From 84a156877e6e401940be255dba78a1aa604e9ae5 Mon Sep 17 00:00:00 2001 From: Max Ogden Date: Sun, 30 Apr 2017 15:15:31 -0700 Subject: reorganize paper --- papers/dat-paper.md | 351 +++++++++++++++++++++++++-------------------------- papers/dat-paper.pdf | Bin 274419 -> 251426 bytes 2 files changed, 172 insertions(+), 179 deletions(-) (limited to 'papers') diff --git a/papers/dat-paper.md b/papers/dat-paper.md index 132218a..8910e57 100644 --- a/papers/dat-paper.md +++ b/papers/dat-paper.md @@ -1,92 +1,77 @@ # Abstract -Dat is a protocol designed for syncing distributed, dynamic datasets. A secure changelog is used to ensure dataset versions are distributed safely. Files are efficiently versioned by checking new file regions against existing ones to duplication of existing similar file regions. Any byte range of any version of any file can be efficiently accessed as a stream from a Dat repository over a network connection. Consumers can choose to fully or partially replicate the contents of a remote Dat repository, and can also subscribe to live changes. Dat uses built-in public key cryptography to encrypt and sign all network traffic, allowing it to make certain privacy and security guarantees. +Dat is a protocol designed for syncing folders of data, even if they are large or changing constantly. A cryptographically secure register of changes to the data over time is used to ensure dataset versions are distributed as exact copies of exact versions. Any byte range of any version of any file can be efficiently streamed from a Dat repository over a network connection. Consumers can choose to fully or partially replicate the contents of a remote Dat repository, and can also subscribe to live changes. Dat uses built-in public key cryptography to encrypt network traffic, allowing it to make certain privacy and security guarantees. A group of Dat clients can connect to each other to form a public or private decentralized network to exchange data amognst each other in a swarm. A reference implementation is provided in JavaScript. # 1. Background -Sharing datasets over the Internet is a subject of much study, but approaches remain relatively limiting. The most widely used approach, sharing files over HTTP, is subject to dead links when files are moved or deleted, as HTTP has no concept of history or versioning built in. E-mailing datasets as attachments is also widely used, and has the concept of history built in, but many email providers limit the maximum attachment size which makes it impractical for many datasets. +Many datasets are shared online today using HTTP and FTP, which lack built in support for version control or content addressing of data. This results in link rot and content drift as files are moved, updated or deleted, leading to an alarming rate of disappearing data references in areas such as [published scientific literature](http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0115253). -Cloud storage services like S3 ensure availability of data, but they have a centralized hub-and-spoke networking model and tend to be limited by their bandwidth, meaning popular files can be come very expensive to share. Services like Dropbox and Google Drive provide version control and synchronization on top of cloud storage services which fixes many issues with broken links but rely on proprietary code and services requiring users to store their data on cloud infrastructure which has implications on cost, transfer speeds, and user privacy. +Cloud storage services like S3 ensure availability of data, but they have a centralized hub-and-spoke networking model and are therefore limited by their bandwidth, meaning popular files can be come very expensive to share. Services like Dropbox and Google Drive provide version control and synchronization on top of cloud storage services which fixes many issues with broken links but rely on proprietary code and services requiring users to store their data on centralized cloud infrastructure which has implications on cost, transfer speeds, vendor lock-in and user privacy. -Distributed file sharing tools can become faster as files become more popular, removing the bandwidth bottleneck and making file distribution cheaper. They also implement discovery systems which can prevent broken links meaning if the original source goes offline other backup sources can be automatically discovered. However these file sharing tools today are not supported by Web browsers, do not have good privacy guarantees, and do not provide a mechanism for updating files without redistributing a new dataset which could mean entire redownloading data you already have. +Distributed file sharing tools can become faster as files become more popular, removing the bandwidth bottleneck and making file distribution cheaper. They also use link resolution and discovery systems which can prevent broken links meaning if the original source goes offline other backup sources can be automatically discovered. However these file sharing tools today are not supported by Web browsers, do not have good privacy guarantees, and do not provide a mechanism for updating files without redistributing a new dataset which could mean entire redownloading data you already have. -Scientists are an example of a group that would benefit from better solutions to these problems. Increasingly scientific datasets are being provided online using one of the above approaches and cited in published literature. Broken links and systems that do not provide version checking or content addressability of data directly limit the reproducibility of scientific analyses based on shared datasets. Services that charge a premium for bandwidth cause monetary and data transfer strain on the users sharing the data, who are often on fast public university networks with effectively unlimited bandwidth that go unused. Version control tools designed for text files do not keep up with the demands of data analysis in science today. +# 2. Dat -# 2. Existing Work +Dat is a dataset synchronization protocol that does not assume a dataset is static or that the entire dataset will be downloaded. The main reference implementation is available from from npm as `npm install dat -g`. -Dat is inspired by a number of features from existing systems. - -## 2.1 Git - -Git popularized the idea of a directed acyclic graph (DAG) combined with a Merkle tree, a way to represent changes to data where each change is addressed by the secure hash of the change plus all ancestor hashes in a graph. This provides a way to trust data integrity, as the only way a specific hash could be derived by another peer is if they have the same data and change history required to reproduce that hash. This is important for reproducibility as it lets you trust that a specific git commit hash refers to a specific source code state. - -Decentralized version control tools for source code like Git provide a protocol for efficiently downloading changes to a set of files, but are optimized for text files and have issues with large files. Solutions like Git-LFS solve this by using HTTP to download large files, rather than the Git protocol. GitHub offers Git-LFS hosting but charges repository owners for bandwidth on popular files. Building a distributed distribution layer for files in a Git repository is difficult due to design of Git Packfiles which are delta compressed repository