Active-Passive Messaging Clusters
Overview
The High Availability (HA) module provides
active-passive, hot-standby
messaging clusters to provide fault tolerant message delivery.
In an active-passive cluster only one broker, known as the
primary, is active and serving clients at a time. The other
brokers are standing by as backups. Changes on the primary
are replicated to all the backups so they are always up-to-date or "hot". Backup
brokers reject client connection attempts, to enforce the requirement that clients
only connect to the primary.
If the primary fails, one of the backups is promoted to take over as the new
primary. Clients fail-over to the new primary automatically. If there are multiple
backups, the other backups also fail-over to become backups of the new primary.
This approach relies on an external cluster resource manager
to detect failures, choose the new primary and handle network partitions. rgmanager is supported
initially, but others may be supported in the future.
Avoiding message loss
In order to avoid message loss, the primary broker delays
acknowledgement of messages received from clients until the
message has been replicated and acknowledged by all of the back-up
brokers, or has been consumed from the primary queue.
This ensures that all acknowledged messages are safe: they have either
been consumed or backed up to all backup brokers. Messages that are
consumed before they are replicated do not need to
be replicated. This reduces the work load when replicating a queue with
active consumers.
Clients keep unacknowledged messages in a buffer
You can control the maximum number of messages in the buffer by setting the
client's capacity. For details of how to set the capacity
in client code see "Using the Qpid Messaging API" in
Programming in Apache Qpid.
until they are acknowledged by the primary. If the primary fails, clients will
fail-over to the new primary and re-send all their
unacknowledged messages.
Clients must use "at-least-once" reliability to enable re-send of unacknowledged
messages. This is the default behaviour, no options need be set to enable it. For
details of client addressing options see "Using the Qpid Messaging API"
in Programming in Apache Qpid.
If the primary crashes, all the acknowledged
messages will be available on the backup that takes over as the new
primary. The unacknowledged messages will be
re-sent by the clients. Thus no messages are lost.
Note that this means it is possible for messages to be
duplicated. In the event of a failure it is possible for a
message to received by the backup that becomes the new primary
and re-sent by the client. The application must take steps
to identify and eliminate duplicates.
When a new primary is promoted after a fail-over it is initially in
"recovering" mode. In this mode, it delays acknowledgement of messages
on behalf of all the backups that were connected to the previous
primary. This protects those messages against a failure of the new
primary until the backups have a chance to connect and catch up.
Not all messages need to be replicated to the back-up brokers. If a
message is consumed and acknowledged by a regular client before it has
been replicated to a backup, then it doesn't need to be replicated.
HA Broker States
Stand-alone
Broker is not part of a HA cluster.
Joining
Newly started broker, not yet connected to any existing primary.
Catch-up
A backup broker that is connected to the primary and downloading
existing state (queues, messages etc.)
Ready
A backup broker that is fully caught-up and ready to take over as
primary.
Recovering
Newly-promoted primary, waiting for backups to connect and catch up.
Clients can connect but they are stalled until the primary is active.
Active
The active primary broker with all backups connected and caught-up.
Limitations
There are a some known limitations in the current implementation. These
will be fixed in future versions.
Transactional changes to queue state are not replicated atomically. If
the primary crashes during a transaction, it is possible that the
backup could contain only part of the changes introduced by a
transaction.
Configuration changes (creating or deleting queues, exchanges and
bindings) are replicated asynchronously. Management tools used to
make changes will consider the change complete when it is complete
on the primary, it may not yet be replicated to all the backups.
Federation links to the primary will fail over
correctly. Federated links from the primary
will be lost in fail over, they will not be re-connected to the new
primary. It is possible to work around this by replacing the
qpidd-primary start up script with a script that
re-creates federation links when the primary is promoted.
Virtual IP Addresses
Some resource managers (including rgmanager) support
virtual IP addresses. A virtual IP address is an IP
address that can be relocated to any of the nodes in a cluster. The
resource manager associates this address with the primary node in the
cluster, and relocates it to the new primary when there is a failure. This
simplifies configuration as you can publish a single IP address rather
than a list.
