Reliable Broker Cluster

This document describes cluster design and implementation as of 19 June 2009.

Overview

A Reliable Broker Clusteror just

Cluster Design Note

A Qpid cluster is a group of brokers co-operating collaborating to provide present the
illusion of a single "virtual broker " with some extra qualities:

...

with multiple addresses. The cluster is active-active, that is to say each member broker maintains the full state of the clustered broker. If any member fails, clients can fail-over to

...

any other member.

New members can be added to a cluster while it is running. An established member volunteers to provide a state update to the new member. Both updater and updatee queue up cluster activity during the update and process it when the update is complete.

The cluster uses the CPG (Closed Process Group) protocol to replicate state. CPG was part of Open AIS package,  it is now part of the  corosync package. To avoid confusion with AMQP messages we will refer to CPG multicast messages as events.

CPG is a virtual synchrony protocol. Members multicast events to the group and CPG ensures that each member receives all the events in the same sequence. Since all members get an identical sequence of events, they can all update their state consistently. To achieve consistency, events must be processed in the order that CPG presents them. In particular members wait for their own events to be re-delivered by CPG before acting on them.

Implementation Approach

The cluster implementation is highly decoupled from the broker. There's no cluster-specific code in the general broker, just a few hooks that the cluster uses to modify broker behavior.

The basic idea is that the cluster treats the broker as a black box and assumes that provided it is fed identical input, it will produce identical results. The cluster::Connection class intercepts data arriving for broker Connections. and sends that data as a CPG event. As data events are delivered by CPG, they are fed to the original broker::Connection objects. Thus each member sees all the data arriving at all the members in the same sequence, so we get the same set of declares, enqueues, dequeues etc. happening on each member.

This approach replicates all broker state: sessions, connections, consumers, wiring etc.  Each broker can have both direct connections and shadow connections. A shadow connection represents a connection on another broker in the cluster. Members use shadow connections to simulate the actions of other brokers, so that all members arrive at the same state. Output for shadow connections is just discarded, brokers only send data to their directly-connected clients.

This approach assumes that the behavior of the broker is determinisitc, that it is completely determined by the input data fed to the broker. There are a number of cases where this does not hold and the cluster has to take steps to ensure consistency:

• Allocating messages: the stand-alone broker allocates messages based on the writability of client connections.
• Client connection disconnects.
• Timers: any action triggered by a timer may happen at an unpredictable point with respect to CPG events.

Allocating messages

The cluster allocates messages to consumers using CPG events rather than writability of client connections. A cluster connection that has potentially got data to write sends a do-output event to itself, allowing it to dequeue N messages. The messages are not actually dequeued until the do-output event is re-delivered in sequence with other events. The value of N is dynamically estimated in an attempt to match it to the rate of writing messages to directly connected clients. All the other members have a shadow connection which allows them to de-queue the same set of messages as the directly connected member.

Client disconnects

When a client disconnects, the directly-connected broker sends a deliver-close event via CPG. It does not actually destroy the connection till that message is re-delivered. This ensures that the direct connection and all the shadows are destroyed at the same point in the event sequence.

 Actions initiated by a timer

The cluster needs to do some extra work at any points where the broker takes action based on a timer (e.g. message expiry, management, producer flow control)  See the source code for details of how each is handled.

Error Handling

There are two types of recoverable error

• Predictable errors occur in the same way on all brokers as a predictable consequence of cluster events. For example binding a queue to a non-existent exchange.
• Unpredictable errors may not occur on all brokers. For example running out of journal space to store a message, or an IO error from the journal.

Unpredictable errors must be handled in such a way that the cluster does not become inconsistent. In a situation where one broker experiences an unpredictable error and the others do not, we want the broker in error to shut down and leave the cluster so its clients can fail over to healthy brokers.

When an error occurs on a cluster member it sends an error-check event to the cluster and stalls processing. If it receives a matching error-check from all other cluster members, it continues. If the error did not occur on some members, those members send an error-check with "no error" status. In this case members that did experience an error shut themselves down as they can no longer consistently update their state. The member that did not have the error continue, clients can fail over to them.