states that do not easily support random access to byte ranges in previous file versions. - -## 2.2 LBFS +The protocol is agnostic to the underlying transport e.g. you could implement Dat over carrier pigeon. The key properties of the Dat design are explained in this section. -LBFS is a networked file system that avoids transferring redundant data by deduplicating common regions of files and only transferring unique regions once. The deduplication algorithm they use is called Rabin fingerprinting and works by hashing the contents of the file using a sliding window and looking for content defined chunk boundaries that probabilistically appear at the desired byte offsets (e.g. every 1kb). +- 2.1 **Content Integrity** - Data and publisher integrity is verified through use of signed hashes of the content. +- 2.2 **Decentralized Mirroring** - Users sharing the same Dat automatically discover each other and exhange data in a swarm. +- 2.3 **Network Privacy** - Dat provides certain privacy guarantees including end-to-end encryption. +- 2.4 **Incremental Versioning** - Datasets can be efficiently synced, even in real time, to other peers. -Content defined chunking has the benefit of being shift resistant, meaning if you insert a byte into the middle of a file only the first chunk boundary to the right of the insert will change, but all other boundaries will remain the same. With a fixed size chunking strategy, such as the one used by rsync, all chunk boundaries to the right of the insert will be shifted by one byte, meaning half of the chunks of the file would need to be retransmitted. +## 2.1 Content Integrity -## 2.3 BitTorrent +Content integrity means being able to verify the data you received is the exact same version of the data that you expected. This is imporant in a distributed system as this mechanism will catch incorrect data sent by bad peers. It also has implications for reproducibility as it lets you refer to a specific version of a dataset. -BitTorrent implements a swarm based file sharing protocol for static datasets. Data is split into fixed sized chunks, hashed, and then that hash is used to discover peers that have the same data. An advantage of using BitTorrent for dataset transfers is that download bandwidth can be fully saturated. Since the file is split into pieces, and peers can efficiently discover which pieces each of the peers they are connected to have, it means one peer can download non-overlapping regions of the dataset from many peers at the same time in parallel, maximizing network throughput. +Link rot, when links online stop resolving, and content drift, when data changes but the link to the data remains the same, are two common issues in data analysis. For example, one day a file called data.zip might change, but a typical HTTP link to the file does not include a hash of the content, or provide a way to get updated metadata, so clients that only have the HTTP link have no way to check if the file changed without downloading the entire file again. Referring to a file by the hash of its content is called content addressability, and lets users not only verify that the data they receive is the version of the data they want, but also lets people cite specific versions of the data by referring to a specific hash. -Fixed sized chunking has drawbacks for data that changes (see LBFS above). BitTorrent assumes all metadata will be transferred up front which makes it impractical for streaming or updating content. Most BitTorrent clients divide data into 1024 pieces meaning large datasets could have a very large chunk size which impacts random access performance (e.g. for streaming video). +Dat uses BLAKE2b cryptographically secure hashes to address content. Hashes are arranged in a Merkle tree, a tree where each non-leaf node is the hash of all child nodes. Leaf nodes contain pieces of the dataset. Due to the properties of secure cryptographic hashes the top hash can only be produced if all data below it matches exactly. If two trees have matching top hashes then you know that all other nodes in the tree must match as well, and you can conclude that your dataset is synchronized. -Another drawback of BitTorrent is due to the way clients advertise and discover other peers in absence of any protocol level privacy or trust. From a user privacy standpoint, BitTorrent leaks what users are accessing or attempting to access, and does not provide the same browsing privacy functions as systems like SSL. +### Dat Links -## 2.4 Kademlia Distributed Hash Table +Dat links are Ed25519 public keys which have a length of 32 bytes (64 characters when Base64 encoded). You can represent your Dat link in the following ways and Dat clients will be able to understand them: -Kademlia is a distributed hash table, a distributed key/value store that can serve a similar purpose to DNS servers but has no hard coded server addresses. All clients in Kademlia are also servers. As long as you know at least one address of another peer in the network, you can ask them for the key you are trying to find and they will either have it or give you some other people to talk to that are more likely to have it. +- The standalone public key: -If you don't have an initial peer to talk to you, most clients use a bootstrap server that randomly gives you a peer in the network to start with. If the bootstrap server goes down, the network still functions as long as other methods can be used to bootstrap new peers (such as sending them peer addresses through side channels like how .torrent files include tracker addresses to try in case Kademlia finds no peers). +`8e1c7189b1b2dbb5c4ec2693787884771201da9...` -Kademlia is distinct from previous DHT designs due to its simplicity. It uses a very simple XOR operation between two keys as its "distance" metric to decide which peers are closer to the data being searched for. On paper it seems like it wouldn't work as it doesn't take into account things like ping speed or bandwidth. Instead its design is very simple on purpose to minimize the amount of control/gossip messages and to minimize the amount of complexity required to implement it. In practice Kademlia has been extremely successful and is widely deployed as the "Mainline DHT" for BitTorrent, with support in all popular BitTorrent clients today. +- Using the dat:// protocol: -Due to the simplicity in the original Kademlia design a number of attacks such as DDOS and/or sybil have been demonstrated. There are protocol extensions (BEPs) which in certain cases mitigate the effects of these attacks, such as BEP 44 which includes a DDOS mitigation technique. Nonetheless anyone using Kademlia should be aware of the limitations. +`dat://8e1c7189b1b2dbb5c4ec2693787884771...` -## 2.5 Peer to Peer Streaming Peer Protocol (PPSPP) +- As part of an HTTP URL: -PPSPP ([IETF RFC 7574](https://datatracker.ietf.org/doc/rfc7574/?include_text=1)) is a protocol for live streaming content over a peer to peer network. In it they define a specific type of Merkle Tree that allows for subsets of the hashes to be requested by a peer in order to reduce the time-till-playback for end users. BitTorrent for example transfers all hashes up front, which is not suitable for live streaming. +`https://datproject.org/8e1c7189b1b2dbb5...