A virtual IP address can be used by clients to connect to the primary. The
following sections will explain how to configure virtual IP addresses for
clients or brokers.
Configuring the Brokers
The broker must load the ha module, it is loaded by
default. The following broker options are available for the HA module.
Broker management is required for HA to operate, it is enabled by
default. The option mgmt-enable must not be set to
"no"
Incorrect security settings are a common cause of problems when
getting started, see .
Broker Options for High Availability Messaging Cluster
Options for High Availability Messaging Cluster
ha-cluster yes|no
Set to "yes" to have the broker join a cluster.
ha-queue-replication yes|no
Enable replication of specific queues without joining a cluster, see .
ha-brokers-url URL
The URL
The full format of the URL is given by this grammar:
url = ["amqp:"][ user ["/" password] "@" ] addr ("," addr)*
addr = tcp_addr / rmda_addr / ssl_addr / ...
tcp_addr = ["tcp:"] host [":" port]
rdma_addr = "rdma:" host [":" port]
ssl_addr = "ssl:" host [":" port]'
used by cluster brokers to connect to each other. The URL should
contain a comma separated list of the broker addresses, rather than a
virtual IP address.
ha-public-url URL
This option is only needed for backwards compatibility if you
have been using the amq.failover exchange.
This exchange is now obsolete, it is recommended to use a
virtual IP address instead.
If set, this URL is advertised by the
amq.failover exchange and overrides the
broker option known-hosts-url
ha-replicate VALUE
Specifies whether queues and exchanges are replicated by default.
VALUE is one of: none,
configuration, all.
For details see .
ha-username USER
ha-password PASS
ha-mechanism MECHANISM
Authentication settings used by HA brokers to connect to each other,
see
ha-backup-timeoutSECONDS
Values specified as SECONDS can be a
fraction of a second, e.g. "0.1" for a tenth of a second.
They can also have an explicit unit,
e.g. 10s (seconds), 10ms (milliseconds), 10us (microseconds), 10ns (nanoseconds)
Maximum time that a recovering primary will wait for an expected
backup to connect and become ready.
link-maintenance-interval SECONDS
HA uses federation links to connect from backup to primary.
Backup brokers check the link to the primary on this interval
and re-connect if need be. Default 2 seconds. Set lower for
faster failover, e.g. 0.1 seconds. Setting too low will result
in excessive link-checking on the backups.
link-heartbeat-interval SECONDS
HA uses federation links to connect from backup to primary.
If no heart-beat is received for twice this interval the primary will consider that
backup dead (e.g. if backup is hung or partitioned.)
This interval is also used to time-out for broker status checks,
it may take up to this interval for rgmanager to detect a hung or partitioned broker.
Clients sending messages may be held up during this time.
Default 120 seconds: you will probably want to set this to a lower value e.g. 10.
If set too low rgmanager may consider a slow broker to have failed and kill it.
To configure a HA cluster you must set at least ha-cluster and
ha-brokers-url.
The Cluster Resource Manager
Broker fail-over is managed by a cluster resource
manager. An integration with rgmanager is
provided, but it is possible to integrate with other resource managers.
The resource manager is responsible for starting the qpidd broker
on each node in the cluster. The resource manager then promotes
one of the brokers to be the primary. The other brokers connect to the primary as
backups, using the URL provided in the ha-brokers-url configuration
option.
Once connected, the backup brokers synchronize their state with the
primary. When a backup is synchronized, or "hot", it is ready to take
over if the primary fails. Backup brokers continually receive updates
from the primary in order to stay synchronized.
If the primary fails, backup brokers go into fail-over mode. The resource
manager must detect the failure and promote one of the backups to be the
new primary. The other backups connect to the new primary and synchronize
their state with it.
The resource manager is also responsible for protecting the cluster from
split-brain conditions resulting from a network partition. A
network partition divide a cluster into two sub-groups which cannot see each other.
Usually a quorum voting algorithm is used that disables nodes
in the inquorate sub-group.
Configuring with <command>rgmanager</command> as resource manager
This section assumes that you are already familiar with setting up and configuring
clustered services using cman and
rgmanager. It will show you how to configure an active-passive,
hot-standby qpidd HA cluster with rgmanager.