Transactions

Transactions are conversational state, allowing a session to collect changes for the shared state and then apply them all at once or not at all.

For TX transactions each broker creates an identical transaction, they all succeed or fail identically since they're all being fed identical input (see Error Handling above for what happens if a broker doesn't reach the same conclusion.)

DTX transactions are not yet supported by the cluster.

Persistence and Asynchronous Journaling

Each cluster member has an independent store, each recording identical state.

A cluster can be configured so that if the cluster is reduced to a single member  (the "last man standing") that member can have transient data queues persisted.

Recovery: after a total cluster shutdown, the state of the new cluster is determined by the store of the first broker started. The second and subsequent brokers will get their state from the cluster, not the store.

At time of writing there is a bug that requires the stores of all but the first broker to be deleted manually before starting the cluste

Limitations of current design

There are several limitations of the current design.

Concurrency: all CPG events are serialized into a single stream and handled by a single thread. This means clustered brokers have limited ability to make use of multiple CPUs. Some of this work is pipelined, so there is some parallelism, but it is limited.

Maintainability:  decoupling the cluster code from the broker and assuming the broker behaves deterministically makes it very easy for developers working on the stand-alone broker to unintentionally break the cluster, for example by adding a feature that depends on timers. This has been the case in particular for management, since the initial cluster code assumed only the queue & exchange state needed to be replicated, whereas in fact all the management state must also be replicated and periodic management actions must be co-ordinated.

Non-replicated state: The current design replicates all state. In some cases however, queues are intended only for directly connected clients, for example management queues, the failover-exchange queues. It would be good to be able to define replicated and non-replicated queues and exchanges in these cases.

Scalability: The current cluster design only addresses reliability. Adding more brokers to a cluster will not increase the cluster's throughput since all brokers are doing all the work. A better approach would move move some of the work to be done only by the directly-connected broker, and to allow messages to "bypass" the cluster when both producer and consumer are connected to the same member

This design discusses clustering at the AMQP protocol layer,
i.e. members of a cluster have distinct AMQP addresses and AMQP
protocol commands are exchanged to negotiate reconnection and
failover. "Transparent" failover in this context means transparent to
the application, the AMQP client will be aware of disconnects and must
take action to fail over.

More precisely we define transparent failover to mean this: In the
event of failover, the sequence of protocol commands sent and received
by the client on each channel excluding failover-related commands is
identical to the sequence that would have been sent/received if no
failure had occured.

Given this definition the failover component of an AMQP client library
can simply hide all failover-related commands from the application
(indeed from the rest of the library) without breaking any semantics.

Requirements

TODO: Define levels of reliability we want to provide - survive one
node failure, survive multiple node failures, survive total failure,
network partitions etc. Does durable/non-durable message distinction
mean anything in a reliable cluster? I.e. can we lose non-durable
messages on a node failure? Can we lose them on orderly shutdown or
total failure?

TODO: The requirements triangle. Concrete performance data.

Clients only need to use standard AMQP to talk to a cluster. They need
to understand some AMQP extensions to fail-over or resume sessions.
We will also use AMQP for point-to-point communication within the
cluster.

Ultimately we may propose extensions to AMQP spec but for the initial
implementation we can use existing extension points:

• Field table parameters to various AMQP methods (declare() arguments etc.)
• Field table in message headers.
• System exchanges and queues.

Abstract Model and Terms

A quick re-cap of AMQP terminology and introduction to some new terms:

A broker is a container for 3 types of broker components: queues,
exchanges and bindings. Broker components represent resources
available to multiple clients, and are not affected by clients
connecting and disconnecting. Persistent broker components are
unaffected by shut-down and re-start of a broker.

A _client_ uses the components contained in a broker via the AMQP
protocol. The _client components_ are _connection_, _channel_, _consumer_ and
_session_\[[Footnote(The "session" concept is not fully defined in AMQP 0-8 or 0-9 but is under discussion. This design note will define a session that may be proposed to AMQP.)]. Client components represent the relationship between a client
and a broker.

