<liclass="toctree-l1"><aclass="reference internal"href="oracles.html#implementing-an-oracle-with-continuously-varying-data">Implementing an oracle with continuously varying data</a></li>
<liclass="toctree-l1"><aclass="reference internal"href="oracles.html#using-an-oracle">Using an oracle</a></li>
<liclass="toctree-l1"><aclass="reference internal"href="setting-up-a-corda-network.html">Introduction - What is a corda network?</a></li>
<liclass="toctree-l1"><aclass="reference internal"href="setting-up-a-corda-network.html#setting-up-your-own-network">Setting up your own network</a></li>
platform’s state machine framework is used, and takes you through the code for a simple 2-party asset trading flow
which is included in the source.</p>
<divclass="section"id="introduction">
<h2>Introduction<aclass="headerlink"href="#introduction"title="Permalink to this headline">¶</a></h2>
<p>Shared distributed ledgers are interesting because they allow many different, mutually distrusting parties to
share a single source of truth about the ownership of assets. Digitally signed transactions are used to update that
shared ledger, and transactions may alter many states simultaneously and atomically.</p>
<p>Blockchain systems such as Bitcoin support the idea of building up a finished, signed transaction by passing around
partially signed invalid transactions outside of the main network, and by doing this you can implement
<em>delivery versus payment</em> such that there is no chance of settlement failure, because the movement of cash and the
traded asset are performed atomically by the same transaction. To perform such a trade involves a multi-step flow
in which messages are passed back and forth privately between parties, checked, signed and so on.</p>
<p>Despite how useful these flows are, platforms such as Bitcoin and Ethereum do not assist the developer with the rather
tricky task of actually building them. That is unfortunate. There are many awkward problems in their implementation
that a good platform would take care of for you, problems like:</p>
<ulclass="simple">
<li>Avoiding “callback hell” in which code that should ideally be sequential is turned into an unreadable mess due to the
desire to avoid using up a thread for every flow instantiation.</li>
<li>Surviving node shutdowns/restarts that may occur in the middle of the flow without complicating things. This
implies that the state of the flow must be persisted to disk.</li>
<li>Error handling.</li>
<li>Message routing.</li>
<li>Serialisation.</li>
<li>Catching type errors, in which the developer gets temporarily confused and expects to receive/send one type of message
when actually they need to receive/send another.</li>
<li>Unit testing of the finished flow.</li>
</ul>
<p>Actor frameworks can solve some of the above but they are often tightly bound to a particular messaging layer, and
we would like to keep a clean separation. Additionally, they are typically not type safe, and don’t make persistence or
writing sequential code much easier.</p>
<p>To put these problems in perspective, the <em>payment channel protocol</em> in the bitcoinj library, which allows bitcoins to
be temporarily moved off-chain and traded at high speed between two parties in private, consists of about 7000 lines of
Java and took over a month of full time work to develop. Most of that code is concerned with the details of persistence,
message passing, lifecycle management, error handling and callback management. Because the business logic is quite
spread out the code can be difficult to read and debug.</p>
<p>As small contract-specific trading flows are a common occurence in finance, we provide a framework for the
construction of them that automatically handles many of the concerns outlined above.</p>
</div>
<divclass="section"id="theory">
<h2>Theory<aclass="headerlink"href="#theory"title="Permalink to this headline">¶</a></h2>
<p>A <em>continuation</em> is a suspended stack frame stored in a regular object that can be passed around, serialised,
unserialised and resumed from where it was suspended. This concept is sometimes referred to as “fibers”. This may
sound abstract but don’t worry, the examples below will make it clearer. The JVM does not natively support
continuations, so we implement them using a library called Quasar which works through behind-the-scenes
bytecode rewriting. You don’t have to know how this works to benefit from it, however.</p>
<p>We use continuations for the following reasons:</p>
<ulclass="simple">
<li>It allows us to write code that is free of callbacks, that looks like ordinary sequential code.</li>
<li>A suspended continuation takes far less memory than a suspended thread. It can be as low as a few hundred bytes.
