[Twisted-Python] Re: communication idioms with Perspective Broker
db3l at fitlinxx.com
Thu Jul 21 12:40:25 EDT 2005
Antony Kummel <antonykummel at yahoo.com> writes:
As a general comment, I would suggest that you may wish to consider if
what appears to be your goal of keeping asynchronous activities
completely in sync is really a constraint you must meet?
That is, your points seem concerned with ensuring that asynchronous
actions or state changes somehow are reflected to users of that state
in some atomic manner or on a transactional basis (no changes exposed
until everything is complete). This is a non-trivial problem in a
distributed form, particularly if you start to consider network
failures that can interfere with communication mid-operation. It
would appear you're looking for a ACID-like transactional layer.
Instead of trying to meet such a tough constraint, it may be better to
see if you can design your system to accept, from the beginning, that
there will be a period of latency between actions taken in the system
and resultant changes to state being apparent to users of that state.
Or, that different parts of the system may be operating with older
(but still valid) data at various points in time, and take more
explicit control of when they update to the latest data. This has its
own form of complexity, but can yield a much more resiliant system.
> 1. Synchronization of events and state/data changes: I want clients
> to be able to receive notifications from a server when certain
> events occur. When such events are related to changes in state or
> data of the server that are also accessible to the user, I want the
> client's interface for retrieving the data and getting notified of
> the event to be coherent.
> The solution I came up with is to always expose associated events
> and data using a single cacheable object, so that when there is a
> change in state/data that should also make an event fire, the data
> change is propagated to the remote cache, and the remote cache fires
> an event locally (at the client's side). This way, the client's
> representation of the server's state is always coherent.
This seems reasonable to me, or at least as close as you can come to
your requirement of keeping a remote client's data/events in sync.
You would need to ensure that your state object managed changes to
itself such that it only reflected changes down the PB channel to any
cacheable observers at appropriate times when its own state was
consistent. E.g., you wouldn't necessarily update on each field or
What you're effectively doing here is using the cacheable object as
your transaction manager, controlling what clients see.
Note that to ensure consistency between clients and server, you'd want
to have everyone working from cacheables of this object and nobody
(even on the server itself) working from the object itself. Otherwise,
you'd need a mechanism (such as with point 2)
You still have an issue of network outages preventing updates from
making it to clients, but in that case you'd only have to deal with a
disconnected client being out of date with respect to the server and
other clients, but in a consistent way, and upon reconnecting it would
receive a new, but again consistent, set of state information.
> 2. Synchronization of remote commands and remote events/data
> changes: I want clients to be able to issue commands to a server
> that make its state or associated data change, and also to have an
> up-to-date representation of the state or data (and possible events
> issued by changes in the state/data). The problem is how to
> synchronize the firing of the deferred returned by the remote method
> call with the change in the client's representation of the server's
> state/data (i.e. to make sure that when the deferred fires, the
> client's representation of the state is coherent with the command,
> i.e. changed).
> Presently I have no solution for this, but I'm pretty sure that it
> requires some combination of referenceable and cacheable that will
> insure one coherent interface to associated data, events and
> commands. I am thinking to achieve this with a copyable that will
> contain both a referenceable and a cacheable.
I think you're going to have an uphill battle here to try to
synchronize what is inherently distributed and asynchronous behavior.
If I absolutely had to do this, it's probably along the lines of what
your considering - I'd consider having a transaction object that
encapsulated change requests and state updates under a single
umbrella. Rather than a simple cacheable though, you'd probably need
to implement a two-phase commit protocol so that the state only
changed on both client/server or not at all. If you didn't do that,
you'd leave yourself open to the operation occurring on the server,
but the network preventing the new state information from getting down
to the client, and the client not knowing what state the server was in.
But I'd much rather just assume that the client needed to work
properly with the state as it currently knew, and respond properly to
any state updates as they occurred, whether due to its own operation
or some other client's operation. E.g., formally decouple the request
to perform an operation, from the state changes that would occur as
that operation was performed.
In other words, try to stick more to a model/controller approach,
where the model state is monitored by clients and they simply react to
its changes, but actions taken always flow through a distinct
controller path that is associated with but decoupled from the model.
> 3. Wrapping remote objects for non-online representation. I want to
> be able to have objects that can be used locally, but may also
> provide an interface to one or more servers that have to do with the
> state of the object. Another reason for these wrappers is that I
> want to be able to pass them freely from process to process, without
> being dependent on the connection or on a specific server.
> The question is, how to represent this to the user of the
> object. Whether to allow access to cached remote data and
> subscription to events when not online, whether to return a deferred
> and try to connect in case we don't have the data or throw an
> exception, etc. I suppose this is mostly a matter of style, but if
> anyone has done something like this before maybe they would have
> some insight
What we've done in a system of ours that is designed to be an
distributed data system is treat the core data objects in the system
as pure state objects as much as possible, generally falling into two
* Pure instance data (data objects) of which multiple copies may exist
simultaneously throughout the system, but which operate locally
(Copyable in twisted). The only methods such objects have are to
manipulate the local representation (also handled by direct
* Shared state objects (model objects) of which multiple observers
throughout the system may exist (Cacheable in twisted). While these
objects may have similar method and attribute access as the above
instance data, they are only ever manipulated by controller objects,
for which the original instance is only the same "node" as the
original instance of the state object. Any client needing to make
changes must use a reference to that controller (a twisted
Referenceable) and not the model itself, even on the local node
where the model/controller are instantiated.
The choice between the two object types is not a hard and fast rule -
we've tended to use more of the former so far.
We then constructed a framework of "manager" objects which are
designed to provide access and manipulation of the above data objects.
