Details
-
Improvement
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
Description
== Background ==
I've been looking at what would be required to get AMQP 1.0 support in
the qpidd broker (using proton-c). In that context I felt there was a
need to refactor the broker code, particularly that part that would be
shared between different protocol versions. Part of the motivation was
clearer separation of 0-10 specific logic, so that 1.0 logic could be
introduced as an alternative. However part of it was also simply the
recognition of some long-standing problems that we have never stopped
to address.
So, here is a patch representing my ideas on what is needed. This is
a little stale again (patch was generated against r13613342) and
needs (yet) another rebase. However it is getting to the point where I'll be asking to commit it soon, so if anyone has feedback, now is the time to give it!
== Key Changes ==
qpid::broker::Message
This is now supposed to be a protocol neutral representation of a
message. It no longer exposes qpid::framing::FrameSet. It can be based
on data received in different encodings (this patch only includes the
existing 0-10 encoding).
The immutable, sharable state is separated from the mutable
queue-specific state. Messages themselves are no longer held through a
shared pointer but are passed by reference or copied if needed. The
immutable state (essentially the data as received) is still shared
and referenced internally through an intrusive pointer. There is no
longer a message level lock. A message instance is 'owned' by
someother entity (usually the queue it is on) which controls
concurrent access/modification if necessary.
The persistence context is a separate part of the message
also. Currently that can be shared between two message instances if
desired.
qpid::broker::Messages
Switched from using qpid::broker::QueuedMessage (which relied on
shared pointer to message itself and made sequence number the explicit
- and only - way to refer to a specific message) to using modified
Message class directly and a new qpid::broker::QueueCursor.
The cursor is opaque outside the Messages implementation to which it
relates. It provides a way to refer to a specific message (without
directly using sequence number, though at present that is what is used
'under the covers') and/or to track progress through a sequence of
messages (for consumers or other iterating entities).
I.e. its an iterator to a Message within its containing Messages
instance that is not invalidated by changes to that container.
A Messages instance owns the Message instances within it. Other
classes access this through a reference or (raw) pointer, or if needed
copy it (the immutable part can be - and is - safely shared).
The codepath for browse/consume is a lot more unified now. You use a
cursor and call Messages::next() in each case. This also lays the
foundation for selectors.
The simplified Messages interface led to a simplied
MessageDistributor. There is still a little more to do to clarify
these separate roles (or indeed perhaps unify them?) but more on that
later.
qpid::broker::amqp_0_10::MessageTransfer
This represents the familiar 0-10 encoding of a message. This class is
broadly similar to the old Message class, based on a FrameSet. However
it represents the shared and essentially immutable state. The
sendHeader() method now explicitly takes a copy of the original
headers and adds to it or otherwise modifies it if needed (e.g. for
redelivered flag, ttl, annotations etc).
[Ideally I'd like to move more of the 0-10 specific classes out of
qpid::broker and into qpid::broker::amqp_0_10, but that has no
functional relevance so I've left existing classes alone for now.]
qpid::broker::Consumer
The deliver() method now takes a QueueCursor (representing a 'handle'
to this message for use in subsequent operations such as accept,
relese etc) and a constant reference to the Message itself
(i.e. consumers can't alter the state of the message on the queue
directly, but only through operations on the queue itself).
qpid::broker::QueueRegistry
The actual queue creation has been pulled out into a base class,
QueueFactory. The actual class of the Queue returned can now be varied
and there are two subclasses in the current patch. The first is a
replacement for the ring policy logic, whereby messages are removed
from the queue in order to keep the queue from growing above a
configured limit. The second is for last value queues and simply pulls
the special case behaviour out of the common code path.
The handling of queue configuration has also been made cleaner and
more uniform, based on the QueueSettings class.
qpid::broker::QueuePolicy
This class has been removed. There is a new QueueDepth utility used
for configuring limits, tracking current depth and testing the latter
against the former. This is used directly by Queue. The behaviour at
the limit can be varied by subclassing queue.
== Limitations etc ==
clustering
This breaks clustering. Indeed it will not compile unless clustering
is disabled (--without-cpg in configure). Keeping the cluster code in
sync was distracting me from the core goal, given its entanglement
with the broker code.
My assumption is that the new ha code will eventually replace the
cluster anyway and the amount of change that would be required to get
the cluster working with this refactor may not be worth it and may in
fact undermine its stability anyway (which seem the only good argument
for using it).
I don't believe there is anything insurmountable to do to re-enable
cluster if that was desired however, just a fair chunk more work that I would rather not do if I can avoid it.
old & nasty features removed
I have removed support for flow to disk, the legacy version of lvq
with two modes (the updated version of lvq is of course still
functional), the last-man-standing persistence in clustering and the
old async queue replication. They are really quite horribly
implemented and/or are no longer necessary in my view.