In contrast a suspended Java thread stack can easily be 1mb in size.</li>
<li>It frees the developer from thinking (much) about persistence and serialisation.</li>
</ul>
<p>A <em>state machine</em> is a piece of code that moves through various <em>states</em>. These are not the same as states in the data
model (that represent facts about the world on the ledger), but rather indicate different stages in the progression
of a multi-stage flow. Typically writing a state machine would require the use of a big switch statement and some
explicit variables to keep track of where you’re up to. The use of continuations avoids this hassle.</p>
</div>
<divclass="section"id="a-two-party-trading-flow">
<h2>A two party trading flow<aclass="headerlink"href="#a-two-party-trading-flow"title="Permalink to this headline">¶</a></h2>
<p>We would like to implement the “hello world” of shared transaction building flows: a seller wishes to sell some
<em>asset</em> (e.g. some commercial paper) in return for <em>cash</em>. The buyer wishes to purchase the asset using his cash. They
want the trade to be atomic so neither side is exposed to the risk of settlement failure. We assume that the buyer
and seller have found each other and arranged the details on some exchange, or over the counter. The details of how
the trade is arranged isn’t covered in this article.</p>
<p>Our flow has two parties (B and S for buyer and seller) and will proceed as follows:</p>
<olclass="arabic simple">
<li>S sends a <codeclass="docutils literal"><spanclass="pre">StateAndRef</span></code> pointing to the state they want to sell to B, along with info about the price they require
B to pay.</li>
<li>B sends to S a <codeclass="docutils literal"><spanclass="pre">SignedTransaction</span></code> that includes the state as input, B’s cash as input, the state with the new
owner key as output, and any change cash as output. It contains a single signature from B but isn’t valid because
it lacks a signature from S authorising movement of the asset.</li>
<li>S signs it and hands the now finalised <codeclass="docutils literal"><spanclass="pre">SignedTransaction</span></code> back to B.</li>
<p>You can find the implementation of this flow in the file <codeclass="docutils literal"><spanclass="pre">finance/src/main/kotlin/net/corda/flows/TwoPartyTradeFlow.kt</span></code>.</p>
<spanclass="k">override</span><spanclass="k">fun</span><spanclass="nf">toString</span><spanclass="p">()</span><spanclass="p">=</span><spanclass="s">"The submitted asset didn't match the expected type: $expectedTypeName vs $typeName"</span>
<spanclass="p">}</span>
<spanclass="c1">// This object is serialised to the network and is the first flow message the seller sends to the buyer.</span>
<p>This code defines several classes nested inside the main <codeclass="docutils literal"><spanclass="pre">TwoPartyTradeFlow</span></code> singleton. Some of the classes are
simply flow messages or exceptions. The other two represent the buyer and seller side of the flow.</p>
<p>Going through the data needed to become a seller, we have:</p>
<li><codeclass="docutils literal"><spanclass="pre">notaryNode:</span><spanclass="pre">NodeInfo</span></code> - the entry in the network map for the chosen notary. See “<aclass="reference internal"href="consensus.html"><spanclass="doc">Consensus model</span></a>” for more
information on notaries.</li>
<li><codeclass="docutils literal"><spanclass="pre">assetToSell:</span><spanclass="pre">StateAndRef<OwnableState></span></code> - a pointer to the ledger entry that represents the thing being sold.</li>
<li><codeclass="docutils literal"><spanclass="pre">price:</span><spanclass="pre">Amount<Currency></span></code> - the agreed on price that the asset is being sold for (without an issuer constraint).</li>
<li><codeclass="docutils literal"><spanclass="pre">myKeyPair:</span><spanclass="pre">KeyPair</span></code> - the key pair that controls the asset being sold. It will be used to sign the transaction.</li>
</ul>
<p>And for the buyer:</p>
<ulclass="simple">