TODO: _Where do transactions fit in the model? They are also a kind of
relationship components but Dtx transactions may span more than just
client/broker._

A client's interaction with a unclustered _individual broker_\[[Footnote(An individual broker by this definition is really a broker behind a single AMQP address. Such a broker might in fact be a cluster using technologies like TCP failover/load balancing. This is outside the scope of this design, which focusses on clustering at the AMQP protocol layer, where cluster members have separate AMQP addresses.)] is
defined by AMQP 0-8/0-9: create a connection to the brokers _address_,
create channels, exchange AMQP commands (which may create consumers),
disconnect. After a disconnect the client can reconnect to the same
broker address. Broker components created by the previous connection
are preserved but client components are not. In the event of a
disorderly disconnect the outcome of commands in flight can only be
determined by their effects on broker components.

A _broker cluster_ (or just _cluster_) is a "virtual" broker implemented
by several _member brokers_ (or just _members_.) A cluster has many AMQP
addresses - the addresses of all its members - all semantically
equivalent for client connection \[[Footnote(They may not be equivalent on other grounds, e.g. network distance from client, load etc.)]. The cluster members co-operate to
ensure:

• all broker components are available via any cluster member.
• all broker components remain available if a member fails, provided at least one member remains active.
• clients disconnected by member failure or network failure can reconnect to another member and resume their session.

A session is an identity for a collection of client components (i.e. a
client-broker relationship) that can outlive a single connection. AMQP
0-9 provides some support for sessions in the `resume` command.

If a connection is closed in an orderly manner by either end, the
session is also closed. However if there is an abrupt disconnect with
no Connection.close command, the session remains viable for some
(possibly long) timeout period. The client can reconnect to a failover
candidate and resume.

A session is like an extended connection: if you cut out the failover
commands the conversation on a session is exactly the conversation the
client would have had on a single connection to an individual broker
with no failures.

If a connection is in a session, events in the AMQP.0-8 spec that are
triggered by closing the connection (e.g. deleting auto-delete queues)
are instead trigged by the close (or timeout) of the session.

Note the session concept could also be used in the absence of
clustering to allow a client to disconnect and resume a long-running
session with the same broker. This is outside the scope of this
design.

Cluster State

We have to consider several kinds of state:

• Wiring: existence of exchanges, Queues and bindings between them.
• Content: messages on queues and references under construction.
• Cluster membership: list of members
• Conversational state: Sessions, channels, prefetch windows, consumers, acks, etc.

Wiring and content needs to be replicated in a fault tolerant way
among the brokers so all brokers can provide access to all
resources. Persistent data also needs to be stored persistently for
recovery from total failure or deliberate shutdown.

Cluster membership needs to be replicated among brokers and passed to
the client. The client needs to know which members are candidates for
failover.

Conversational state relates to a client-broker relationship:

• session identity.
• open channels, channel attributes (qos).
• consumers.
• commands "in flight" - sent but not acknowledged.
• acknowledged command history.

To resume successfully the converstaional state must be re-established
on both sidess. There are several options about how much state is
stored where. We'll outline the solution that minimizes broker-side
replication, but it's not clear yet if this is the optimal choice.

To minimize converstaional replication on the broker, the broker must
replicate at least:

• session identities: to recognize the session on reconnect.
• history of acknowleded commands: to ignore duplicates resent by client.
• commands in flight: to resend to client and/or handle client responses.

Everything else can be re-established by the client:

• re-open the channels.
• re-set qos settings.
• re-create consumers.
• ignore incoming duplicates during reconnect.
• re-send all unacknowledged commands.

Note that all of the above can be accomplished with normal AMQP commands.

The broker could replicate more converstaional state to relieve the
client from re-creating it. It's not clear yet where the best tradeoff
lies since it's not possible to have 0 conversational state on the
broker. Minimizing the brokers role seems like a good approach since
replicating data affects the whole cluster, keeping it on the
client affects only one client-broker connection.

Replication

The different types of state are replicated in different ways to
strike the best performance/reliability balance.

