W3C home > Mailing lists > Public > public-ws-chor@w3.org > March 2003

RE: Choreography and Orchestration

From: Assaf Arkin <arkin@intalio.com>
Date: Sat, 15 Mar 2003 17:52:37 -0800
To: "Ricky Ho" <riho@cisco.com>, <public-ws-chor@w3.org>
I want to offer an alternative definition.

I like to talk in terms of generic models that do not exclude any particular
type of application, before drilling down into a specific application. When
you talk in generic terms it may seem hard at first to understand 'how my
specific problem is solved in that case', but the benefit is that it allows
more problems to be solved, not just specific ones.

For example, an OO language is one that allows you to define the common
behavior of multiple objects, and allows the behavior to relate to data
owned by the object (or referencing other objects). Another definition is
that an OO language allows me to build business objects. What about
non-business objects like a window or a counter? What about utility objects
like a line item list or an address object? A generic definition has more
utility then an application-specific definition.

The term choreography was interesting because it was borrowed from
performance arts where it usually refers to a dance or a ballet. It places
no restriction on how many dancers may participate in the dance. Tango is a
choreographed dance that has just two dancers. So while Choreography is a
good term for any dance involving any number of dancers, if there is an
intent to allow only two dancers, I would prefer to use Tango instead.

In OO we talk about objects and classes. We don't talk much about class
instances. Similary in the WS world we talk about service and interface. We
don't talk about service instance (though such a thing does exist), nor do
we talk about service being an interface instance. Similarly, I would like
to propose that we distinguish between Choreography as the definition and
Conversation as what actually happens when these services communicate with
each other. While we're using two different terms, I think it will actually
reduce some of the confusion when we get to discuss more complex use cases.

So what is a choreography: a choreography is a specification of how multiple
things interact with each other over space and time.

In our case space means different services in different places, time means
the ordering of activities with relation to each other. I would like to
suggest the following definition of WS Choreography (as opposed to just
about any choreograph):

WS Choreography: A specification of the behavior of multiple servies that
interact with each other by exchanging messages

You will notice that in my presentation I've tried, and hopefully succeeded,
to make that distinction clear. I was talking about a generic concept first,
and then tried to place it in the scope of WS. So I never imagined WS as the
only way to choreograph things, but I do think that WS-Choreography would
talk only about Web services, as defined by the WSA. This is also important
because there are some choreography definitions that cannot be defined by
WS-Choreography and in my opinion are not interesting for us, e.g. updating
shared states in a database, or using some non-WS means to do that. A load
balancing engine has a way to manage shared states that involve multiple
servers, but since this is done below the WS level it's not exposable in
WS-Choreography (and quite frankly not of interest to this group).

Let's assume that two (or more) Web services are performing a dance with
each other based on some WS Choreography scripted for them by some language.
What is each of these services doing? Each of these services is talking to
other services. It's arranging or controling elements (in this case WS
operations) that allow it to achieve some overall effect or goal. That's
pretty much the dictionary definition of orchestration (conducting a
symphony is another unrelated definition).

So I would like to propose a definition for WS Orchestration that does
something like:

WS Orchestration: A specification of how a service interacts with other
services by exchanging messages

As I pointed out on the whiteboard, a choreography is a combination of
orchestrations. Let's look at a simple example using a bi-party

Service A sends some message (m) to service B. Following that, service B
sends some message (m1) to service A. That's a very simple choreography of
how two services interact with each other. It's also a specification of
their orchestration. Service A sends message m0 and waits to receive message
m1. Service B receives message m0 and then sends message m1.

Let's look at a more complex example involving multi-party choreography with
services A, B, C and D. In the choreography we define that service A sends
messages to B and C. C sends messages to D. So we also have a definition of
what the orchestration in this context would look like: A talks to B,C. B
talks to A. C talks to A,D. D talks to C. Magically, the definition of
choreography is nothing more than the composed definition of orchestrations,
and as I'll show below vice versa.

But there's an interesting question. In defining the orchestration of some
service, say service D in the example above, did we also define the
implementation? Not necessarily.

Let's say that service D receives a message from C, and based on some
information sends back a message to D. How did D get that information? It
can have some piece of code in there that does it, whether we write it in
Java, BPEL or Perl makes no difference. We've expressed it's orchestration
but not it's implementation. In this case, which covers the majority of
cases, we've expressed the service interface, one coarse grained
orchestration is must perform to participate in the choreography, but in no
way did we express its implementation.