Once all components are installed it is important to take the following step:
chkconfig rgmanager on
chkconfig cman on
chkconfig qpidd off
The qpidd service must be off in
chkconfig because rgmanager will
start and stop qpidd. If the normal system init
process also attempts to start and stop qpidd it can cause rgmanager to
lose track of qpidd processes. The symptom when this happens is that
clustat shows a qpidd service to
be stopped when in fact there is a qpidd process
running. The qpidd log will show errors like this:
critical Unexpected error: Daemon startup failed: Cannot lock /var/lib/qpidd/lock: Resource temporarily unavailable
You must provide a cluster.conf file to configure
cman and rgmanager. Here is
an example cluster.conf file for a cluster of 3 nodes named
node1, node2 and node3. We will go through the configuration step-by-step.
status_poll_interval is the interval in seconds that the resource manager checks the status
of managed services. This affects how quickly the manager will detect failed services.
-->
]]>
There is a failoverdomain for each node containing just that
one node. This lets us stipulate that the qpidd service should always run on all
nodes.
The resources section defines the qpidd
script used to start the qpidd service. It also defines the
qpid-primary script which does not
actually start a new service, rather it promotes the existing
qpidd broker to primary status.
The resources section also defines a virtual IP
address for clients: 20.0.20.200.
qpidd.conf should contain these lines:
ha-cluster=yes
ha-brokers-url=20.0.20.1,20.0.20.2,20.0.20.3
The brokers connect to each other directly via the addresses
listed in ha-brokers-url. Note the client and broker
addresses are on separate sub-nets, this is recommended but not required.
The service section defines 3 qpidd
services, one for each node. Each service is in a restricted fail-over
domain containing just that node, and has the restart
recovery policy. The effect of this is that rgmanager will run
qpidd on each node, restarting if it fails.
There is a single qpidd-primary-service using the
qpidd-primary script which is not restricted to a
domain and has the relocate recovery policy. This means
rgmanager will start qpidd-primary on one of the nodes
when the cluster starts and will relocate it to another node if the
original node fails. Running the qpidd-primary script
does not start a new broker process, it promotes the existing broker to
become the primary.
Shutting down qpidd on a HA node
As explained above both the per-node qpidd service
and the re-locatable qpidd-primary service are
implemented by the same qpidd daemon.
As a result, stopping the qpidd service will not stop
a qpidd daemon that is acting as primary, and
stopping the qpidd-primary service will not stop a
qpidd process that is acting as backup.
To shut down a node that is acting as primary you need to shut down the
qpidd service and relocate the
primary:
clusvcadm -d somenode-qpidd-service
clusvcadm -r qpidd-primary-service
This will shut down the qpidd daemon on that node and
prevent the primary service service from relocating back to the node
because the qpidd service is no longer running there.
Broker Administration Tools
Normally, clients are not allowed to connect to a backup broker. However
management tools are allowed to connect to a backup brokers. If you use
these tools you must not add or remove messages from
replicated queues, nor create or delete replicated queues or exchanges as
this will disrupt the replication process and may cause message loss.
qpid-ha allows you to view and change HA configuration settings.
The tools qpid-config, qpid-route and
qpid-stat will connect to a backup if you pass the flag ha-admin on the
command line.
Controlling replication of queues and exchanges
By default, queues and exchanges are not replicated automatically. You can change
the default behaviour by setting the ha-replicate configuration
option. It has one of the following values:
all: Replicate everything automatically: queues,
exchanges, bindings and messages.
configuration: Replicate the existence of queues,
exchange and bindings but don't replicate messages.
none: Don't replicate anything, this is the default.
You can over-ride the default for a particular queue or exchange by passing the
argument qpid.replicate when creating the queue or exchange. It
takes the same values as ha-replicate
Bindings are automatically replicated if the queue and exchange being bound both
have replication all or configuration, they
are not replicated otherwise.