Cluster membership and Wiring

Membership changes are replicated to entire cluster symmetrically
using a virtual synchrony protocol.

Connected clients are also notified of changes to the set of fail-over
candidates for that client. Clients are notified over AMQP by binding
a queue to a special system exchange.

Wiring is low volume so and can also be replicated via virtual
synchrony. Cluster membership + wiring make up the common "picture"
that every member has of the cluster.

Queue Content

Message content is too high volume to replicate to the entire cluster.
To limit the extent of replication each queue or reference plays
exactly one of the following roles in each broker:

• Primary: content owner, all modifications are done by primary.
• Proxy: Forwards client requests to the primary.
• Backup: A proxy that also receives a replica of the content and can take over if the primary fails.
• Fragment: special case for shared queues, see below.

A single cluster member may contain a mix of primary, backup and proxy
queues. (TODO: mechanism for establishing primary, backup etc.)

The client is unaware of the distinction, it sees an identical picture
regardless of what broker it connects to.

TODO: Ordering issues with proxys and put-back messages (reject,
transaction rollback) or selectors.

Fragmented Shared Queues

A shared queue has reduced ordering requirements and increased
distribution requirements. "Fragmenting" a shared queue is a special
type of replication where each of a set of brokers holds a disjoint
subset (fragment) of the messages on the queue. The idea is to
distribute load over independent fragments hosted in separate brokers.

Each fragment's content is replicated to backups independently just
like a normal queue.

The appearance of a single queue is created by collaboration between
fragments. Fragments store incomging messges in the local queue, and
server local consumers from the local queue whenever possible. Only
when a fragment cannot satisfy its consumers does it consume messages
from other fragments in the group.

Proxies to a fragmented queue will consume from the "nearest" fragment
if possible.

TODO: Proxies can play a more active role. Ordering guarantees, we can
provide "same producer to same consumer preserves order" since
messages from the same producer always go on the same fragment
queue. May break down in the presence of failover unless we remember
which fragment received messges from the client and proxy to the same
one on the failover replica.

Conversational State

The minimum converstaional state a broker needs to replicate for failover is

• session identities: to recognize the session on reconnect.
• history of processed request-ids: to ignore duplicates resent by client.
• all requests in flight: to resend to client and/or handle client responses.
• open incoming references.

Session identities and processed requiest history is very low volume
and could be replicated with virtual synchrony. However commands in flight
and open incoming references are probably too large to be replicated this
way.

Enter the session backup: the broker directly connected to a client is
the session primary. The session primary has one or more session backup
brokers. Conversational state is replicated only to the session
backups, not to the entire group. On failure the client must reconnect
to one of the session backups, other members will not accept the
session.

Shared storage

In the case of high-performance, reliable shared storage (e.g. GFS)
queue content can be stored instead of replicated to backups. In that
case the conversations with content backups can be dehydrated, on
failover the backup can recover data from shared store.

For commodity hardware cases we need a solution with no shared store
as described above.

TODO: Would we wan to use shared store for conversational state?

Client-Broker Protocol

TODO: How it looks in client protocol terms - system queues &
exchanges, connect argument tables.

Membership updates, choosing a replica, reconnection,
building conversational state => client stores representation of state
unacked messages & duplicates, hiding failover.

Broker-Broker Protocol

Broker-broker communication uses normal AMQP over specially identified
connections and channels (identified in the connection negotiation
argument table.)

Proxying: Proxies simply forward methods between client and primary
and create consumers on behalf of the client. TODO: _any issues with
message put-back and transactions?_

Queue/fragment replication: Use AMQP transfer commands to transfer content
to backup(s). TODO: propagating transactions.

Session replication:
must replicate a command (and get confirmation it was
replicated) before responding. However this can be done in async
streams - forward commands to replica

Session replication: AMQP on special connections. Primary forwards all
outgoing requests and incoming responses to session backup. Backup can
track the primary request/response tables and retransmit messages.

Note the
dehydrated requests to the session backup should reference the content
backup so the backup can recover content in a failure. Alternatively
content could be sent to both - what's the tradeoff?

