corda/docs/source/whitepaper/corda-technical-whitepaper.tex

\documentclass{article}
\author{Mike Hearn}
\date{December, 2016}
\title{Corda: A distributed ledger}
%%\setlength{\parskip}{\baselineskip}
\usepackage{amsfonts}
\usepackage{listings}
\usepackage{color}
\usepackage{epigraph}
\usepackage{graphicx}
\graphicspath{ {images/} }
\usepackage[export]{adjustbox}
\usepackage{float}
\usepackage{hyperref}
\usepackage[super,comma,sort&compress]{natbib}
\usepackage[nottoc]{tocbibind}
\usepackage[parfill]{parskip}
\usepackage{textcomp}
%\usepackage[natbibapa]{apacite}
\renewcommand{\thefootnote}{\alph{footnote}}

%\epigraphfontsize{\small\itshape}
\setlength\epigraphwidth{4.5cm}
\setlength\epigraphrule{0pt}

\begin{document}

\maketitle
%\epigraphfontsize{\small\itshape}

%\renewcommand{\abstractname}{An introduction}
\begin{center}
Version 0.1

\emph{Confidential: Pre-Publication Draft For R3 DLG}
\end{center}

\vspace{10mm}

\begin{abstract}

A decentralised database with minimal trust between nodes would allow for the creation of a global ledger. Such a ledger
would not only be capable of implementing cryptocurrencies but also have many useful applications in finance, trade,
supply chain tracking and more. We present Corda, a decentralised global database, and describe in detail how it
achieves the goal of providing a robust and easy to use platform for decentralised app development. We elaborate on the
high level description provided in the paper \emph{Corda: An introduction}\cite{CordaIntro} and provide a detailed
technical overview, but assume no prior knowledge of the platform.

\end{abstract}
\newpage
\tableofcontents
\newpage
\section{Introduction}

In many industries significant effort is needed to keep organisation-specific databases in sync with each
other. In the financial sector the effort of keeping different databases synchronised, reconciling them to ensure
they actually are synchronised and resolving the `breaks' that occur when they are not represents a significant
fraction of the total work a bank actually does!

Why not just use a shared relational database? This would certainly solve a lot of problems with only existing technology,
but it would also raise more questions than answers:

\begin{itemize}
\item Who would run this database? Where would we find a sufficient supply of angels to own it?
\item In which countries would it be hosted? What would stop that country abusing the mountain of sensitive information it would have?
\item What if it got hacked?
\item Can you actually scale a relational database to fit the entire financial system within it?
\item What happens if The Financial System\texttrademark~needs to go down for maintenance?
\item What kind of nightmarish IT bureaucracy would guard changes to the database schemas?
\item How would you manage access control?
\end{itemize}

We can imagine many other questions. A decentralised database attempts to answer them.

In this paper we differentiate between a \emph{decentralised} database and a \emph{distributed} database. A distributed
database like BigTable\cite{BigTable} scales to large datasets and transaction volumes by spreading the data over many
computers. However it is assumed that the computers in question are all run by a single homogenous organisation and that
the nodes comprising the database all trust each other not to misbehave or leak data. In a decentralised database, such
as the one underpinning Bitcoin\cite{Bitcoin}, the nodes make much weaker trust assumptions and actively cross-check
each other's work. Such databases trade off performance and usability in order to gain security and global acceptance.

\emph{Corda} is a decentralised database platform with the following novel features:

\begin{itemize}
\item New transaction types can be defined using JVM\cite{JVM} bytecode.
\item Transactions may execute in parallel, on different nodes, without either node being aware of the other's transactions.
\item Nodes are arranged in an authenticated peer to peer network. All communication is direct.
\item There is no block chain\cite{Bitcoin}. Transaction races are deconflicted using pluggable \emph{notaries}. A single
Corda network may contain multiple notaries that provide their guarantees using a variety of different algorithms. Thus
Corda is not tied to any particular consensus algorithm.
\item Data is shared on a need-to-know basis. Nodes provide the dependency graph of a transaction they are sending to
another node on demand, but there is no global broadcast of \emph{all} transactions.
\item Bytecode-to-bytecode transpilation is used to allow complex, multi-step transaction building protocols called
\emph{flows} to be modelled as blocking code. The code is transformed into an asynchronous state machine, with
checkpoints written to the node's backing database when messages are sent and received. A node may potentially have
millions of flows active at once and they may last days, across node restarts and even upgrades. Flows expose progress
information to node administrators and users and may interact with people as well as other nodes.
\item The data model allows for arbitrary object graphs to be stored in the ledger. These graphs are called \emph{states} and are the atomic unit of data.
\item The platform provides a rich type system for the representation of things like dates, currencies, legal entities and so on.
\item States can declare a relational mapping and can be queried using SQL.
\item Integration with existing systems is considered from the start. The network can support rapid bulk data imports
from other database systems without placing load on the network. Global ledger data can be joined with existing,
internal RDBMS tables thanks to slots in the state definitions that are reserved for join keys. Events on the ledger
are exposed via an embedded JMS compatible message broker.
\item States can declare scheduled events. For example a bond state may declare an automatic transition to a ``in default'' state if it is not repaid in time.
\end{itemize}

Comparisons with Bitcoin and Ethereum will be provided throughout.