The key attribute of a manager object is that it is both referenceable
(twisted Referenceable) and all of its methods are deferrable
interfaces - even if used locally.
Although I found I had a desire to try to somehow just always pass
object references (referenceable) around to everything and let PB
handle everything transparently - in practical terms I didn't find
that workable. There are just too many nooks and crannies you can get
into as things get distributed that taking some more explicit control
became necessary to ensure robustness. And you just can't always
assume that making changes to state on what is ostensibly a shared
object will magically get reflected everywhere reliably. But that's
actually where we found PB to be just at the right level as it was
easy enough to wrap to behave how we wanted.
We provided an extra layer of wrapping for the networking, both for
the basic connection as well as our referenceables. A Client object
encapsulates making a connection to a server, along with reconnecting
as necessary and generating local signals (we use pydispatcher) on
connection state changes. A matching Server object provides client
access to local managers - through a simple Registry object - upon a
A general purpose wrapper object is used to wrap each manager
referenceable so that they appear to be local (it uses the manager's
interface definition to automatically translate method calls into
callRemote), as well as to automatically re-establish contact with the
remote manager if needed, by listening for the Client signals and on a
reconnect, re-obtaining the remote Registry handle and refetching
remote references to the manager it had previously wrapped.
The wrapping may be multi-layer. This allows us for example, to have
a remote site have a master server, which maintains the client link to
a central server. That site server, therefore, has what it considers
a local Registry with a whole set of local managers - all of which are
technically wrappers around the twisted referenceable to the main
central server. But then other machines at the site themselves become
clients of the site server, with their own references to the site
server's references. So when the other site machines make requests
they flow to the site server, and then up to the central server and
back down. But the same source code works whether running on the
central server, or at any level of the site servers, without knowing
the difference (since all registry and manager interfaces are
While the wrapper isolates users from need to worry about
reconnecting, we don't attempt to hide the fact that an outage is
occurring. Attempts to make a call on a wrapped manager during an
outage generates normal PB exceptions, with one change that we modified
Twisted to always return exceptions up the deferred change (even for a
dead reference) so clients wouldn't have to deal with both local
exceptions and errbacks.
In practice, the client applications generally have some application
level object that is also listening to the Client objects connection
signals, and either blocking access to the user when the network is
(with an appropriate message), or adjusting behavior accordingly.
So, in operation, client code works something like:
* Instantiate a Client object, give it connection info and start the
connection. Request the registry from the client object (which is
deferrable and only fires once the overall connection cycle is
* Using the Registry object (which itself is a remote wrapper version
on the client side), query references for any manager objects
* Using the manager objects, retrieve any data objects needed.
Changes to model objects occur through their controllers, while
changes to data objects are performed locally and changes updated
via explicit "save" calls to the managers.
The last point is where we run into similar issues as yourself, I
think. By choosing this route we do not provide for other clients of
that same data object to automatically see changes made by other
clients. They would continue to run with the copy they had previously
received, although any subsequent retrieval would get a new copy with
the new data.
To handle crossing state changes, the originator (actual manager
object on whatever node it exists on) of the data object maintains an
internal tag (we use a UUID but it could also be a hash of the
contents) in the object representing its unique state and will raise a
SaveConflict exception of our own if someone else attempts to store
changes to an outdated copy. It is up to clients to handle such
issues, should they occur (typically by requerying the information and
then re-applying their changes), although in practice we really don't
have scenarios where this happens yet due to typical usage patterns.
Some of this could change if we moved a data object to a model object,
but then we're requiring that even simple users of the data object
maintain a remote cacheable reference to the object, which is
relatively heavyweight. Thus my comment about it being a grey area
above as to which sort of object we decide to place such state in. In
our environment, our User object (which contains identifying and
control information about users) is just a data object, as the need
for simultaneous manipulation and monitoring of it is reasonably low.
We do expect to have many copies of it around, but mostly on a
In your context, I would think that the user object itself need not be
something that constantly updates, but the state of which users were
currently online would fit better as a model (and the controller to
feed it would have methods for a given user to go online or offline).
In our structure, we would separate out the concept of generating a
system message - probably into a messaging manager - which would then
receive requests to transmit messages to identified users. But I
don't think I'd try to tie those three things (current User object
contents, currently online user set, generating a message) into any
sort of guaranteed state ... I'd leave them very loosely coupled.
On the issue of distribute events, that's the area we're currently
working on, and to us the hardest part is how to handle events that
may be generated during outages for which the disconnected clients
have subscriptions. If it's just for changes to state objects (such
as cacheables) that's not so bad, since the reconnection process will
re-query the current state information. But if it's for more general
notifications (we might have our own bit for "user updated" like your
"user came online") you have a question of how long do you queue up
such events for clients that might never show up again.
Currently we are targetting such events being handled by a signal or
event manager, which will maintain an ongoing history of such events.
Subscribers to the event manager will get copies of appropriate
events. When a client connects it's local wrapper for the remote
event manager will actually handle all local subscriptions,
maintaining a single remote set of subscriptions to minimize network
I/O. It will also track the delivery of any events. Upon being
disconnected/reconnecting (per the standard mechanisms), the client
event manager wrapper will request any signals that may have been
generated since the last event seen prior to the disconnect. We'll
have to bound this somehow for prolonged outages. But a key point is
still decoupling the event handling from other operations, and we
won't be trying to force everything to stay in sync with other clients
and/or servers at all times.
If you've put up with me until here, I hope that this at least gives
you some other approaches to think about, even if some or all of it
isn't directly applicable to your problem domain.
More information about the Twisted-Python