<li><codeclass="docutils literal"><spanclass="pre">acceptablePrice:</span><spanclass="pre">Amount<Currency></span></code> - the price that was agreed upon out of band. If the seller specifies
a price less than or equal to this, then the trade will go ahead.</li>
<li><codeclass="docutils literal"><spanclass="pre">typeToBuy:</span><spanclass="pre">Class<out</span><spanclass="pre">OwnableState></span></code> - the type of state that is being purchased. This is used to check that the
sell side of the flow isn’t trying to sell us the wrong thing, whether by accident or on purpose.</li>
</ul>
<p>Alright, so using this flow shouldn’t be too hard: in the simplest case we can just create a Buyer or Seller
with the details of the trade, depending on who we are. We then have to start the flow in some way. Just
calling the <codeclass="docutils literal"><spanclass="pre">call</span></code> function ourselves won’t work: instead we need to ask the framework to start the flow for
us. More on that in a moment.</p>
</div>
<divclass="section"id="suspendable-functions">
<h2>Suspendable functions<aclass="headerlink"href="#suspendable-functions"title="Permalink to this headline">¶</a></h2>
<p>The <codeclass="docutils literal"><spanclass="pre">call</span></code> function of the buyer/seller classes is marked with the <codeclass="docutils literal"><spanclass="pre">@Suspendable</span></code> annotation. What does this mean?</p>
<p>As mentioned above, our flow framework will at points suspend the code and serialise it to disk. For this to work,
any methods on the call stack must have been pre-marked as <codeclass="docutils literal"><spanclass="pre">@Suspendable</span></code> so the bytecode rewriter knows to modify
the underlying code to support this new feature. A flow is suspended when calling either <codeclass="docutils literal"><spanclass="pre">receive</span></code>, <codeclass="docutils literal"><spanclass="pre">send</span></code> or
<codeclass="docutils literal"><spanclass="pre">sendAndReceive</span></code> which we will learn more about below. For now, just be aware that when one of these methods is
invoked, all methods on the stack must have been marked. If you forget, then in the unit test environment you will
get a useful error message telling you which methods you didn’t mark. The fix is simple enough: just add the annotation
and try again.</p>
<divclass="admonition note">
<pclass="first admonition-title">Note</p>
<pclass="last">Java 9 is likely to remove this pre-marking requirement completely.</p>
</div>
</div>
<divclass="section"id="starting-your-flow">
<h2>Starting your flow<aclass="headerlink"href="#starting-your-flow"title="Permalink to this headline">¶</a></h2>
<p>The <codeclass="docutils literal"><spanclass="pre">StateMachineManager</span></code> is the class responsible for taking care of all running flows in a node. It knows
how to register handlers with the messaging system (see “<aclass="reference internal"href="messaging.html"><spanclass="doc">Networking and messaging</span></a>”) and iterate the right state machine
when messages arrive. It provides the send/receive/sendAndReceive calls that let the code request network
interaction and it will save/restore serialised versions of the fiber at the right times.</p>
<p>Flows can be invoked in several ways. For instance, they can be triggered by scheduled events,
see “<aclass="reference internal"href="event-scheduling.html"><spanclass="doc">Event scheduling</span></a>” to learn more about this. Or they can be triggered directly via the Java-level node RPC
APIs from your app code.</p>
<p>You request a flow to be invoked by using the <codeclass="docutils literal"><spanclass="pre">CordaRPCOps.startFlowDynamic</span></code> method. This takes a
Java reflection <codeclass="docutils literal"><spanclass="pre">Class</span></code> object that describes the flow class to use (in this case, either <codeclass="docutils literal"><spanclass="pre">Buyer</span></code> or <codeclass="docutils literal"><spanclass="pre">Seller</span></code>).
