Title = "LBRY: A Decentralized Digital Content Marketplace"
area = "Internet"
[seriesInfo]
name = "Internet-Draft"
value = "draft-grintsvayg-00"
stream = "IETF"
status = "informational"
date = 2018-08-21T00:00:00Z
[[author]]
initials="A."
surname="Grintsvayg"
fullname="Alex Grintsvayg"
%%%
# LBRY: A Decentralized Digital Content Marketplace
A> Please excuse the unfinished state of this paper. It is being actively worked on. The content here is made available early because it contains useful information for developers.
A> For more technical information about LBRY, visit [lbry.tech](https://lbry.tech).
LBRY is a protocol for accessing and publishing digital content in a global, decentralized marketplace. Clients can use LBRY to publish, host, find, download, and pay for content — books, movies, music, or anything else. Anyone can participate and no permission is required, nor can anyone be blocked from participating. The system is distributed, so no single entity has unilateral control, nor will the removal of any single entity prevent the system from functioning.
This document defines the LBRY protocol, its components, and how they fit together. At its core, LBRY consists of several discrete components that are used together in order to provide the end-to-end capabilities of the protocol. There are two distributed data stores (blockchain and DHT), a peer-to-peer protocol for exchanging data, and several specifications for data structure, transformation, and retrieval.
This document assumes that the reader is familiar with Bitcoin and blockchain technology. It does not attempt to document the Bitcoin protocol or explain how it works. The [Bitcoin developer reference](https://bitcoin.org/en/developer-reference) is recommended for anyone wishing to understand the technical details.
<dd>The output of a cryptographic hash function is applied to a blob. Hashes are used to uniquely identify blobs and to verify that the contents of the blob are correct. Unless otherwise specified, LBRY uses SHA384 as the hash function.</dd>
The LBRY blockchain is a fork of the [Bitcoin](https://bitcoin.org/bitcoin.pdf) blockchain, with substantial modifications. This document will not cover or specify any aspects of LBRY that are identical to Bitcoin, and will instead focus on the differences.
A support contains a claim ID, and amount, and nothing else. Supports function analogously to claims in terms of [Claim Operations](#claim-operations) and [Claim Statuses](#claim-statuses), with the exception that they cannot be updated.
The claimtrie is implemented as a [Merkle tree](https://en.wikipedia.org/wiki/Merkle_tree) that maps names to claims. Claims are stored as leaf nodes in the tree. Names are stored as the path from the root node to the leaf node.
The _root hash_ is the hash of the root node. It is stored in the header of each block in the blockchain. Nodes in the LBRY network use the root hash to efficiently and securely validate the state of the claimtrie.
Multiple claims can exist for the same name. They are all stored in the leaf node for that name, sorted in decreasing order by the total amount of credits backing each claim.
An _abandoned_ claim is one that was withdrawn by its creator or current owner. Spending a transaction that contains a claim will cause that claim to become abandoned.
While data related to abandoned claims technically still resides in the blockchain, it is improper to use this data to fetch the associated content, and active claims signed by abandoned identities will no longer be reported as valid.
An _active_ claim is an accepted and non-abandoned claim that has been in the blockchain for an algorithmically determined number of blocks. This length of time required is called the _activation delay_.
If the claim is an update to an already active claim, is the first claim for a name, or does not affect the sort order at the leaf for a name, the activation delay is 0 (i.e. the claim becomes active in the same block it is accepted).
Otherwise, the activation delay is determined by a formula covered in [Claimtrie Transitions](#claimtrie-transitions). The formula's variable inputs are the height of the current block, the height at which the claim was accepted, and the height at which the relevant claimtrie state for the name being considered last changed.
The sum of the amount of an active claim and all active supports is called it's _effective amount_. Only the effective amount affects the sort order of a claimtrie leaf.
A _controlling_ claim is the active claim that is first in the sort order at a leaf. That is, it has the highest total effective amount of all claims with the same name.
1. Otherwise, a takeover is occurring. Set the takeover height for this name to the current height, recalculate which claims and supports are now active, and return to step 1.
The purpose of 3 is to handle the case when multiple competing claims are made on the same name in different blocks, and one of those claims becomes active but another still-inactive claim has the greatest amount. Step 3 will cause the greater claim to also activate and become the controlling claim.
If a claim does not become active immediately, it becomes active at the block heigh determined by the following formula:
```
C + min(4032, floor((H-T) / 32))
```
Where:
- C = claim height (height when the claim was accepted)
- H = current height
- T = takeover height (the most recent height at which the relevant claimtrie state for the name changed)
In written form, the delay before a claim becomes active is equal to the claim’s height minus height of the last takeover, divided by 32. The delay is capped at 4032 blocks, which is 7 days of blocks at 2.5 minutes per block (our target block time). The max delay is reached 224 (7x32) days after the last takeover.
The purpose of this delay function is to give long-standing claimants time to respond to changes, while still keeping takeover times reasonable and allowing recent or contentious claims to change state quickly.
**Block 1031:** Claim B activates. It has 20LBC, while claim A has 24LBC (10 original + 14 from support X). There is no takeover, and claim A remains controlling.
**Block 1051:** Claim C activates. It has 50LBC, while claim A has 24LBC, so a takeover is initiated. The takeover height for this name is set to 1051, and therefore the activation delay for all the claims becomes `min(4032, floor((1051-1051) / 32)) = 0`. All the claims become active. The totals for each claim are recalculated, and claim D becomes controlling because it has the highest total.
Names in the claimtrie are normalized to avoid confusion due to Unicode equivalence or casing. All names are converted using [Unicode Normalization Form D](http://unicode.org/reports/tr15/#Norm_Forms) (NFD), then lowercased using the en_US locale when possible.