` -Their Merkle trees are ordered using a scheme they call "bin numbering", which is a method for deterministically arranging an append-only log of leaf nodes into an in-order layout tree where non-leaf nodes are derived hashes. If you want to verify a specific node, you only need to request its sibling's hash and all its uncle hashes. PPSPP is very concerned with reducing round trip time and time-till-playback by allowing for many kinds of optimizations, such as to pack as many hashes into datagrams as possible when exchanging tree information with peers. +All messages in the Dat protocol are encrypted and signed using the public key during transport. This means that unless you know the public key (e.g. unless the Dat link was shared with you) then you will not be able to discover or communicate with any member of the swarm for that Dat. Anyone with the public key can verify that messages (such as entries in a Dat Stream) were created by a holder of the private key. -Although PPSPP was designed with streaming video in mind, the ability to request a subset of metadata from a large and/or streaming dataset is very desirable for many other types of datasets. +Every Dat repository has corresponding a private key that kept in your home folder and never shared. Dat never exposes either the public or private key over the network. During the discovery phase the BLAKE2b hash of the public key is used as the discovery key. This means that the original key is impossible to discover (unless it was shared publicly through a separate channel) since only the hash of the key is exposed publicly. -## 2.6 WebTorrent +Dat does not provide an authentication mechanism at this time. Instead it provides a capability system. Anyone with the Dat link is currently considered able to discover and access data. Do not share your Dat links publicly if you do not want them to be accessed. -With WebRTC browsers can now make peer to peer connections directly to other browsers. BitTorrent uses UDP sockets which aren't available to browser JavaScript, so can't be used as-is on the Web. +### Hypercore and Hyperdrive -WebTorrent implements the BitTorrent protocol in JavaScript using WebRTC as the transport. This includes the BitTorrent block exchange protocol as well as the tracker protocol implemented in a way that can enable hybrid nodes, talking simultaneously to both BitTorrent and WebTorrent swarms (if a client is capable of making both UDP sockets as well as WebRTC sockets, such as Node.js). Trackers are exposed to web clients over HTTP or WebSockets. - -## 2.7 InterPlanetary File System - -IPFS is a family of application and network protocols that have peer to peer file sharing and data permanence baked in. IPFS abstracts network protocols and naming systems to provide an alternative application delivery platform to todays Web. For example, instead of using HTTP and DNS directly, in IPFS you would use LibP2P streams and IPNS in order to gain access to the features of the IPFS platform. - -## 2.8 Certificate Transparency/Secure Registers +The Dat storage, content integrity, and networking protocols are implemented in a module called [Hypercore](https://npmjs.org/hypercore). Hypercore is agnostic to the format of the input data, it operates on any stream of binary data. For the Dat use case of synchronizing datasets we use a file system module on top of Hypercore called [Hyperdrive](https://npmjs.org/hyperdrive). -The UK Government Digital Service have developed the concept of a register which they define as a digital public ledger you can trust. In the UK government registers are beginning to be piloted as a way to expose essential open data sets in a way where consumers can verify the data has not been tampered with, and allows the data publishers to update their data sets over time. +Dat has a layered abstraction so that users can use Hypercore directly to have full control over how they model their data. Hyperdrive works well when your data can be represented as files on a filesystem, which is the main use case with Dat. -The design of registers was inspired by the infrastructure backing the Certificate Transparency project, initated at Google, which provides a service on top of SSL certificates that enables service providers to write certificates to a distributed public ledger. Anyone client or service provider can verify if a certificate they received is in the ledger, which protects against so called "rogue certificates". +### Hypercore Registers -# 3. Dat +Hypercore Registers are the core mechanism used in Dat. They are binary append-only streams whose contents are cryptographically hashed and signed and therefore can be verified by anyone with access to the public key of the writer. They are an implemenation of the concept known as a register, a digital ledger you can trust -Dat is a dataset synchronization protocol that does not assume a dataset is static or that the entire dataset will be downloaded. The protocol is agnostic to the underlying transport e.g. you could implement Dat over carrier pigeon. The key properties of the Dat design are explained in this section. +Dat uses two registers, `content` and `metadata`. The `content` register contains the files in your repository and `metadata` contains the metadata about the files including name, size, last modified time, etc. Dat replicates them both when synchronizing with another peer. -- 1. **Mirroring** - Any participant in the network can simultaneously share and consume data. -- 2. **Content Integrity** - Data and publisher integrity is verified through use of signed hashes of the content. -- 3. **Parallel Replication** - Subsets of the data can be accessed from multiple peers simultaneously, improving transfer speeds. -- 4. **Efficient Versioning** - Datasets can be efficiently synced, even in real time, to other peers using Dat Streams. -- 5. **Network Privacy** - Dat employs a capability system whereby anyone with a Dat link can connect to the swarm, but the link itself is very difficult to guess. +When files are added to Dat, each file gets split up into some number of chunks, and the chunks are then arranged into a Merkle tree, which is used later for version control and replication processed. -## 3.1 Mirroring +## 2.2 Decentralized Mirroring -Dat is a peer to peer protocol designed to exchange pieces of a dataset amongst a swarm of peers. As soon as a peer acquires their first piece of data in the dataset they can choose to become a partial mirror for the dataset. If someone else contacts them and needs the piece they have, they can choose to share it. This can happen simultaneously while the peer is still downloading the pieces they want. +Dat is a peer to peer protocol designed to exchange pieces of a dataset amongst a swarm of peers. As soon as a peer acquires their first piece of data in the dataset they can choose to become a partial mirror for the dataset. If someone else contacts them and needs the piece they have, they can choose to share it. This can happen simultaneously while the peer is still downloading the pieces they want from others. -### 3.1.1 Source Discovery +### Source Discovery -An important aspect of mirroring is source discovery, the techniques that peers use to find each other. Source discovery means finding the IP and port of data sources online that have a copy of that data you are looking for. You can then connect to them and begin exchanging data using a Dat Stream. By using source discovery techniques Dat is able to create a