It also takes a set of arguments to pass to the constructor. Because it’s possible for flow invocations to
be requested by untrusted code (e.g. a state that you have been sent), the types that can be passed into the
flow are checked against a whitelist, which can be extended by apps themselves at load time. There are also a series
of inlined extension functions of the form <codeclass="docutils literal"><spanclass="pre">CordaRPCOps.startFlow</span></code> which help with invoking flows in a type
safe manner.</p>
<p>The process of starting a flow returns a <codeclass="docutils literal"><spanclass="pre">FlowHandle</span></code> that you can use to either observe
the result, observe its progress and which also contains a permanent identifier for the invoked flow in the form
of the <codeclass="docutils literal"><spanclass="pre">StateMachineRunId</span></code>.</p>
<p>In a two party flow only one side is to be manually started using <codeclass="docutils literal"><spanclass="pre">CordaRPCOps.startFlow</span></code>. The other side
has to be registered by its node to respond to the initiating flow via <codeclass="docutils literal"><spanclass="pre">PluginServiceHub.registerFlowInitiator</span></code>.
<p>This is telling the buyer node to fire up an instance of <codeclass="docutils literal"><spanclass="pre">Buyer</span></code> (the code in the lambda) when the initiating flow
is a seller (<codeclass="docutils literal"><spanclass="pre">Seller::class</span></code>).</p>
</div>
<divclass="section"id="implementing-the-seller">
<h2>Implementing the seller<aclass="headerlink"href="#implementing-the-seller"title="Permalink to this headline">¶</a></h2>
<p>Let’s implement the <codeclass="docutils literal"><spanclass="pre">Seller.call</span></code> method. This will be run when the flow is invoked.</p>
<spanclass="n">logger</span><spanclass="p">.</span><spanclass="n">trace</span><spanclass="p">{</span><spanclass="s">"Received partially signed transaction: ${it.id}"</span><spanclass="p">}</span>
<spanclass="c1">// Download and check all the things that this transaction depends on and verify it is contract-valid,</span>
<spanclass="c1">// even though it is missing signatures.</span>
<spanclass="k">throw</span><spanclass="n">IllegalArgumentException</span><spanclass="p">(</span><spanclass="s">"Transaction is not sending us the right amount of cash"</span><spanclass="p">)</span>
<p>Let’s break this down. We fill out the initial flow message with the trade info, and then call <codeclass="docutils literal"><spanclass="pre">sendAndReceive</span></code>.
This function takes a few arguments:</p>
<ulclass="simple">
<li>The party on the other side.</li>
<li>The thing to send. It’ll be serialised and sent automatically.</li>
<li>Finally a type argument, which is the kind of object we’re expecting to receive from the other side. If we get
back something else an exception is thrown.</li>
</ul>
<p>Once <codeclass="docutils literal"><spanclass="pre">sendAndReceive</span></code> is called, the call method will be suspended into a continuation and saved to persistent
storage. If the node crashes or is restarted, the flow will effectively continue as if nothing had happened. Your
code may remain blocked inside such a call for seconds, minutes, hours or even days in the case of a flow that
needs human interaction!</p>
<divclass="admonition note">
<pclass="first admonition-title">Note</p>
<p>There are a couple of rules you need to bear in mind when writing a class that will be used as a continuation.