So I would also like to formulate three axiomatic restrictions:

1. Closure: A choreography of n service types includes only messages the are
exchanged by these n service types.

2. Completeness: Any orchestration for one service type that involves n
other service types can be expressed as a choreography of n+1 service types.

3. Non-reflexive: An implementation is a super set of some (possibly more
than one) orchestration.

These restrictions are important because they allow us to attain some of the
properties of choreography that we all care about, namely:

1. Abstraction: A choreography asks the implementation to do certain things,
but does not strive to restrict the number of possible implementations. I
argue that a choreography model that does not observe restriction #1 or #3
is not useful.

2. Validation: Since the choreography expresses a subset of the
orchestration of a service (#2 and #3), it allows us to determine that the
service actually observes it's role obligations in the choreography,
regardless of how we choose to implement it. (In other words, we can ensure
that the implementation matches the contract, but we don't decide on what
the implementation looks like and we can change it as often as we like)

3. Utility: Since we placed no other restrictions we can allow a service to
participate in any number of choreographies, and we can allow choreographies
that are applicable to B2B, choreographies that are applicable to A2A,
simple back & forth and complex end-to-end, reuse and recursive composition,

Note that throughout this discussion I've been talking about services and
not business entities. How we map services to business entities is
orthogonal. For example, a choreography involving n service types may in
fact be a choreography involving two business entities. But since one of
these business entities elected to reuse multiple service types in that
choreography the choreography involves more than two service types. This
makes choreographies more interesting because they support reuse, recursive
composition and more than a limited set of scenarios.

Also I did not limit bindings intentionally. In WS terms we talk about
service types (or interfaces) and then bind services during the message
exchange, possibly passing service references in the messages. Practical
examples may be more restricting, depending on what you want to achieve. In
some cases services must be fixed as soon as possible and never changed, in
other cases the "interesting" services are those that are bound later on
(e.g. marketplace scenario). There is no technical restriction that imposes
that all services be bound in advance, and a variety of mechanisms that
allow them to be bound later on (WSDL, UDDI, WS-Addressing, WS-Callback,
WSCI's locator, etc).

I also did not express choreography as a sequence of state alignments for a
very simple reason. Choreography is a solution for addressing state
alignment, and allows a variety of models (e.g. point-to-point, group
consensus, transaction contexts, state expiration). On the other hand, a
language for expressing state alignment may be useful but does not specify
how WS are used to achieve this. It is quite likely that two services would
have a perfect understanding of the states they want to align, but not a
perfect understanding of how they intend to do that, resulting in misaligned
states. On the other hand, a choreography gives a framework for expressing
the deterministic interaction that leads to proper state alignment for any
number of executions.


  -----Original Message-----
  From: public-ws-chor-request@w3.org
[mailto:public-ws-chor-request@w3.org]On Behalf Of Ricky Ho
  Sent: Saturday, March 15, 2003 8:10 AM
  To: public-ws-chor@w3.org
  Subject: Choreography and Orchestration

  I try to put up my own definition of "Choreography" and "Orchestration"
and use a simple buyer/seller use case to illustrate what I mean.
  I'm particularly interested to see how the "Choreography" portion of this
simple example get represented by WSCI and BPSS.

     1       Definitions

  1.1     Choreography
  CHOREOGRAPHY defines the public part of a bi-lateral interaction between
two communicating parties.  It formalize a contractual agreement between
these parties.

  CHOREOGRAPHY defines TWO communicating parties in terms of ROLES, which
will be bound to the actual business entity when the choreography instance
starts.  The binding doesn’t change throughout the lifecycle of the

  CHOREOGRAPHY defines a set of “SHARED STATES” between the TWO
communicating parties.

where one ROLE sends a message to another ROLE.  In other words, the purpose
of MEP is to align the SHARED STATES between the two ROLES.

  CHOREOGRAPHY does NOT reflect the perspective of a single party.  It can
be taken by any parties who wants to play a role within it.

  The CHOREOGRAPHY INSTANCE starts when the following occurs
  •       One party sends the first message (which propose the initial
SHARED STATES) to another party.
  •       This another party verifies that the initial SHARED STATES meets
the pre-requisite to start the CHOREOGRAPHY

  1.2     Orchestration
  ORCHESTRATION defines the private part of the implementation of a
particular party who plays a ROLE in the CHOREOGRAPHY.  It formalize the
execution logic of that party throughout the message exchanges.

  ORCHESTRATION realize a particular ROLE of a CHOREOGRAPHY.  Therefore,
ORCHESTRATION needs to be conformed with the CHOREOGRAPHY.