network where data can be discovered even if the original data source disappears. +An important aspect of mirroring is source discovery, the techniques that peers use to find each other. Source discovery means finding the IP and port of data sources online that have a copy of that data you are looking for. You can then connect to them and begin exchanging data. By using source discovery techniques Dat is able to create a network where data can be discovered even if the original data source disappears. Source discovery can happen over many kinds of networks, as long as you can model the following actions: @@ -94,143 +79,61 @@ Source discovery can happen over many kinds of networks, as long as you can mode - `leave(key, [port])` - Stop looking for `key`. Specify `port` to stop announcing that you share `key` as well. - `foundpeer(key, ip, port)` - Called when a peer is found by a lookup -In the Dat implementation we implement the above actions on top of four types of discovery networks: +In the Dat implementation we implement the above actions on top of three types of discovery networks: - DNS name servers - An Internet standard mechanism for resolving keys to addresses - Multicast DNS - Useful for discovering peers on local networks -- Kademlia Mainline Distributed Hash Table - Zero point of failure, increases probability of Dat working even if DNS servers are unreachable -- [Signalhub](https://npmjs.org/signalhub) - An HTTP key resolving service, non-distributed. Used by web browser clients who can't form raw UDP/TCP packets. +- Kademlia Mainline Distributed Hash Table - Less central points of failure, increases probability of Dat working even if DNS servers are unreachable -Additional discovery networks can be implemented as needed. We chose the above four as a starting point to have a complementary mix of strategies to increase the probability of source discovery. +Additional discovery networks can be implemented as needed. We chose the above three as a starting point to have a complementary mix of strategies to increase the probability of source discovery. Additionally you can specify a Dat via HTTPS link, which runs the Dat protocol in "single-source" mode, meaning the above discovery networks are not used, and instead only that one HTTPS server is used as the only peer. Our implementation of source discovery is called [discovery-channel](https://npmjs.org/discovery-channel). We also run a [custom DNS server](https://www.npmjs.com/package/dns-discovery) that Dat clients use (in addition to specifying their own if they need to), as well as a [DHT bootstrap](https://github.com/bittorrent/bootstrap-dht) server. These discovery servers are the only centralized infrastructure we need for Dat to work over the Internet, but they are redundant, interchangeable, never see the actual data being shared, anyone can run their own and Dat will still work even if they all are unavailable. If this happens discovery will just be manual (e.g. manually sharing IP/ports). -TODO detail each discovery mechanism - -### 3.1.2 Peer Connections +### Peer Connections -After the discovery phase, Dat should have a list of potential data sources to try and contact. Dat uses either [TCP](https://en.wikipedia.org/wiki/Transmission_Control_Protocol), [UTP](https://en.wikipedia.org/wiki/Micro_Transport_Protocol), WebSockets or WebRTC for the network connections. UTP is designed to not take up all available bandwidth on a network (e.g. so that other people sharing wifi can still use the Internet). WebSockets and WebRTC makes Dat work in modern web browsers. Note that these are the protocols we support in the reference Dat implementation, but the Dat protocol itself is transport agnostic. +After the discovery phase, Dat should have a list of potential data sources to try and contact. Dat uses either TCP, [UTP](https://en.wikipedia.org/wiki/Micro_Transport_Protocol), or HTTP. UTP is designed to not take up all available bandwidth on a network (e.g. so that other people sharing wifi can still use the Internet). HTTP is support for compatibility with static file servers and web browser clients. Note that these are the protocols we support in the reference Dat implementation, but the Dat protocol itself is transport agnostic. -When Dat gets the IP and port for a potential source it tries to connect using all available protocols and hopes one works. If one connects first, Dat aborts the other ones. If none connect, Dat will try again until it decides that source is offline or unavailable and then stops trying to connect to them. Sources Dat is able to connect to go into a list of known good sources, so that the Internet connection goes down Dat can use that list to reconnect to known good sources again quickly. +If an HTTP source is specified Dat will prefer that one over other sources. Otherwise when Dat gets the IP and port for a potential TCP or UTP source it tries to connect using both protocols. If one connects first, Dat aborts the other one. If none connect, Dat will try again until it decides that source is offline or unavailable and then stops trying to connect to them. Sources Dat is able to connect to go into a list of known good sources, so that the Internet connection goes down Dat can use that list to reconnect to known good sources again quickly. If Dat gets a lot of potential sources it picks a handful at random to try and connect to and keeps the rest around as additional sources to use later in case it decides it needs more sources. The connection logic is implemented in a module called [discovery-swarm](https://www.npmjs.com/package/discovery-swarm). This builds on discovery-channel and adds connection establishment, management and statistics. It provides statistics such as how many sources are currently connected, how many good and bad behaving sources have been talked to, and it automatically handles connecting and reconnecting to sources. UTP support is implemented in the module [utp-native](https://www.npmjs.com/package/utp-native). -Once a duplex binary connection to a remote source is open Dat then layers on its own protocol on top called a Dat Stream. - -## 3.2 Content Integrity - -Content integrity means being able to verify the data you received is the exact same version of the data that you expected. This is imporant in a distributed system as this mechanism will catch incorrect data sent by bad peers. It also has implications for reproducibility as it lets you refer to a specific version of a dataset. - -Link rot, when links online stop resolving, and content drift, when data changes but the link to the data remains the same, are two common issues in data analysis. For example, one day a file called data.zip might change, but a typical HTTP link to the file does not include a hash of the content, or provide a way to get updated metadata, so clients that only have the HTTP link have no way to check if the file changed without downloading the entire file again. Referring to a file by the hash of its content is called content addressability, and lets users not only verify that the data they receive is the version of the data they want, but also lets people cite specific versions of the data by referring to a specific hash. - -Dat uses SHA256 hashes to address content. Hashes are arranged in a Merkle tree, a tree where each non-leaf node is the hash of all child nodes. Leaf nodes contain pieces