The first is that anything on the stack when the function is suspended will be stored into the heap and kept alive by
<p>The second is that as well as being kept on the heap, objects reachable from the stack will be serialised. The state
of the function call may be resurrected much later! Kryo doesn’t require objects be marked as serialisable, but even so,
doing things like creating threads from inside these calls would be a bad idea. They should only contain business
logic and only do I/O via the methods exposed by the flow framework.</p>
<pclass="last">It’s OK to keep references around to many large internal node services though: these will be serialised using a
special token that’s recognised by the platform, and wired up to the right instance when the continuation is
loaded off disk again.</p>
</div>
<p>The buyer is supposed to send us a transaction with all the right inputs/outputs/commands in response to the opening
message, with their cash put into the transaction and their signature on it authorising the movement of the cash.</p>
<p>You get back a simple wrapper class, <codeclass="docutils literal"><spanclass="pre">UntrustworthyData<SignedTransaction></span></code>, which is just a marker class that reminds
us that the data came from a potentially malicious external source and may have been tampered with or be unexpected in
other ways. It doesn’t add any functionality, but acts as a reminder to “scrub” the data before use.</p>
<p>Our “scrubbing” has three parts:</p>
<olclass="arabic simple">
<li>Check that the expected signatures are present and correct. At this point we expect our own signature to be missing,
because of course we didn’t sign it yet, and also the signature of the notary because that must always come last.</li>
<li>We resolve the transaction, which we will cover below.</li>
<li>We verify that the transaction is paying us the demanded price.</li>
</ol>
</div>
<divclass="section"id="sub-flows">
<h2>Sub-flows<aclass="headerlink"href="#sub-flows"title="Permalink to this headline">¶</a></h2>
<p>Flows can be composed via nesting. Invoking a sub-flow looks similar to an ordinary function call:</p>
<p>In this code snippet we are using the <codeclass="docutils literal"><spanclass="pre">NotaryFlow.Client</span></code> to request notarisation of the transaction.
We simply create the flow object via its constructor, and then pass it to the <codeclass="docutils literal"><spanclass="pre">subFlow</span></code> method which
returns the result of the flow’s execution directly. Behind the scenes all this is doing is wiring up progress
tracking (discussed more below) and then running the objects <codeclass="docutils literal"><spanclass="pre">call</span></code> method. Because this little helper method can
be on the stack when network IO takes place, we mark it as <codeclass="docutils literal"><spanclass="pre">@Suspendable</span></code>.</p>
<p>Going back to the previous code snippet, we use a sub-flow called <codeclass="docutils literal"><spanclass="pre">ResolveTransactionsFlow</span></code>. This is
responsible for downloading and checking all the dependencies of a transaction, which in Corda are always retrievable
from the party that sent you a transaction that uses them. This flow returns a list of <codeclass="docutils literal"><spanclass="pre">LedgerTransaction</span></code>
objects, but we don’t need them here so we just ignore the return value.</p>
<divclass="admonition note">
<pclass="first admonition-title">Note</p>
<pclass="last">Transaction dependency resolution assumes that the peer you got the transaction from has all of the
dependencies itself. It must do, otherwise it could not have convinced itself that the dependencies were themselves
valid. It’s important to realise that requesting only the transactions we require is a privacy leak, because if
we don’t download a transaction from the peer, they know we must have already seen it before. Fixing this privacy
leak will come later.</p>
</div>
<p>After the dependencies, we check the proposed trading transaction for validity by running the contracts for that as
well (but having handled the fact that some signatures are missing ourselves).</p>
<spanclass="n">logger</span><spanclass="p">.</span><spanclass="n">trace</span><spanclass="p">{</span><spanclass="s">"Built finished transaction, sending back to secondary!"</span><spanclass="p">}</span>
<p>It’s all pretty straightforward from now on. Here <codeclass="docutils literal"><spanclass="pre">id</span></code> is the secure hash representing the serialised
transaction, and we just use our private key to calculate a signature over it. As a reminder, in Corda signatures do
not cover other signatures: just the core of the transaction data.</p>
<p>In <codeclass="docutils literal"><spanclass="pre">sendSignatures</span></code>, we take the two signatures we obtained and add them to the partial transaction we were sent.
