.0.pre12 and 0.9.0.pre13
Channels Without Consumers Now Tear Down Consumer Pools
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.
Consumer and Channel Available In Delivery Handlers
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
end
This allows using Bunny::Queue#subscribe
for one-off consumers
much easier, including when used with the :block
option.
Bunny::Exchange#wait_for_confirms
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.
Changes between Bunny 0.9.0.pre11 and 0.9.0.pre12
Ruby 1.8 Compatibility Regression Fix
Bunny::Socket
no longer uses Ruby 1.9-specific constants.
Bunny::Channel#wait_for_confirms Return Value Regression Fix
Bunny::Channel#wait_for_confirms
returns true
or false
again.
Changes between Bunny 0.9.0.pre10 and 0.9.0.pre11
Bunny::Session#create_channel Now Accepts Consumer Work Pool Size
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)
Heartbeat Fix For Long Running Consumers
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.
Time-bound continuations
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.
Logging Support
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)info
warn
error
fatal
(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.
Changes between Bunny 0.9.0.pre9 and 0.9.0.pre10
This release contains a breaking API change.
Concurrency Improvements On JRuby
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.
Explicitly Closed Sockets
Bunny now will correctly close the socket previous connection had when recovering from network issues.
Bunny::Exception Now Extends StandardError
Bunny::Exception
now inherits from StandardError
and not Exception
.
Naked rescue like this
begin
# ...
rescue => e
# ...
end
catches 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.
Changes between Bunny 0.9.0.pre8 and 0.9.0.pre9
Bunny::Session#start Now Returns a Session
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.
More Reliable Heartbeat Sender
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 Recovery After Delay
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.
Default Heartbeat Value Is Now Server-Defined
Bunny will now use heartbeat value provided by RabbitMQ by default.
Changes between Bunny 0.9.0.pre7 and 0.9.0.pre8
Stability Improvements
Several stability improvements in the network layer, connection error handling, and concurrency hazards.
Automatic Connection Recovery Can Be Disabled
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.
No Timeout Control For Publishing
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.
Changes between Bunny 0.9.0.pre6 and 0.9.0.pre7
Bunny::Channel#on_error
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
end
Fixed Framing of Larger Messages With Unicode Characters
Larger (over 128K) messages with non-ASCII characters are now always encoded
correctly with amq-protocol 1.2.0
.
Efficiency Improvements
Publishing of large messages is now done more efficiently.
Contributed by Greg Brockman.
API Reference
Bunny API reference is now up online.
Bunny::Channel#basic_publish Support For :persistent
Bunny::Channel#basic_publish
now supports both
:delivery_mode
and :persistent
options.
Bunny::Channel#nacked_set
Bunny::Channel#nacked_set
is a counter-part to Bunny::Channel#unacked_set
that contains basic.nack
-ed (rejected) delivery tags.
Single-threaded Network Activity Mode
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.
Changes between Bunny 0.9.0.pre5 and 0.9.0.pre6
Automatic Network Failure Recovery
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.
Confirms Listeners
Bunny now supports listeners (callbacks) on
ch.confirm_select do |delivery_tag, multiple, nack|
# handle confirms (e.g. perform retries) here
end
Contributed by Greg Brockman.
Publisher Confirms Improvements
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.
Publisher Confirms Concurrency Fix
Bunny::Channel#wait_for_confirms
will now correctly block the calling
thread until all pending confirms are received.
Changes between Bunny 0.9.0.pre4 and 0.9.0.pre5
Channel Errors Reset
Channel error information is now properly reset when a channel is (re)opened.
GH issue: #83.
Bunny::Consumer#initial Default Change
the default value of Bunny::Consumer
noack argument changed from false to true
for consistency.
Bunny::Session#prefetch Removed
Global prefetch is not implemented in RabbitMQ, so Bunny::Session#prefetch
is gone from the API.
Queue Redeclaration Bug Fix
Fixed a problem when a queue was not declared after being deleted and redeclared
GH issue: #80
Channel Cache Invalidation
Channel queue and exchange caches are now properly invalidated when queues and exchanges are deleted.
Changes between Bunny 0.9.0.pre3 and 0.9.0.pre4
Heartbeats Support Fixes
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.
Eliminate Race Conditions When Registering Consumers
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.
Support for Alternative Authentication Mechanisms
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.start
Bunny::Consumer#cancel
A 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 Behavior Change
Bunny::Exchange#delete
will no longer delete pre-declared exchanges
that cannot be declared by Bunny (amq.*
and the default exchange).
Bunny::DeliveryInfo#redelivered?
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.
Corrected Bunny::DeliveryInfo#delivery_tag Name
Bunny::DeliveryInfo#delivery_tag
had a typo which is now fixed.
Changes between Bunny 0.9.0.pre2 and 0.9.0.pre3
Client Capabilities
Bunny now correctly lists RabbitMQ extensions it currently supports in client capabilities:
basic.nack
- exchange-to-exchange bindings
- consumer cancellation notifications
- publisher confirms
Publisher Confirms Support
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 RabbitMQ
Consumers as Objects
It 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_delivery
with 3 arguments:handle_cancellation
with 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)
RABBITMQ_URL ENV variable support
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.
Changes between Bunny 0.9.0.pre1 and 0.9.0.pre2
Change Bunny::Queue#pop default for :ack to false
It makes more sense for beginners that way.
Bunny::Queue#subscribe now support the new :block option
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 again
Bunny::Exchange#publish
now supports :key
as an alias for
:routing_key
.
Bunny::Session#queue et al.
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, Bunny::Session#exchange
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)
Bunny::Queue#subscribe now properly takes 3 arguments
q.subscribe(:exclusive => false, :ack => false) do |delivery_info, properties, payload|
# ...
end
Changes between Bunny 0.8.x and 0.9.0.pre1
New convenience functions: Bunny::Channel#fanout, Bunny::Channel#topic
Bunny::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::Queue#pop and consumer handlers (Bunny::Queue#subscribe) signatures have changed
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.pop
The improve is both in that Ruby has positional destructuring, e.g.
delivery_info, _, content = q.pop
but 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
Bunny::Client#write now raises Bunny::ConnectionError
instead of Bunny::ServerDownError
when network
I/O operations fail.
Bunny::Client.create_channel now uses a bitset-based allocator
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 Bunny::TCPConnectionFailed
Bunny::ServerDownError
is now an alias for Bunny::TCPConnectionFailed