of the dataset. This means that in order to verify the integrity of some subset of content only the top most common ancestors of the leaf nodes that contain that content must be fetched. For example to verify all content in a Merkle tree the top level node of the tree can be used. Due to the behavior of secure cryptographic hashes the top hash can only be produced if all data below it matches exactly. If two trees have matching top hashes then you know that all other nodes in the tree must match as well, and you can conclude that your dataset is synchronized. - -### 3.2.1 Hypercore and Hyperdrive - -The Dat storage, content integrity, and networking protocols are implemented in a module called [Hypercore](https://npmjs.org/hypercore). Hypercore is agnostic to the format of the input data, it operates on any stream of binary data. For the Dat use case of synchronizing datasets we use a file system module on top of Hypercore called [Hyperdrive](https://npmjs.org/hyperdrive). - -We have a layered abstraction so that if someone wishes they can use Hypercore directly to have full control over how they model their data. Hyperdrive works well when your data can be represented as files on a filesystem, which is our main use case with Dat. - -### 3.2.2 Dat Streams +Once a duplex binary connection to a remote source is open Dat then layers on the Hypercore protocol, a message based replication protocol that allows two peers to communicate over a stateless channel to request and exchange data. You open separate replication channels with many peers at once which allows clients to parallelize data requests across the entire pool of peers they have established connections with. -Dat Streams are binary append-only stream whose contents are cryptographically hashed and signed and therefore can be verified by anyone with access to the public key of the writer. They are an implemenation of the concept known as a register, a digital ledger you can trust. Dat lets you create many Streams, and replicates them when synchronizing with another peer. +## 2.3 Network Privacy -Dat Streams use a specific method of encoding a Merkle tree where hashes are positioned by a scheme called binary interval numbering or just simply "bin" numbering. This is just a specific, deterministic way of laying out the nodes in a tree. For example a tree with 7 nodes will always be arranged like this: - -``` -0 - 1 -2 - 3 -4 - 5 -6 -``` - -In our use case, the hashes of the actual content are always even numbers, at the wide end of the tree. So the above tree had four original values that become the even numbers: - -``` -value0 -> 0 -value1 -> 2 -value2 -> 4 -value3 -> 6 -``` - -A Dat Stream contains two pieces of information: - -Evens: List of binary values with their hash and size: [value0, value1, value2, ...] -Odds: List of Merkle hashes with the size of all their children: [hash0, hash1, hash2, ...] - -These two lists get interleaved into a single register such that the indexes (position) in the register are the same as the bin numbers from the Merkle tree. - -All odd hashes are derived by hashing the two child nodes, e.g. given hash0 is `hash(value0)` and hash2 is `hash(value1)`, hash1 is `hash(hash0 + hash2)`. - -For example a Dat Stream with two data entries would look something like this (pseudocode): - -``` -0. hash(value0) -1. hash(hash(value0) + hash(value1)) -2. hash(value1) -``` +On the Web today, with SSL, there is a guarantee that the traffic between your computer and the server is private. As long as you trust the server to not leak your logs, attackers who intercept your network traffic will not be able to read the HTTP traffic exchanged between you and the server. This is a fairly straightforward model as clients only have to trust a single server for some domain. -## 3.3 Parallel Replication +There is an inherent tradeoff in peer to peer systems of source discovery vs. user privacy. The more sources you contact and ask for some data, the more sources you trust to keep what you asked for private. Our goal is to have Dat be configurable in respect to this tradeoff to allow application developers to meet their own privacy guidelines. -Dat Streams include a message based replication protocol so two peers can communicate over a stateless channel to discover and exchange data. Once you have received the Stream metadata, you can make individual requests for chunks from any peer you are connected to. This allows clients to parallelize data requests across the entire pool of peers they have established connections with. +It is up to client programs to make design decisions around which discovery networks they trust. For example if a Dat client decides to use the BitTorrent DHT to discover peers, and they are searching for a publicly shared Dat key with known contents, then because of the privacy design of the BitTorrent DHT it becomes public knowledge what key that client is searching for. -## 3.4 Efficient Versioning +A client could choose to only use discovery networks with certain privacy guarantees. For example a client could only connect to an approved list of sources that they trust, similar to SSL. As long as they trust each source, the encryption built into the Dat network protocol will prevent the Dat key they are looking for from being leaked. -Given a stream of binary data, Dat splits the stream into chunks using Rabin fingerprints, hashes each chunk, and arranges the hashes in a specific type of Merkle tree that allows for certain replication properties. Dat uses the chunk boundaries provided by Rabin fingerprinting to decide where to slice up the binary input stream. The Rabin implementation in Dat is tuned to produce a chunk every 16kb on average. This means for a 1MB file the initial chunking will produce around 64 chunks. +## 2.4 Incremental Versioning -If a 1 byte edit is made to the file, chunking again should produce 63 existing chunks and 1 new chunk. This allows for deduplication of similar file regions across versions, which means Dat can avoid retransmitting or storing the same chunk twice even if it appears in multiple files. +Given a stream of binary data, Dat splits the stream into chunks, hashes each chunk, and arranges the hashes in a specific type of Merkle tree that allows for certain replication properties. Dat is also able to fully or partially synchronize streams in a distributed setting even if the stream is being appended to. This is accomplished by using the messaging protocol to traverse the Merkle tree of remote sources and fetch a strategic set of nodes. Due to the low level message oriented design of the replication protocol different node traversal strategies can be implemented. -TODO example of using protocol messages to request a subset of nodes in a live sync scenario +There are two types of versioning performed automatically by Dat. Metadata is stored in a folder called `.dat` in the root folder of a repository, and data is stored as normal files in the root folder. -```js -var log = [ -{ - hash: hash(value + size), - size: value.length - value: -}, -{ - hash: hash(log[0].hash+log[2].hash+size), - size: log[0].size + log[1].size -}, -{ - hash: hash(value + size), - size: value.length - value: -} -] -``` +### Metadata Versioning -## 3.6 Network Privacy +Dat tries as much as possible to act as a one-to-one mirror of the state of a folder and all it's contents. When importing