There is an overload for the + operator so signatures can be added to a SignedTransaction easily. Finally, we wrap the
two signatures in a simple wrapper message class and send it back. The send won’t block waiting for an acknowledgement,
but the underlying message queue software will retry delivery if the other side has gone away temporarily.</p>
<p>You can also see that every flow instance has a logger (using the SLF4J API) which you can use to log progress
messages.</p>
<divclass="admonition warning">
<pclass="first admonition-title">Warning</p>
<pclass="last">This sample code is <strong>not secure</strong>. Other than not checking for all possible invalid constructions, if the
seller stops before sending the finalised transaction to the buyer, the seller is left with a valid transaction
but the buyer isn’t, so they can’t spend the asset they just purchased! This sort of thing will be fixed in a
future version of the code.</p>
</div>
</div>
<divclass="section"id="implementing-the-buyer">
<h2>Implementing the buyer<aclass="headerlink"href="#implementing-the-buyer"title="Permalink to this headline">¶</a></h2>
<p>OK, let’s do the same for the buyer side:</p>
<li>We create a cash spend in the normal way, by using <codeclass="docutils literal"><spanclass="pre">VaultService.generateSpend</span></code>. See the vault documentation if this
<li>We access the <em>service hub</em> when we need it to access things that are transient and may change or be recreated
whilst a flow is suspended, things like the wallet or the network map.</li>
<li>Finally, we send the unfinished, invalid transaction to the seller so they can sign it. They are expected to send
back to us a <codeclass="docutils literal"><spanclass="pre">SignaturesFromSeller</span></code>, which once we verify it, should be the final outcome of the trade.</li>
</ol>
<p>As you can see, the flow logic is straightforward and does not contain any callbacks or network glue code, despite
the fact that it takes minimal resources and can survive node restarts.</p>
<divclass="admonition warning">
<pclass="first admonition-title">Warning</p>
<pclass="last">In the current version of the platform, exceptions thrown during flow execution are not propagated
back to the sender. A thorough error handling and exceptions framework will be in a future version of the platform.</p>
</div>
</div>
<divclass="section"id="progress-tracking">
<h2>Progress tracking<aclass="headerlink"href="#progress-tracking"title="Permalink to this headline">¶</a></h2>
<p>Not shown in the code snippets above is the usage of the <codeclass="docutils literal"><spanclass="pre">ProgressTracker</span></code> API. Progress tracking exports information
from a flow about where it’s got up to in such a way that observers can render it in a useful manner to humans who
may need to be informed. It may be rendered via an API, in a GUI, onto a terminal window, etc.</p>
<p>A <codeclass="docutils literal"><spanclass="pre">ProgressTracker</span></code> is constructed with a series of <codeclass="docutils literal"><spanclass="pre">Step</span></code> objects, where each step is an object representing a
stage in a piece of work. It is therefore typical to use singletons that subclass <codeclass="docutils literal"><spanclass="pre">Step</span></code>, which may be defined easily
in one line when using Kotlin. Typical steps might be “Waiting for response from peer”, “Waiting for signature to be
approved”, “Downloading and verifying data” etc.</p>
<p>Each step exposes a label. By default labels are fixed, but by subclassing <codeclass="docutils literal"><spanclass="pre">RelabelableStep</span></code>
you can make a step that can update its label on the fly. That’s useful for steps that want to expose non-structured
progress information like the current file being downloaded. By defining your own step types, you can export progress
in a way that’s both human readable and machine readable.</p>
<p>Progress trackers are hierarchical. Each step can be the parent for another tracker. By altering the
<codeclass="docutils literal"><spanclass="pre">ProgressTracker.childrenFor[step]</span><spanclass="pre">=</span><spanclass="pre">tracker</span></code> map, a tree of steps can be created. It’s allowed to alter the hierarchy
at runtime, on the fly, and the progress renderers will adapt to that properly. This can be helpful when you don’t
fully know ahead of time what steps will be required. If you _do_ know what is required, configuring as much of the
hierarchy ahead of time is a good idea, as that will help the users see what is coming up.</p>
<p>Every tracker has not only the steps given to it at construction time, but also the singleton