You can create replicated queues and exchanges with the
qpid-config management tool like this:
qpid-config add queue myqueue --replicate all
To create replicated queues and exchanges via the client API, add a
node entry to the address like this:
"myqueue;{create:always,node:{x-declare:{arguments:{'qpid.replicate':all}}}}"
There are some built-in exchanges created automatically by the broker, these
exchanges are never replicated. The built-in exchanges are the default (nameless)
exchange, the AMQP standard exchanges (amq.direct, amq.topic, amq.fanout and
amq.match) and the management exchanges (qpid.management, qmf.default.direct and
qmf.default.topic)
Note that if you bind a replicated queue to one of these exchanges, the
binding will not be replicated, so the queue will not
have the binding after a fail-over.
Client Connection and Fail-over
Clients can only connect to the primary broker. Backup brokers reject any
connection attempt by a client. Clients rejected by a backup broker will
automatically fail-over until they connect to the primary.
Clients are configured with the URL for the cluster (details below for
each type of client). There are two possibilities
The URL contains multiple addresses, one for each broker in the cluster.
The URL contains a single virtual IP address
that is assigned to the primary broker by the resource manager.
This is the recommended configuration.
In the first case, clients will repeatedly re-try each address in the URL
until they successfully connect to the primary. In the second case the
resource manager will assign the virtual IP address to the primary broker,
so clients only need to re-try on a single address.
When the primary broker fails, clients re-try all known cluster addresses
until they connect to the new primary. The client re-sends any messages
that were previously sent but not acknowledged by the broker at the time
of the failure. Similarly messages that have been sent by the broker, but
not acknowledged by the client, are re-queued.
TCP can be slow to detect connection failures. A client can configure a
connection to use a heartbeat to detect connection
failure, and can specify a time interval for the heartbeat. If heartbeats
are in use, failures will be detected no later than twice the heartbeat
interval. The following sections explain how to enable heartbeat in each
client.
Note: the following sections explain how to configure clients with
multiple dresses, but if you are using a virtual IP address you only need
to configure that one address for clients, you don't need to list all the
addresses.
Suppose your cluster has 3 nodes: node1,
node2 and node3 all using the
default AMQP port, and you are not using a virtual IP address. To connect
a client you need to specify the address(es) and set the
reconnect property to true. The
following sub-sections show how to connect each type of client.
C++ clients
With the C++ client, you specify multiple cluster addresses in a single URL
The full grammar for the URL is:
url = ["amqp:"][ user ["/" password] "@" ] addr ("," addr)*
addr = tcp_addr / rmda_addr / ssl_addr / ...
tcp_addr = ["tcp:"] host [":" port]
rdma_addr = "rdma:" host [":" port]
ssl_addr = "ssl:" host [":" port]'
You also need to specify the connection option
reconnect to be true. For example:
qpid::messaging::Connection c("node1,node2,node3","{reconnect:true}");
Heartbeats are disabled by default. You can enable them by specifying a
heartbeat interval (in seconds) for the connection via the
heartbeat option. For example:
qpid::messaging::Connection c("node1,node2,node3","{reconnect:true,heartbeat:10}");
Python clients
With the python client, you specify reconnect=True
and a list of host:port addresses as
reconnect_urls when calling
Connection.establish or
Connection.open
connection = qpid.messaging.Connection.establish("node1", reconnect=True, reconnect_urls=["node1", "node2", "node3"])
Heartbeats are disabled by default. You can
enable them by specifying a heartbeat interval (in seconds) for the
connection via the 'heartbeat' option. For example:
connection = qpid.messaging.Connection.establish("node1", reconnect=True, reconnect_urls=["node1", "node2", "node3"], heartbeat=10)
Java JMS Clients
In Java JMS clients, client fail-over is handled automatically if it is
enabled in the connection. You can configure a connection to use
fail-over using the failover property:
connectionfactory.qpidConnectionfactory = amqp://guest:guest@clientid/test?brokerlist='tcp://localhost:5672'&failover='failover_exchange'
This property can take three values:
Fail-over Modes
failover_exchange
If the connection fails, fail over to any other broker in the cluster.
roundrobin
If the connection fails, fail over to one of the brokers specified in the brokerlist.
singlebroker
Fail-over is not supported; the connection is to a single broker only.