files, Dat uses a sorted depth-first recursion to list all the files in the tree. For each file it finds, it grabs the filesystem metadata (filename, Stat object, etc) and checks if there is already an entry for this filename with this exact metadata already represented in the Dat repository metadata. If the file with this metadata matches exactly the newest version of the file metadata stored in Dat, then this file will be skipped (no change). -On the Web today, with SSL, there is a guarantee that the traffic between your computer and the server is private. As long as you trust the server to not leak your logs, attackers who intercept your network traffic will not be able to read the HTTP traffic exchanged between you and the server. This is a fairly straightforward model as clients only have to trust a single server for some domain. +If the metadata differs from the current existing one (or there are no entries for this filename at all in the history), then this new metadata entry will be appended as the new 'latest' version for this file in the append-only SLEEP metadata content feed (described below). -There is an inherent tradeoff in peer to peer systems of source discovery vs. user privacy. The more sources you contact and ask for some data, the more sources you trust to keep what you asked for private. Our goal is to have Dat be configurable in respect to this tradeoff to allow application developers to meet their own privacy guidelines. +### Content Versioning -It is up to client programs to make design decisions around which discovery networks they trust. For example if a Dat client decides to use the BitTorrent DHT to discover peers, and they are searching for a publicly shared Dat key with known contents, then because of the privacy design of the BitTorrent DHT it becomes public knowledge what key that client is searching for. +In addition to storing a historical record of filesystem metadata, the content of the files themselves are also capable of being stored in a version controlled manner. The default storage system used in Dat simply stores the files as files. This has the advantage of being very straightforward for users to understand, but the downside of not storing old versions of content by default. -A client could choose to only use discovery networks with certain privacy guarantees. For example a client could only connect to an approved list of sources that they trust, similar to SSL. As long as they trust each source, the encryption built into the Dat network protocol will prevent the Dat key they are looking for from being leaked. +In contrast to other version control systems like Git, Dat by default only stores the current set of checked out files on disk in the repository folder, not old versions. It does store all previous metadata for old versions in `.dat`. Git for example stores all previous content versions and all previous metadata versions in the `.git` folder. Because Dat is designed for larger datasets, if it stored all previous file versions in `.dat`, then the `.dat` folder could easily fill up the users hard drive inadverntently. Therefore Dat has multiple storage modes based on usage. -### 3.6.2 Security +The default behavior is to store the current files only, but you can also run Dat using other storage provider modules. If you want to run an 'archival' node that keeps all previous versions, you can use the `content-addressable-blob-store` module instead of the default `dat-storage` module. For example, on a shared server with lots of storage you probably want to store all versions. However on a workstation machine that is only accessing a subset of one version, the default mode of storing all metadata plus the current set of downloaded files is acceptable, because you know the server has the full history. -Dat links are Ed25519 public keys which have a length of 32 bytes (64 characters when Base64 encoded). Every Dat repository has corresponding a private key that kept internally in the Dat metadata and never shared. - -Dat never exposes either the public or private key over the network. During the discovery phase the SHA256 hash of the public key is used as the discovery key. This means that the original key is impossible to discover (unless it was shared publicly through a separate channel) since only the hash of the key is exposed publicly. - -All messages in the Dat protocol are encrypted using the public key during transport. This means that unless you know the public key (e.g. unless the Dat link was shared with you) then you will not be able to discover or communicate with any member of the swarm for that Dat. Anyone with the public key can verify that messages (such as entries in a Dat Stream) were created by a holder of the private key. - -Dat does not provide an authentication mechanism. Instead it provides a capability system. Anyone with the Dat link is currently considered able to discover and access data. Do not share your Dat links publicly if you do not want them to be accessed. - -### SLEEP - -### What is SLEEP? +## 3. SLEEP SLEEP is the the on-disk format that Dat produces and uses. It is a set of 9 files that hold all of the metadata needed to list the contents of a Dat repository and verify the integrity of the data you receive. SLEEP is designed to work with REST, allowing servers to be plain HTTP file servers serving the static SLEEP files, meaning you can implement a Dat protocol client using HTTP with a static HTTP file server as the backend. @@ -238,11 +141,11 @@ SLEEP files contain metadata about the data inside a Dat repository, including c The acronym SLEEP is a slumber related pun on REST and stands for Syncable Lightweight Event Emitting Persistence. The Event Emitting part refers to how SLEEP files are append-only in nature, meaning they grow over time and new updates can be subscribed to as a realtime feed of events through the Dat protocol. -The SLEEP version used in Dat as of 2017 is SLEEP V2. The previous version is documented at http://specs.okfnlabs.org/sleep. +The SLEEP version described here, used in Dat as of 2017 is SLEEP V2. SLEEP V1 is documented at http://specs.okfnlabs.org/sleep. ### SLEEP Files -SLEEP is a set of 9 files that should be stored in a folder with the following names. In Dat, the files are stored in a folder called `.dat` in the top level of the repository. +SLEEP is a set of 9 files that should be stored with the following names. In Dat, the files are stored in a folder called `.dat` in the top level of the repository. ``` metadata.key @@ -256,9 +159,49 @@ content.bitfield content.tree ``` -The files prefixed with `content` store metadata about the primary data in a Dat repository, for example the raw binary contents of the files. The files prefixed with `metadata` store metadata about the files in the repository, for example the filenames, file sizes, file permissions. The `content` and `metadata` files are both serialized representations of Hypercore feeds, making SLEEP a set of two Hypercore feeds to represent a set of files, one for file data and one for file metadata. +The files prefixed with `content` store metadata about the primary data in a Dat repository, for example the raw binary contents of the files. The files prefixed with `metadata` store metadata about the files in the repository, for example the filenames, file sizes, and file permissions. The `content` and `metadata` files are both Hypercore feeds, making SLEEP simply a set of two Hypercore feeds. + +### Merkle Trees + +Registers in Dat use a specific method of encoding a Merkle tree where hashes are positioned by a scheme called binary in-order interval numbering or just simply "bin" numbering. This is just a specific, deterministic way of laying out the nodes in a tree. For example a tree with 7 nodes will always be arranged like this: + +``` +0 + 1 +2 + 3 +4 + 5 +6 +``` + +In Dat, the hashes of the chunks of files are always even numbers, at the wide end of the tree. So the above tree had four original values that become the even numbers: + +``` +chunk0 -> 0 +chunk1 -> 2 +chunk2 -> 4 +chunk3 -> 6 +``` + +In the resulting Merkle tree, the even and odd nodes store different information: + +- Evens - List of data hashes [chunk0, chunk1, chunk2, ...] +- Odds - List of Merkle hashes (hashes of child even nodes) [hash0, hash1, hash2, ...] + +These two lists get interleaved into a single register such that the indexes (position) in the register are the same as the bin numbers from the Merkle tree. + +All odd hashes are derived by hashing the two child nodes, e.g. given hash0 is `hash(chunk0)` and hash2 is `hash(chunk1)`, hash1 is `hash(hash0 + hash2)`. -#### SLEEP File Headers +For example a register with two data entries would look something like this (pseudocode): + +``` +0. hash(value0) +1. hash(hash(chunk0) + hash(chunk1)) +2. hash(value1) +``` + +### SLEEP File Headers The following structured binary format is used for `signatures`, `bitfield`, and `tree` files. The header contains metadata as well as information needed to decode the rest of the files after the header. SLEEP files are designed to be easy to append new data to at the end, easy to read arbitrary byte offsets in the middle, and are relatively flat, simple files that rely on the filesystem for the heavy lifting. @@ -410,7 +353,7 @@ This one has one root (1): 2 ``` -##### data +#### data The `data` file is only included in the SLEEP format for the `metadata.*` prefixed files which contains filesystem metadata and not actual file data. For the `content.*` files, the data is stored externally (in Dat it is stored as normal files on the filesystem and not in a SLEEP file). However you can configure Dat to use a `content.data` file if you want and it will still work. @@ -542,7 +485,7 @@ Using this scheme, to represent 32 bytes of data it takes at most 8 bytes of Ind If you write 4GB of data using on average 64KB data chunk size, your bitfield will be at most 32KB. -##### metadata.data +#### metadata.data This file is used to store content described by the rest of the `metadata.*` hypercore SLEEP files. Whereas the `content.*` SLEEP files describe the data stored in the actual data cloned in the Dat repository filesystem, the `metadata` data feed is stored inside the `.dat` folder along with the rest of the SLEEP files. @@ -672,17 +615,15 @@ These are the field defintions: - `mtime` - posix modified_at time - `mtime` - posix created_at time -#### Replication +## 4. Dat Network Protocol -The above file formats are designed to allow for sparse replication, meaning you can efficiently download only the metadata and data required to resolve a single byte region of a single file, which makes Dat suitable for a wide variety of streaming, real time and large dataset use cases. +The SLEEP format is designed to allow for sparse replication, meaning you can efficiently download only the metadata and data required to resolve a single byte region of a single file, which makes Dat suitable for a wide variety of streaming, real time and large dataset use cases. -##### Replication Protocol - -The Dat replication protocol is message based and stateless, making it possible to implement on a variety of network transport protocols including UDP and TCP. Both metadata and content feeds in SLEEP share the exact same replication protocol. +To take advantage of this, Dat includes a network protocol. It is message based and stateless, making it possible to implement on a variety of network transport protocols including UDP and TCP. Both metadata and content feeds in SLEEP share the exact same replication protocol. Individual messages are encoded using Protocol Buffers and there are ten message types using the following schema: -##### Wire Protocol +### Wire Protocol Over the wire messages are packed in the following lightweight container format @@ -696,7 +637,7 @@ The `header` value is a single varint that has two pieces of information, the in To generate this varint, you bitshift the 4-bit type integer onto the end of the channel identifier, e.g. `channel << 4 | <4-bit-type>`. -##### Feed +### Feed Type 0, should be the first message sent on a channel. @@ -710,7 +651,7 @@ message Feed { } ``` -##### Handshake +### Handshake Type 1. Overall connection handshake. Should be sent just after the feed message on the first channel only (metadata). @@ -724,7 +665,7 @@ message Handshake { } ``` -##### Status +### Status Type 2. Message indicating state changes. Used to indicate whether you are uploading and/or downloading. @@ -737,7 +678,7 @@ message Status { } ``` -##### Have +### Have Type 3. How you tell the other peer what blocks of data you have or don't have. You should only send Have messages to peers who have expressed interest in this region with Want messages. @@ -772,7 +713,7 @@ uncompressed-sequence = varint( ) + (bitfield) ``` -##### Unhave +### Unhave Type 4. How you communicate that you deleted or removed a block you used to have. @@ -784,7 +725,7 @@ message Unhave { } ``` -##### Want +### Want Type 5. How you ask the other peer to subscribe you to Have messages for a region of blocks. The `length` value defaults to Infinity or feed.length (if not live). @@ -795,7 +736,7 @@ message Want { } ``` -##### Unwant +### Unwant Type 6. How you ask to unsubscribe from Have messages for a region of blocks from the other peer. You should only Unwant previously Wanted regions, but if you do Unwant something that hasn't been Wanted it won't have any effect. The `length` value defaults to Infinity or feed.length (if not live). @@ -806,7 +747,7 @@ message Unwant { } ``` -##### Request +### Request Type 7. Request a single block of data. @@ -826,7 +767,7 @@ message Request { } ``` -##### Cancel +### Cancel Type 8. Cancel a previous Request message that you haven't received yet. @@ -838,7 +779,7 @@ message Cancel { } ``` -##### Data +### Data Type 9. Sends a single block of data to the other peer. You can send it in response to a Request or unsolicited on it's own as a friendly gift. The data includes all of the Merkle tree parent nodes needed to verify the hash chain all the way up to the Merkle roots for this block. Because you can produce the direct parents by hashing the block, only the roots and 'uncle' hashes are included (the siblings to all of the parent nodes). @@ -864,3 +805,55 @@ message Data { } } ``` + +# 5. Existing Work + +Dat is inspired by a number of features from existing systems. + +## Git + +Git popularized the idea of a directed acyclic graph (DAG) combined with a Merkle tree, a way to represent changes to data where each change is addressed by the secure hash of the change plus all ancestor hashes in a graph. This provides a way to trust data integrity, as the only way a specific hash could be derived by another peer is if they have the