<codeclass="docutils literal"><spanclass="pre">ProgressTracker.UNSTARTED</span></code> step and the <codeclass="docutils literal"><spanclass="pre">ProgressTracker.DONE</span></code> step. Once a tracker has become <codeclass="docutils literal"><spanclass="pre">DONE</span></code> its
position may not be modified again (because e.g. the UI may have been removed/cleaned up), but until that point, the
position can be set to any arbitrary set both forwards and backwards. Steps may be skipped, repeated, etc. Note that
rolling the current step backwards will delete any progress trackers that are children of the steps being reversed, on
the assumption that those subtasks will have to be repeated.</p>
<p>Trackers provide an <aclass="reference external"href="http://reactivex.io/">Rx observable</a> which streams changes to the hierarchy. The top level
observable exposes all the events generated by its children as well. The changes are represented by objects indicating
whether the change is one of position (i.e. progress), structure (i.e. new subtasks being added/removed) or some other
aspect of rendering (i.e. a step has changed in some way and is requesting a re-render).</p>
<p>The flow framework is somewhat integrated with this API. Each <codeclass="docutils literal"><spanclass="pre">FlowLogic</span></code> may optionally provide a tracker by
overriding the <codeclass="docutils literal"><spanclass="pre">flowTracker</span></code> property (<codeclass="docutils literal"><spanclass="pre">getFlowTracker</span></code> method in Java). If the
<codeclass="docutils literal"><spanclass="pre">FlowLogic.subFlow</span></code> method is used, then the tracker of the sub-flow will be made a child of the current
step in the parent flow automatically, if the parent is using tracking in the first place. The framework will also
automatically set the current step to <codeclass="docutils literal"><spanclass="pre">DONE</span></code> for you, when the flow is finished.</p>
<p>Because a flow may sometimes wish to configure the children in its progress hierarchy _before_ the sub-flow
is constructed, for sub-flows that always follow the same outline regardless of their parameters it’s conventional
to define a companion object/static method (for Kotlin/Java respectively) that constructs a tracker, and then allow
the sub-flow to have the tracker it will use be passed in as a parameter. This allows all trackers to be built
and linked ahead of time.</p>
<p>In future, the progress tracking framework will become a vital part of how exceptions, errors, and other faults are
surfaced to human operators for investigation and resolution.</p>
</div>
<divclass="section"id="versioning">
<h2>Versioning<aclass="headerlink"href="#versioning"title="Permalink to this headline">¶</a></h2>
<p>Fibers involve persisting object-serialised stack frames to disk. Although we may do some R&D into in-place upgrades
in future, for now the upgrade process for flows is simple: you duplicate the code and rename it so it has a
new set of class names. Old versions of the flow can then drain out of the system whilst new versions are
initiated. When enough time has passed that no old versions are still waiting for anything to happen, the previous
copy of the code can be deleted.</p>
<p>Whilst kind of ugly, this is a very simple approach that should suffice for now.</p>
<divclass="admonition warning">
<pclass="first admonition-title">Warning</p>
<pclass="last">Flows are not meant to live for months or years, and by implication they are not meant to implement entire deal
lifecycles. For instance, implementing the entire life cycle of an interest rate swap as a single flow - whilst
technically possible - would not be a good idea. The platform provides a job scheduler tool that can invoke
flows for this reason (see “<aclass="reference internal"href="event-scheduling.html"><spanclass="doc">Event scheduling</span></a>”)</p>
</div>
</div>
<divclass="section"id="future-features">
<h2>Future features<aclass="headerlink"href="#future-features"title="Permalink to this headline">¶</a></h2>
<p>The flow framework is a key part of the platform and will be extended in major ways in future. Here are some of
the features we have planned:</p>
<ulclass="simple">
<li>Identity based addressing</li>
<li>Exception propagation and management, with a “flow hospital” tool to manually provide solutions to unavoidable
problems (e.g. the other side doesn’t know the trade)</li>
Built with <ahref="http://sphinx-doc.org/">Sphinx</a> using a <ahref="https://github.com/snide/sphinx_rtd_theme">theme</a> provided by <ahref="https://readthedocs.org">Read the Docs</a>.
