<liclass="toctree-l2"><aclass="reference internal"href="#rationale-for-and-tradeoffs-in-adopting-a-utxo-style-model">Rationale for and tradeoffs in adopting a UTXO-style model</a></li>
differs from existing systems in <aclass="reference external"href="https://r3-cev.atlassian.net/wiki/display/AWG/Platform+Stream%3A+Corda">the R3 wiki</a>,
<li>The basic notion of immutable states that are consumed and created by transactions is the same.</li>
<li>The notion of transactions having multiple inputs and outputs is the same. Bitcoin sometimes refers to the ledger
as the unspent transaction output set (UTXO set) as a result.</li>
<li>Like in Bitcoin, a contract is pure function. Contracts do not have storage or the ability to interact with anything.
Given the same transaction, a contract’s accept function always yields exactly the same result.</li>
<li>Bitcoin output scripts are parameterised by the input scripts in the spending transaction. This is somewhat similar
to our notion of a <em>command</em>.</li>
<li>Bitcoin transactions, like ours, refer to the states they consume by using a (txhash, index) pair. The Bitcoin
protocol calls these “outpoints”. In our prototype code they are known as <codeclass="docutils literal"><spanclass="pre">StateRefs</span></code> but the concept is identical.</li>
<li>Bitcoin transactions have an associated timestamp (the time at which they are mined).</li>
</ul>
<p>Differences:</p>
<ulclass="simple">
<li>A Bitcoin transaction has a single, rigid data format. A “state” in Bitcoin is always a (quantity of bitcoin, script)
pair and cannot hold any other data. Some people have been known to try and hack around this limitation by embedding
data in semi-standardised places in the contract code so the data can be extracted through pattern matching, but this
is a poor approach. Our states can include arbitrary typed data.</li>
<li>A Bitcoin transaction’s acceptance is controlled only by the contract code in the consumed input states. In practice
this has proved limiting. Our transactions invoke not only input contracts but also the contracts of the outputs.</li>
<li>A Bitcoin script can only be given a fixed set of byte arrays as the input. This means there’s no way for a contract
to examine the structure of the entire transaction, which severely limits what contracts can do.</li>
<li>Our contracts are Turing-complete and can be written in any ordinary programming language that targets the JVM.</li>
<h2>Rationale for and tradeoffs in adopting a UTXO-style model<aclass="headerlink"href="#rationale-for-and-tradeoffs-in-adopting-a-utxo-style-model"title="Permalink to this headline">¶</a></h2>
<p>As discussed above, Corda uses the so-called “UTXO set” model (unspent transaction output). In this model, the database
does not track accounts or balances. Instead all database entries are immutable. An entry is either spent or not spent
but it cannot be changed. In Bitcoin, spentness is implemented simply as deletion – the inputs of an accepted transaction
are deleted and the outputs created.</p>
<p>This approach has some advantages and some disadvantages, which is why some platforms like Ethereum have tried
(or are trying) to abstract this choice away and support a more traditional account-like model. We have explicitly
chosen <em>not</em> to do this and our decision to adopt a UTXO-style model is a deliberate one. In the section below,
the rationale for this decision and its pros and cons of this choice are outlined.</p>
</div>
<divclass="section"id="rationale">
<h2>Rationale<aclass="headerlink"href="#rationale"title="Permalink to this headline">¶</a></h2>
<p>Corda, in common with other blockchain-like platforms, is designed to bring parties to shared sets of data into
consensus as to the existence, content and allowable evolutions of those data sets. However, Corda is designed with the
explicit aim of avoiding, to the extent possible, the scalability and privacy implications that arise from those platforms’
decisions to adopt a global broadcast model.</p>
<p>Whilst the privacy implications of a global consensus model are easy to understand, the scalability implications are
perhaps more subtle, yet serious. In a consensus system, it is critical that all processors of a transaction reach
precisely the same conclusion as to its effects. In situations where two transactions may act on the same data set,
it means that the two transactions must be processed in the same <em>order</em> by all nodes. If this were not the case then it
would be possible to devise situations where nodes processed transactions in different orders and reached different
conclusions as to the state of the system. It is for this reason that systems like Ethereum effectively run
single-threaded, meaning the speed of the system is limited by the single-threaded performance of the slowest
machine on the network.</p>
<p>In Corda, we assume the data being processed represents financial agreements between identifiable parties and that these
institutions will adopt the system only if a significant number of such agreements can be managed by the platform.
As such, the system has to be able to support parallelisation of execution to the greatest extent possible,
whilst ensuring correct transaction ordering when two transactions seek to act on the same piece of shared state.</p>
<p>To achieve this, we must minimise the number of parties who need to receive and process copies of any given
transaction and we must minimise the extent to which two transactions seek to mutate (or supersede) any given piece
of shared state.</p>
<p>A key design decision, therefore, is what should be the most atomic unit of shared data in the system. This decision
also has profound privacy implications: the more coarsely defined the shared data units, the larger the set of
actors who will likely have a stake in its accuracy and who must process and observe any update to it.</p>
<p>This becomes most obvious when we consider two models for representing cash balances and payments.</p>
<p>A simple account model for cash would define a data structure that maintained a balance at a particular bank for each
“account holder”. Every holder of a balance would need a copy of this structure and would thus need to process and
validate every payment transaction, learning about everybody else’s payments and balances in the process.
All payments across that set of accounts would have to be single-threaded across the platform, limiting maximum
throughput.</p>
<p>A more sophisticated example might create a data structure per account holder.