In a Connection URL, heartbeat is set using the heartbeat property, which is an integer corresponding to the heartbeat period in seconds. For instance, the following line from a JNDI properties file sets the heartbeat time out to 3 seconds:
connectionfactory.qpidConnectionfactory = amqp://guest:guest@clientid/test?brokerlist='tcp://localhost:5672'&heartbeat='3'
Security and Access Control.
This section outlines the HA specific aspects of security configuration.
Please see for
more details on enabling authentication and setting up Access Control Lists.
Unless you disable authentication with auth=no in
your configuration, you must set the options below
and you must have an ACL file with at least the
entry described below.
Backups will be unable to connect to the primary if
the security configuration is incorrect. See also
When authentication is enabled you must set the credentials used by HA
brokers with following options:
HA Security Options
HA Security Options
ha-username USER
User name for HA brokers. Note this must not include the @QPID suffix.
ha-password PASS
Password for HA brokers.
ha-mechanism MECHANISM
Mechanism for HA brokers. Any mechanism you enable for
broker-to-broker communication can also be used by a client, so
do not use ha-mechanism=ANONYMOUS in a secure environment.
This identity is used to authorize federation links from backup to
primary. It is also used to authorize actions on the backup to replicate
primary state, for example creating queues and exchanges.
When authorization is enabled you must have an Access Control List with the
following rule to allow HA replication to function. Suppose
ha-username=USER
acl allow USER@QPID all all
Integrating with other Cluster Resource Managers
To integrate with a different resource manager you must configure it to:
Start a qpidd process on each node of the cluster.
Restart qpidd if it crashes.
Promote exactly one of the brokers to primary.
Detect a failure and promote a new primary.
The qpid-ha command allows you to check if a broker is
primary, and to promote a backup to primary.
To test if a broker is the primary:
qpid-ha -b broker-address status --expect=primary
This will return 0 if the broker at broker-address is the primary,
non-0 otherwise.
To promote a broker to primary:
qpid-ha --cluster-manager -b broker-address promote
Note that promote is considered a "cluster manager
only" command. Incorrect use of promote outside of the
cluster manager could create a cluster with multiple primaries. Such a
cluster will malfunction and lose data. "Cluster manager only" commands
are not accessible in qpid-ha without the
--cluster-manager option.
To list the full set of commands use:
qpid-ha --cluster-manager --help
Using a message store in a cluster
If you use a persistent store for your messages then each broker in a
cluster will have its own store. If the entire cluster fails and is
restarted, the *first* broker that becomes primary will recover from its
store. All the other brokers will clear their stores and get an update
from the primary to ensure consistency.
Troubleshooting a cluster
This section applies to clusters that are using rgmanager as the
cluster manager.
No primary broker
When you initially start a HA cluster, all brokers are in
joining mode. The brokers do not automatically select
a primary, they rely on the cluster manager rgmanager
to do so. If rgmanager is not running or is not
configured correctly, brokers will remain in the
joining state. See
Authentication and ACL failures
If a broker is unable to establish a connection to another broker in the
cluster due to authentication or ACL problems the logs may contain
errors like the following:
info SASL: Authentication failed: SASL(-13): user not found: Password verification failed
warning Client closed connection with 320: User anonymous@QPID federation connection denied. Systems with authentication enabled must specify ACL create link rules.
warning Client closed connection with 320: ACL denied anonymous@QPID creating a federation link.
Set the HA security configuration and ACL file as described in . Once the cluster is running and the primary is
promoted , run:
qpid-ha status --all
to make sure that the brokers are running as one cluster.
Slow recovery times
The following configuration settings affect recovery time. The
values shown are examples that give fast recovery on a lightly
loaded system. You should run tests to determine if the values are
appropriate for your system and load conditions.
cluster.conf:
<rm status_poll_interval=1>
status_poll_interval is the interval in seconds that the
resource manager checks the status of managed services. This
affects how quickly the manager will detect failed services.