Persistence and Recovery

Use cases for persistence

Durable messages: must be stored to disk across broker shutdowns.

Flow-to-disk: to cope with memory overflow on full queues.

Reliability: recover after a crash.

We need a common solution - it would make no sense to be writing the
same message to disk multiple times for different purposes. The
solution may be tunable to offer better performance for individual
cases but should cover the case where all 3 are required.

Competing failure modes:

Tibco: fast when running clean but performance over time has GC
"spikes" Single journal for all queues. "holes" in log have to be
garbage collected to re-use the log. 1 slow consumer affects everyone
because it causes fragmentation of the log.

MQ: write to journal, write journal to DB, read from DB.
Consistent & reliable but slow.

Street homegrown solutions: transient MQ with home grown
persistence. Can we get more design details for these solutions?

Persistence approach

Try to avoid the problems with:

• journal-per-queue
• multiple segment files per journal
• GC to "snapshot" (per queue? per broker? think about queue migration.)
• Disk thrashing - how do we distribute the parts to avoid it?

Need to think about how we consolidate journal into a "snapshot" or
"database" to reclaim space.

No write on fast consume: Optimization - if we can deliver (and get
ack) faster than we write then no need to write. How does this
interact with HA?

Async journalling: writing to client, writing to journal, acks from
client, acks from journal are separate async streams? So if we get
client ack before the journalling stream has written the journal we
cancel the write? But what kind of ack info do we need? Need a diagram
of interactions, failure points and responses at each point. Start
simple and optimize, but dont rule out optimizations.

What about persistence-free reliability?

Is memory-only replication with no disk a viable option for high-speed
transient message flow? We will lose messages in total failure or
multiple failures where all backups fail, but we can survive single
failures and will run a lot faster than diskful.

Virtual synchrony

TODO: Wiring & membership via virtual synchrony
TODO: journaling, speed. Will file-per-q really help with disk burnout?

Configuration

Connection issues: potentiall for big N*N connection web if primaries
and backups are not grouped.

Simplifying patterns:

• Virtual hosts as units of replication.
• Backup rings: all primary components in a broker use the same backup broker and vice-versa. Backups form rings.
• Disk management issues?
• Shared storage issues?

Dynamic cluster configuration

• Failover: the primary use case.
• Add node: backup, proxy, primary case?
• Redirect clients from loaded broker (pretend failure)
• Move queue primary from loaded broker/closer to consumers?
• Re-start after failover.

Issue: unit of failover/redirect is connection/channel but "working
set" of queues and exchanges is unrelated. Use virtual host as unit
for failover/relocation? It's also a queue namespace...

If a queue moves we have to redirect its consumers, can't redirect
entire channels! Channels in the same session may move between
connections. Or rather we depend on broker to proxy?

Backups: chained backups rather than multi-backup? Ring backup?
What about split brain, elections, quorums etc.

Should new backups acquire state from primary, from disk or possibly
both? Depends on GFS/SAN vs. commodity hw?

Open Questions

Issues: double failure in backup ring: A -> B -> C. Simultaneous
failure of A and B. C doesn't have the replica data to take over for A.

Java/C++ interworking - is there a requirement? Fail over from C++ to Java?
Common persistence formats?

Implementation breakdown.

The following are independently useful units of work that combine to
give the full story:

Proxy Queues: Useful in federation. Pure-AMQP proxies for exchanges
might also be useful but are not needed for current purpose as we will
use virtual synchrony to replicate wiring.

Fragmented queues: Over pure AMQP (no VS) useful by itself for unreliable
high volume shared queue federation.

Virtual Synchrony Cluster: Multicast membership and total ordering
protocol for brokers. Not useful alone, but useful with proxies and/or
fragments for dynamic federations.

Primary-backup replication: Over AMQP, no persistence. Still offers some
level of reliability in a simple primary-backup pair.

Persistence: Useful on its own for flow-to-disk and durable messages.
Must meet the performance requirements of reliable journalling.

<muse:fn-sep?/> Exclusive or auto-delete queues are deleted on disconnect, we'll return to this point.