\newpage

\section{Overview}

Corda is a platform for the writing of ``CorDapps'': applications that extend the global database with new capabilities.
Such apps define new data types, new inter-node protocols and the ``smart contracts'' that determine allowed changes.

What is a smart contract? That depends on the model of computation we are talking about. There are two competing
computational models used in decentralised databases: the virtual computer model and the UTXO model. The virtual
computer model is used by Ethereum\cite{Ethereum}. It models the database as the in-memory state of a
global computer with a single thread of execution determined by the block chain. In the UTXO model, as used in
Bitcoin, the database is a set of immutable rows keyed by \texttt{(hash:output index)}. Transactions define
outputs that append new rows and inputs which consume existing rows. The term ``smart contract'' has a different
meaning in each model. A deeper discussion of the tradeoffs and terminology in the different approaches can
be found in the Corda introductory paper\cite{CordaIntro}.

We use the UTXO model and as a result our transactions are structurally similar to Bitcoin transactions: they have
inputs, outputs and signatures. Unlike Bitcoin, Corda database rows can contain arbitrary data, not just a value field.
Because the data consumed and added by transactions is not necessarily a set of key/value pairs, we don't talk about rows
but rather \emph{states}. Like Bitcoin, Corda states are associated with bytecode programs that must accept a transaction
for it to be valid, but unlike Bitcoin, a transaction must satisfy the programs for both the input and output states
at once. \emph{Issuance transactions} may append new states to the database without consuming any existing states but
unlike in Bitcoin these transactions are not special and may be created at any time, by anyone.

In contrast to both Bitcoin and Ethereum, Corda does not order transactions using a block chain and by implication does
not use miners or proof-of-work. Instead each state points to a \emph{notary}, which is a service that guarantees it
will sign a transaction only if all the input states are un-consumed. A transaction is not allowed to consume states
controlled by multiple notaries and thus there is never any need for two-phase commit between notaries. If a combination of
states would cross notaries then a special transaction type is used to move them onto a single notary first.

Notaries are expected to be composed of multiple mutually distrusting parties who use a byzantine fault
tolerant algorithm like HoneyBadgerBFT\cite{HBBFT} to reach consensus. Notaries are identified by and sign with compound
public keys that conceptually follow the Interledger Crypto-Conditions specification\cite{ILPCC}. Note that whilst it
would be conventional to use a BFT algorithm for a notary service, there is no requirement to do so and in cases where
the legal system is sufficient to ensure protocol compliance a higher performance algorithm like RAFT may be used.
Because multiple notaries can co-exist a single network may provide a single global BFT notary for
general use and region-specific RAFT notaries for low latency trading within a unified regulatory area, for example
London or New York.

The Corda transaction format has various other features which are described in later sections.

\section{The peer to peer network}

\subsection{Network overview}
A Corda network consists of the following components:

\begin{itemize}
\item Nodes, communicating using AMQP/1.0 over TLS. Nodes use a relational database for data storage.
\item A permissioning service that automates the process of provisioning TLS certificates.
\item A network map service that publishes information about nodes on the network.
\item One or more notary services. A notary may itself be distributed over multiple nodes.
\item Zero or more oracle services. An oracle is a well known service that signs transactions if they state a fact
and that fact is considered to be true. This is how the ledger can be connected to the real world, despite being
fully deterministic.
\end{itemize}

A purely in-memory implementation of the messaging subsystem is provided which can inject simulated latency between
nodes and visualise communications between them. This can be useful for debugging, testing and educational purposes.

Oracles and notaries are covered in later sections.

\subsection{Identity and the permissioning service}

Unlike Bitcoin and Ethereum, Corda is designed for semi-private networks in which admission requires obtaining an
identity signed by a root authority. This assumption is pervasive - the flow API provides messaging in terms of identities,
with routing and delivery to underlying nodes being handled automatically. There is no global broadcast at any point.