<ip address="20.0.20.200" monitor_link="yes" sleeptime="0"/>
This is a virtual IP address for client traffic.
monitor_link="yes" means monitor the health of the network interface
used for the VIP. sleeptime="0" means don't delay when
failing over the VIP to a new address.
qpidd.conf
link-maintenance-interval=0.1
Interval for backup brokers to check the link to the primary
re-connect if need be. Default 2 seconds. Can be set lower for
faster fail-over. Setting too low will result in excessive
link-checking activity on the broker.
link-heartbeat-interval=5
Heartbeat interval for federation links. The HA cluster uses
federation links between the primary and each backup. The
primary can take up to twice the heartbeat interval to detect a
failed backup. When a sender sends a message the primary waits
for all backups to acknowledge before acknowledging to the
sender. A disconnected backup may cause the primary to block
senders until it is detected via heartbeat.
This interval is also used as the timeout for broker status
checks by rgmanager. It may take up to this interval for
rgmanager to detect a hung broker.
The default of 120 seconds is very high, you will probably want
to set this to a lower value. If set too low, under network
congestion or heavy load, a slow-to-respond broker may be
re-started by rgmanager.
Total cluster failure
Note: for definition of broker states joining,
catch-up, ready,
recovering and active see
The cluster can only guarantee availability as long as there is at
least one active primary broker or ready backup broker left alive.
If all the brokers fail simultaneously, the cluster will fail and
non-persistent data will be lost.
While there is an active primary broker, clients can get service.
If the active primary fails, one of the "ready" backup
brokers will take over, recover and become active. Note a backup
can only be promoted to primary if it is in the "ready"
state (with the exception of the first primary in a new cluster
where all brokers are in the "joining" state)
Given a stable cluster of N brokers with one active primary and
N-1 ready backups, the system can sustain up to N-1 failures in
rapid succession. The surviving broker will be promoted to active
and continue to give service.
However at this point the system cannot
sustain a failure of the surviving broker until at least one of
the other brokers recovers, catches up and becomes a ready backup.
If the surviving broker fails before that the cluster will fail in
one of two modes (depending on the exact timing of failures)
1. The cluster hangs
All brokers are in joining or catch-up mode. rgmanager tries to
promote a new primary but cannot find any candidates and so
gives up. clustat will show that the qpidd services are running
but the the qpidd-primary service has stopped, something like
this:
Service Name Owner (Last) State
------- ---- ----- ------ -----
service:mrg33-qpidd-service 20.0.10.33 started
service:mrg34-qpidd-service 20.0.10.34 started
service:mrg35-qpidd-service 20.0.10.35 started
service:qpidd-primary-service (20.0.10.33) stopped
Eventually all brokers become stuck in "joining" mode,
as shown by: qpid-ha status --all
At this point you need to restart the cluster in one of the
following ways:
Restart the entire cluster:
In luci:your-cluster:Nodes
click reboot to restart the entire cluster
Stop and restart the cluster with
ccs --stopall; ccs --startall
Restart just the Qpid services:In luci:your-cluster:Service Groups
Select all the qpidd (not qpidd-primary) services, click restart
Select the qpidd-primary service, click restart
Stop the qpidd-primary and
qpidd services with clusvcadm,
then restart (qpidd-primary last)
2. The cluster reboots
A new primary is promoted and the cluster is functional but all
non-persistent data from before the failure is lost.
Fencing and network partitions
A network partition is a a network failure that divides the
cluster into two or more sub-clusters, where each broker can
communicate with brokers in its own sub-cluster but not with
brokers in other sub-clusters. This condition is also referred to
as a "split brain".
Nodes in one sub-cluster can't tell whether nodes in other
sub-clusters are dead or are still running but disconnected. We
cannot allow each sub-cluster to independently declare its own
qpidd primary and start serving clients, as the cluster will
become inconsistent. We must ensure only one sub-cluster continues
to provide service.
A quorum determines which sub-cluster
continues to operate, and power fencing
ensures that nodes in non-quorate sub-clusters cannot attempt to
provide service inconsistently. For more information see:
https://access.redhat.com/site/documentation/en-US/Red_Hat_Enterprise_Linux/6/html-single/High_Availability_Add-On_Overview/index.html,
chapter 2. Quorum and 4. Fencing.