same data and change history required to reproduce that hash. This is important for reproducibility as it lets you trust that a specific git commit hash refers to a specific source code state. + +Decentralized version control tools for source code like Git provide a protocol for efficiently downloading changes to a set of files, but are optimized for text files and have issues with large files. Solutions like Git-LFS solve this by using HTTP to download large files, rather than the Git protocol. GitHub offers Git-LFS hosting but charges repository owners for bandwidth on popular files. Building a distributed distribution layer for files in a Git repository is difficult due to design of Git Packfiles which are delta compressed repository states that do not easily support random access to byte ranges in previous file versions. + +## BitTorrent + +BitTorrent implements a swarm based file sharing protocol for static datasets. Data is split into fixed sized chunks, hashed, and then that hash is used to discover peers that have the same data. An advantage of using BitTorrent for dataset transfers is that download bandwidth can be fully saturated. Since the file is split into pieces, and peers can efficiently discover which pieces each of the peers they are connected to have, it means one peer can download non-overlapping regions of the dataset from many peers at the same time in parallel, maximizing network throughput. + +Fixed sized chunking has drawbacks for data that changes (see LBFS above). BitTorrent assumes all metadata will be transferred up front which makes it impractical for streaming or updating content. Most BitTorrent clients divide data into 1024 pieces meaning large datasets could have a very large chunk size which impacts random access performance (e.g. for streaming video). + +Another drawback of BitTorrent is due to the way clients advertise and discover other peers in absence of any protocol level privacy or trust. From a user privacy standpoint, BitTorrent leaks what users are accessing or attempting to access, and does not provide the same browsing privacy functions as systems like SSL. + +## Kademlia Distributed Hash Table + +Kademlia is a distributed hash table, a distributed key/value store that can serve a similar purpose to DNS servers but has no hard coded server addresses. All clients in Kademlia are also servers. As long as you know at least one address of another peer in the network, you can ask them for the key you are trying to find and they will either have it or give you some other people to talk to that are more likely to have it. + +If you don't have an initial peer to talk to you, most clients use a bootstrap server that randomly gives you a peer in the network to start with. If the bootstrap server goes down, the network still functions as long as other methods can be used to bootstrap new peers (such as sending them peer addresses through side channels like how .torrent files include tracker addresses to try in case Kademlia finds no peers). + +Kademlia is distinct from previous DHT designs due to its simplicity. It uses a very simple XOR operation between two keys as its "distance" metric to decide which peers are closer to the data being searched for. On paper it seems like it wouldn't work as it doesn't take into account things like ping speed or bandwidth. Instead its design is very simple on purpose to minimize the amount of control/gossip messages and to minimize the amount of complexity required to implement it. In practice Kademlia has been extremely successful and is widely deployed as the "Mainline DHT" for BitTorrent, with support in all popular BitTorrent clients today. + +Due to the simplicity in the original Kademlia design a number of attacks such as DDOS and/or sybil have been demonstrated. There are protocol extensions (BEPs) which in certain cases mitigate the effects of these attacks, such as BEP 44 which includes a DDOS mitigation technique. Nonetheless anyone using Kademlia should be aware of the limitations. + +## Peer to Peer Streaming Peer Protocol (PPSPP) + +PPSPP ([IETF RFC 7574](https://datatracker.ietf.org/doc/rfc7574/?include_text=1)) is a protocol for live streaming content over a peer to peer network. In it they define a specific type of Merkle Tree that allows for subsets of the hashes to be requested by a peer in order to reduce the time-till-playback for end users. BitTorrent for example transfers all hashes up front, which is not suitable for live streaming. + +Their Merkle trees are ordered using a scheme they call "bin numbering", which is a method for deterministically arranging an append-only log of leaf nodes into an in-order layout tree where non-leaf nodes are derived hashes. If you want to verify a specific node, you only need to request its sibling's hash and all its uncle hashes. PPSPP is very concerned with reducing round trip time and time-till-playback by allowing for many kinds of optimizations, such as to pack as many hashes into datagrams as possible when exchanging tree information with peers. + +Although PPSPP was designed with streaming video in mind, the ability to request a subset of metadata from a large and/or streaming dataset is very desirable for many other types of datasets. + +## WebTorrent + +With WebRTC browsers can now make peer to peer connections directly to other browsers. BitTorrent uses UDP sockets which aren't available to browser JavaScript, so can't be used as-is on the Web. + +WebTorrent implements the BitTorrent protocol in JavaScript using WebRTC as the transport. This includes the BitTorrent block exchange protocol as well as the tracker protocol implemented in a way that can enable hybrid nodes, talking simultaneously to both BitTorrent and WebTorrent swarms (if a client is capable of making both UDP sockets as well as WebRTC sockets, such as Node.js). Trackers are exposed to web clients over HTTP or WebSockets. + +## InterPlanetary File System + +IPFS is a family of application and network protocols that have peer to peer file sharing and data permanence baked in. IPFS abstracts network protocols and naming systems to provide an alternative application delivery platform to todays Web. For example, instead of using HTTP and DNS directly, in IPFS you would use LibP2P streams and IPNS in order to gain access to the features of the IPFS platform. + +## Certificate Transparency/Secure Registers + +The UK Government Digital Service have developed the concept of a register which they define as a digital public ledger you can trust. In the UK government registers are beginning to be piloted as a way to expose essential open data sets in a way where consumers can verify the data has not been tampered with, and allows the data publishers to update their data sets over time. + +The design of registers was inspired by the infrastructure backing the Certificate Transparency project, initated at Google, which provides a service on top of SSL certificates that enables service providers to write certificates to a distributed public ledger. Anyone client or service provider can verify if a certificate they received is in the ledger, which protects against so called "rogue certificates". \ No newline at end of file diff --git a/papers/dat-paper.pdf b/papers/dat-paper.pdf index 002e301..03fc9c4 100644 Binary files a/papers/dat-paper.pdf and b/papers/dat-paper.pdf differ -- cgit v1.2.3