This `identity' does not have to be a legal or true identity. In the same way that an email address is a globally
unique pseudonym that is ultimately rooted by the top of the DNS hierarchy, so too can a Corda network work with
arbitrary self-selected usernames. The permissioning service can implement any policy it likes as long as the
identities it signs are globally unique. Thus an entirely anonymous Corda network is possible if a suitable
IP obfuscation system like Tor is also used.

Whilst simple string identities are likely sufficient for some networks, the financial industry typically requires some
level of \emph{know your customer} checking, and differentiation between different legal entities that may share
the same brand name. Corda reuses the standard PKIX infrastructure for connecting public keys to identities and thus
names are actually X.500 names. When a single string is sufficient the \emph{common name} field can be used alone,
similar to the web PKI. In more complex deployments the additional structure X.500 provides may be useful to
differentiate between entities with the same name. For example there are at least five different companies called
\emph{American Savings Bank} and in the past there may have been more than 40 independent banks with that name.

More complex notions of identity that may attest to many time-varying attributes are not handled at this layer of the
system: the base identity is always just an X.500 name. Note that even though messaging is always identified, transactions
themselves may still contain anonymous public keys.

\subsection{The network map}

Every network require a network map service, which may itself be composed of multiple cooperating nodes. This is
similar to Tor's concept of \emph{directory authorities}. The network map publishes the IP addresses through which
every node on the network can be reached, along with the identity certificates of those nodes and the services they
provide. On receiving a connection nodes check that the connecting node is in the network map.

The network map abstracts the underlying IP addresses of the nodes from more useful business concepts like identities
and services. Each participant on the network, called a \emph{party}, publishes one or more IP addresses in the
network map. Equivalent domain names may be helpful for debugging but are not required. User interfaces and APIs
always work in terms of identities - there is thus no equivalent to Bitcoin's notion of an address (hashed public key),
and user-facing applications rely on auto-completion and search rather than QRcodes to identify a logical recipient.

It is possible to subscribe to network map changes and registering with the map is the first thing a node does at
startup. Nodes may optionally advertise their nearest city for load balancing and network visualisation purposes.

The map is a document that may be cached and distributed throughout the network. The map is therefore not required
to be highly available: if the map service becomes unreachable new nodes may not join the network and existing nodes
may not change their advertised service set, but otherwise things continue as normal.

\subsection{Message delivery}

The network is structurally similar to the email network. Nodes are expected to be long lived but may depart
temporarily due to crashes, connectivity interruptions or maintenance. Messages are written to disk
and delivery is retried until the remote node has acknowledged a message, at which point it is expected to have
either reliably stored the message or processed it completely. Connections between nodes are built and torn down as
needed: there is no assumption of constant connectivity. An ideal network would be entirely flat with high quality
connectivity between all nodes, but Corda recognises that this is not always compatible with common network
setups and thus the message routing component of a node can be separated from the rest and run outside the firewall.
In this way nodes that do not have duplex connectivity can still take part in the network as first class citizens.
Additionally a single node may have multiple advertised IP addresses.

The reference implementation provides this functionality using the Apache Artemis message broker, through which it
obtains journalling, load balancing, flow control, high availability clustering, streaming of messages too large to fit
in RAM and many other useful features. The network uses the \emph{AMQP/1.0}\cite{AMQP} protocol which is a widely
implemented binary messaging standard, combined with TLS to secure messages in transit and authenticate the endpoints.

\subsection{Serialization, sessioning, deduplication and signing}

All messages are encoded using a compact binary format. Each message has a UUID set in an AMQP header which is used
as a deduplication key, thus accidentally redelivered messages will be ignored.

% TODO: Describe the serialization format in more detail once finalised.

Messages may also have an associated organising 64-bit \emph{session ID}. Note that this is distinct from the AMQP
notion of a session. Sessions can be long lived and persist across node restarts and network outages. They exist in order
to group messages that are part of a \emph{flow}, described in more detail below.

Messages that are successfully processed by a node generate a signed acknowledgement message called a `receipt'. Note that
this is distinct from the unsigned acknowledgements that live at the AMQP level and which simply flag that a message was
successfully downloaded over the wire. A receipt may be generated some time after the message is processed in the case
where acknowledgements are being batched to amortise signing overhead, and the receipt identifies the message by the hash
of its content. The purpose of the receipts is to give a node undeniable evidence that a counterparty received a
notification that would stand up later in a dispute mediation process. Corda does not attempt to support deniable
messaging.