But, even here, I would leak my account balance to anybody to whom I ever made
a payment and I could only ever make one payment at a time, for the same reasons above.</p>
<p>A UTXO model would define a data structure that represented an <em>instance</em> of a claim against the bank. An account
holder could hold <em>many</em> such instances, the aggregate of which would reveal their balance at that institution. However,
the account holder now only needs to reveal to their payee those instances consumed in making a payment to that payee.
This also means the payer could make several payments in parallel. A downside is that the model is harder to understand.
However, we consider the privacy and scalability advantages to overwhelm the modest additional cognitive load this places
on those attempting to learn the system.</p>
<p>In what follows, further advantages and disadvantages of this design decision are explored.</p>
</div>
<divclass="section"id="pros">
<h2>Pros<aclass="headerlink"href="#pros"title="Permalink to this headline">¶</a></h2>
<p>The UTXO model has these advantages:</p>
<ulclass="simple">
<li>Immutable ledger entries gives the usual advantages that a more functional approach brings: it’s easy to do analysis
on a static snapshot of the data and reason about the contents.</li>
<li>Because there are no accounts, it’s very easy to apply transactions in parallel even for high traffic legal entities
assuming sufficiently granular entries.</li>
<li>Transaction ordering becomes trivial: it is impossible to mis-order transactions due to the reliance on hash functions
to identify previous states. There is no need for sequence numbers or other things that are hard to provide in a
fully distributed system.</li>
<li>Conflict resolution boils down to the double spending problem, which places extremely minimal demands on consensus
algorithms (as the variable you’re trying to reach consensus on is a set of booleans).</li>
</ul>
</div>
<divclass="section"id="cons">
<h2>Cons<aclass="headerlink"href="#cons"title="Permalink to this headline">¶</a></h2>
<p>It also comes with some pretty serious complexities that in practice must be abstracted from developers:</p>
<ulclass="simple">
<li>Representing numeric amounts using immutable entries is unnatural. For instance, if you receive $1000 and wish
to send someone $100, you have to consume the $1000 output and then create two more: a $100 for the recipient and
$900 back to yourself as change. The fact that this happens can leak private information to an observer.</li>
<li>Because users do need to think in terms of balances and statements, you have to layer this on top of the
underlying ledger: you can’t just read someone’s balance out of the system. Hence, the “wallet” / position manager.
Experience from those who have developed wallets for Bitcoin and other systems is that they can be complex pieces of code,
although the bulk of wallets’ complexity in public systems is handling the lack of finality (and key management).</li>
<li>Whilst transactions can be applied in parallel, it is much harder to create them in parallel due to the need to
strictly enforce a total ordering.</li>
</ul>
<p>With respect to parallel creation, if the user is single threaded this is fine, but in a more complex situation
where you might want to be preparing multiple transactions in flight this can prove a limitation – in
the worst case where you have a single output that represents all your value, this forces you to serialise
the creation of every transaction. If transactions can be created and signed very fast that’s not a concern.
If there’s only a single user, that’s not a concern.</p>
<p>Both cases are typically true in the Bitcoin world, so users don’t suffer from this much. In the context of a
complex business with a large pool of shared funds, in which creation of transactions may be very slow due to the
need to get different humans to approve a tx using a signing device, this could quickly lead to frustrating
conflicts where someone approves a transaction and then discovers that it has become a double spend and
they must sign again. In the absolute worst case you could get a form of human livelock.</p>
<p>The tricky part about solving these problems is that the simplest way to express a payment request
(“send me $1000 to public key X”) inherently results in you receiving a single output, which then can
prove insufficiently granular to be convenient. In the Bitcoin space Mike Hearn and Gavin Andresen designed “BIP 70”
to solve this: it’s a simple binary format for requesting a payment and specifying exactly how you’d like to get paid,
including things like the shape of the transaction. It may seem that it’s an over complex approach: could you not
just immediately respend the big output back to yourself in order to split it? And yes, you could, until you hit
scenarios like “the machine requesting the payment doesn’t have the keys needed to spend it”,
which turn out to be very common. So it’s really more effective for a recipient to be able to say to the
sender, “here’s the kind of transaction I want you to send me”. The <aclass="reference internal"href="protocol-state-machines.html"><spanclass="doc">protocol framework</span></a>
may provide a vehicle to make such negotiations simpler.</p>
<p>A further challenge is privacy. Whilst our goal of not sending transactions to nodes that don’t “need to know”
helps, to verify a transaction you still need to verify all its dependencies and that can result in you receiving
lots of transactions that involve random third parties. The problems start when you have received lots of separate
payments and been careful not to make them linkable to your identity, but then you need to combine them all in a
single transaction to make a payment.</p>
<p>Mike Hearn wrote an article about this problem and techniques to minimise it in
<aclass="reference external"href="https://medium.com/@octskyward/merge-avoidance-7f95a386692f">this article</a> from 2013. This article
coined the term “merge avoidance”, which has never been implemented in the Bitcoin space,
although not due to lack of practicality.</p>
<p>A piece of future work for the wallet implementation will be to implement automated “grooming” of the wallet
to “reshape” outputs to useful/standardised sizes, for example, and to send outputs of complex transactions
back to their issuers for reissuance to “sever” long privacy-breaching chains.</p>
<p>Finally, it should be noted that some of the issues described here are not really “cons” of
the UTXO model; they’re just fundamental.
If you used many different anonymous accounts to preserve some privacy and then needed to
spend the contents of them all simultaneously, you’d hit the same problem, so it’s not
something that can be trivially fixed with data model changes.</p>
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>.