  ORCHESTRATION can potentially span across multiple CHOREOGRAPHIES.
Therefore, CHOREOGRAPHY INSTANCES can form inter-dependent relationship at

  Note here that I try to restrict choreography to 2 parties and disallow
changes of role binding throughout the lifecycle of choreography instance.
The downside is now a multi-party interaction needs to be broken down into
multiple bi-lateral choreographies and their inter-dependencies is not
externalized at the choreography level.  It is up to the implementation
(which is the orchestration) to determine such inter-dependencies.  The
purpose of these restrictions is to simplify the choreography model which I
think still address 80% of the real life use cases.  I would like to see
where it breaks before remove this restriction.

  2       Use Case Example

  Lets look at a very simple example of the product purchase interaction
between a BUYER, a SELLER, and a COURIER.
  •       The buyer send a PURCHASE ORDER message to the seller.
  •       The seller check the credit history of the seller as well as the
product availability and decide either to accept or reject the purchase
  •       If the seller decide to reject the order, he send an ORDER
REJECTION message back to the buyer.  The interaction ends here.
  •       If the seller decide to accept the order, he will arrange shipment
of the purchased product by selecting one of his preferred couriers.
  •       The selected courier pickup the product from the seller and
deliver to the buyer’s location.  The courier start a new interaction with
the buyer by sending a SHIPMENT NOTICATION message.
  •       The buyer verifies the product is delivered in good shape and send
back a SHIPMENT RECEIVED message to the courier as well as FULFILLMENT
COMPLETE message to the seller.  Otherwise, the buyer sends back a SHIPMENT
REJECTED message to the courier as well as FULFILLMENT FAILED message to the

  3       Illustration

  3.1     Choreography
  There are four possible CHOREOGRAPHIES in this example
  •       Product Purchase (Buyer and Seller)
  •       Credit Checking (Seller and CreditCheck provider)
  •       Arrange Delivery (Seller and Courier)
  •       Shipment Delivery (Courier and Buyer)

  For the purpose of this discussion, I’ll focus in the first one.

  3.1.1   Public / Shared States
  Product Purchase choreography defines the following “PUBLIC STATES”
  •       OrderNo
  •       OrderStatus (“Submitted”, “Accepted”, “Rejected”, “Delivered”,
“Returned”, “Terminated”)
  •       ShipmentNo

  3.1.2   Message Exchanges
  Product Purchase Choreography defines the PUBLIC STATE TRANSITIONS in
terms of the following message exchanges …

  Start State = “submitted”:  (OrderStatus=“submitted”)
  Triggered by: when buyer send a “PurchaseOrder” to sender

  State = “accepted”:  (OrderStatus= “accepted”, OrderNo, ShipmentNo)
  From State “submitted”
  Triggered by when seller send a “OrderAcceptance” message to buyer

  State = “rejected”:  (OrderStatus= “rejected”)
  From State “submitted”
  Triggered by when seller send a “OrderRejection” message to buyer

  End State = “delivered”:  (OrderStatus= “delivered”, OrderNo, ShipmentNo)
  From State “accepted”
  Triggered by when buyer send a “FulfillmentCompleted” message to seller

  End State = “returned”:  (OrderStatus= “returned”, OrderNo, ShipmentNo)
  From State “accepted”
  Triggered by when buyer send a “FulfillmentFailed” message to seller

  3.2     Orchestration
  Here is the Orchestration of the Seller within the Product Purchase

  Wait for receiving a PurchaseOrder message from buyer
  Starts a new instance of “Credit Check” choreography by invoke the
CreditCheck web services.  After receiving the response, this “Credit Check”
choreography instance is terminated.
  Invoke an internal web service to check the stock level of product
  If (credit is OK and product is available) {
      Invoke an UDDI search to lookup shipping companies.
      Select one courier based on company specific decision logic
      Starts a new instance of “Arrange Shipment” choreography by invoking
the ShipmentHandling web services.
      Send an “OrderAcceptance” message (which include the shipment No) to
the buyer
      Wait for receiving either “FulfillmentComplete” or “FulfillmentFailed”
message from the buyer and update the OrderStatus correspondingly.  The
choreography instance ends here.
      If the OrderStatus is “return”, log into the customer care DB.
  } else {
      Send an “OrderRejection” message to the buyer

  As you can see, some activities within the orchestration is not visible by
the buyer and hence is the private part of the seller.
  - Check the credit history
  - Check the product availability
  - Start another choreography with the courier

  Comments and Thoughts ?

  Best regards,
Received on Saturday, 15 March 2003 20:53:34 UTC

This archive was generated by hypermail 2.3.1 : Tuesday, 6 January 2015 21:00:56 UTC