\section{Flow framework}

It is common in decentralised ledger systems for complex multi-party protocols to be needed. The Bitcoin payment channel
protocol\cite{PaymentChannels} involves two parties putting money into a multi-signature pot, then iterating with your
counterparty a shared transaction that spends that pot, with extra transactions used for the case where one party or the
other fails to terminate properly. Such protocols typically involve reliable private message passing, checkpointing to
disk, signing of transactions, interaction with the p2p network, reporting progress to the user, maintaining a complex
state machine with timeouts and error cases, and possibly interaction with internal systems on either side. All
this can become quite involved. The implementation of Bitcoin payment channels in the bitcoinj library is approximately
9000 lines of Java, very little of which involves cryptography.

As another example, the core Bitcoin protocol only
allows you to append transactions to the ledger. Transmitting other information that might be useful such as a text message,
refund address, identity information and so on is not supported and must be handled in some other way - typically by
wrapping the raw ledger transaction bytes in a larger message that adds the desired metadata and giving responsibility
for broadcasting the embedded transaction to the recipient, as in Bitcoin's BIP 70\cite{BIP70}.

In Corda transaction data is not globally broadcast. Instead it is transmitted to the relevant parties only when they
need to see it. Moreover even quite simple use cases - like sending cash - may involve a multi-step negotiation between
counterparties and the involvement of a third party such as a notary. Additional information that isn't put into the
ledger is considered essential, as opposed to nice-to-have. Thus unlike traditional blockchain systems in which the primary
form of communication is global broadcast, in Corda \emph{all} communication takes the form of small multi-party sub-protocols
called flows.

The flow framework presents a programming model that looks to the developer as if they have the ability to run millions
of long lived threads which can survive node restarts, and even node upgrades. APIs are provided to send and receive
object graphs to and from other identities on the network, embed sub-flows, and report progress to observers. In this
way business logic can be expressed at a very high level, with the details of making it reliable and efficient
abstracted away. This is achieved with the following components.

\paragraph{Just-in-time state machine compiler.}Code that is written in a blocking manner typically cannot be stopped
and transparently restarted later. The first time a flow's \texttt{call} method is invoked a bytecode-to-bytecode
transformation occurs that rewrites the classes into a form that implements a resumable state machine. These state
machines are sometimes called fibers or coroutines, and the transformation engine Corda uses is capable of rewriting
code arbitrarily deep in the stack on the fly. The developer may thus break his or her logic into multiple methods and
classes, use loops, and generally structure their program as if it were executing in a single blocking thread. There's only a
small list of things they should not do: sleeping, directly accessing the network APIs, or doing other tasks that might
block outside of the framework.

\paragraph{Transparent checkpointing.}When a flow wishes to wait for a message from another party (or input from a
human being) the underlying stack frames are suspended onto the heap, then crawled and serialized into the node's
underlying relational database using an object serialization framework. The written objects are prefixed with small
schema definitions that allow some measure of portability across changes to the layout of objects, although
portability across changes to the stack layout is left for future work. Flows are resumed and suspended on demand, meaning
it is feasible to have far more flows active at once than would fit in memory. The checkpointing process is atomic with
changes to local storage and acknowledgement of network messages.

\paragraph{Identity to IP address mapping.}Flows are written in terms of identities. The framework takes care of routing
messages to the right IP address for a given identity, following movements that may take place whilst the flow is active
and handling load balancing for multi-homed parties as appropriate.

\paragraph{A library of subflows.}Flows can invoke sub-flows, and a library of flows is provided to automate common tasks
like notarising a transaction or atomically swapping ownership of two assets.

\paragraph{Progress reporting.}Flows can provide a progress tracker that indicates which step they are up to. Steps can
have human-meaningful labels, along with other tagged data like a progress bar. Progress trackers are hierarchical and
steps can have sub-trackers for invoked sub-flows.

\paragraph{Flow hospital.}Flows can pause if they throw exceptions or explicitly request human assistance. A flow that
has stopped appears in the \emph{flow hospital} where the node's administrator may decide to kill the flow or provide it
with a solution. The ability to request manual solutions is useful for cases where the other side isn't sure why you
are contacting them, for example, the specified reason for sending a payment is not recognised, or when the asset used for
a payment is not considered acceptable.

\section{Data model}
\subsection{Commands}
\subsection{Identity lookups}
\subsection{Attachments, legal prose and bytecode}
\subsection{Merkle-structured transactions}
\subsection{Encumbrances}
\subsection{Contract constraints}
\section{Cash and Obligations}
\section{Integration with existing infrastructure}
\section{Deterministic JVM}
\section{Notaries}
\section{Secure signing devices}
\section{Client RPC and reactive collections}
\section{Event scheduling}

\section{Conclusion}

\bibliographystyle{unsrt}
\bibliography{Ref}

\end{document}