<!-- fix me - @grin does SPV need a mention inside of the document? ->
URLs are human-readable references to claims. All URLs:
1. must contain a name (see [Claim Properties](#claim-properties))
2. and resolve to a single, specific claim for that name
The ultimate purpose of much of the claim and blockchain design is to provide short, human-readable URLs to all active claims that can be trustfully resolved by clients that have don't have a full copy of the blockchain (i.e. [Simplified Payment Verification](https://lbry.tech/glossary#spv) wallets).
#### Components
<!-- done -->
A URL is a name with one or more modifiers. A bare name on its own will resolve to the [controlling claim](#controlling) at the latest block height. Common URL structures are:
##### Stream Claim Name
A basic claim for a name.
```
lbry://meet-lbry
```
##### Channel Claim Name
A basic claim for a channel.
```
lbry://@lbry
```
##### Channel Claim Name and Stream Claim Name
A URL containing both a channel and a stream claim name. URLs containing both are resolved in two steps. First, the channel is resolved to it's associated claim. Then the stream claim name is resolved to get the appropriate claim from among the claims in the channel.
```
lbry://@lbry/meet-lbry
```
##### Claim ID
A claim for this name with this claim ID. Partial prefix matches are allowed (see [Resolution](#resolution)).
```
lbry://meet-lbry#7a0aa95c5023c21c098
lbry://meet-lbry#7a
lbry://@lbry#3f/meet-lbry
```
##### Claim Sequence
The Nth claim for this name, in the order the claims entered the blockchain. N must be a positive number. This can be used to determine which claim came first, rather than which claim has the most support.
```
lbry://meet-lbry:1
lbry://@lbry:1/meet-lbry
```
**Bid Position:**### URLs
<!-- fix me - @grin does SPV need a mention inside of the document? ->
URLs are human-readable references to claims. All URLs:
1. must contain a name (see [Claim Properties](#claim-properties))
2. and resolve to a single, specific claim for that name
The ultimate purpose of much of the claim and blockchain design is to provide human-readable URLs that can be trustfully resolved by clients that have don't have a full copy of the blockchain (i.e. [Simplified Payment Verification](https://lbry.tech/glossary#spv) wallets).
A URL is a name with one or more modifiers. A bare name on its own will resolve to the [controlling claim](#controlling) at the latest block height. Common URL structures are:
A URL containing both a channel and a stream claim name. URLs containing both are resolved in two steps. First, the channel is resolved to it's associated claim. Then the stream claim name is resolved to get the appropriate claim from among the claims in the channel.
The _n_th claim for this name, in the order the claims entered the blockchain. _n_ must be a positive number. This can be used to resolve claims in the order in which they were recorded, rather than by which claim has the most support.
The _n_th claim for this name, in order of most support to least total support. _n_ must be a positive number. This is useful for resolving non-winning bids in bid order.
Get all claims for the claim name whose IDs start with the given `ClaimID`. Sort the claims in ascending order by block height and position within the block. Return the first claim.
Get all claims for the claim name. Sort the claims in ascending order by block height and position within the block. Return the Nth claim, where N is the given `ClaimSequence` value.
Get all claims for the claim name. Sort the claims in descending order by total effective amount. Return the Nth claim, where N is the given `BidSequence` value.
If both a channel name and a claim name is present, resolution happens in two steps. First, remove the `/` and `StreamClaimNameAndModifier` from the path, and resolve the URL as if it only had a `ChannelClaimNameAndModifier`. Then get the list of all claims in that channel. Finally, resolve the `StreamClaimNameAndModifier` as if it was its own URL, but instead of considering all claims, only consider the set of claims in the channel.
Technically, if multiple claims for the same name exist inside the same channel, they are resolved via the same resolution rules applied entirely within the sub-scope of the channel. Pragmatically, it rarely makes sense for channels to use the same name twice and support for this functionality may be unreliable in current tooling.
The most contentious aspect of this design has been the choice to resolve naked names (sometimes called _vanity names_) to the claim with the largest number of staked credits.
First, it is important to note the problems in existing domain allocation design. Most existing public name schemes are first-come, first-serve with a fixed price. This leads to several bad outcomes:
1. Speculation and extortion. Entrepreneurs are incentivized to register common names even if they don't intend to use them, in hopes of selling them to the proper owner in the future for an exorbitant price. While speculation in general can have positive externalities, in this case it is pure value extraction. It also harms the user experience of users, who will see the vast majority of URLs sitting unused (c.f. Namecoin).
1. Bureaucracy and transaction costs. While a centralized system can allow for an authority to use a process to reassign names based on trademark or other common use reasons, this system is also imperfect. Most importantly, it is a censorship point and an avenue for complete exclusion. Additionally, such processes are often arbitrary, change over time, involve significant transaction costs, and _still_ lead to names being used in ways that are contrary to user expectation (e.g. [nissan.com](http://nissan.com)).
1. Inefficencies from price controls. Any system that does not allow a price to float completely freely creates inefficiencies. If the set price is too low, we facilitate speculation and rent-seeking. If the price is too high, we see people excluded from a good that it would otherwise be beneficial for them to purchase.
Thus, we need an algorithmic design built into consensus that allows URLs to flow to their highest valued use. Following [Coase](https://en.wikipedia.org/wiki/Coase_theorem), this design allows for clearly defined rules, low transaction costs, and no information asymmetry, ensuring minimal inefficiency in URL allocation.
It's also important to note that _only_ vanity URLs have this property. Extremely short, memorable URLs like `lbry://myclaimname#a` exist and are available for the minimal cost of issuing a transaction.
To enable [claim operations](#claim-operations), three new opcodes were added to the blockchain scripting language: `OP_CLAIM_NAME`, `OP_SUPPORT_CLAIM`, and `OP_UPDATE_CLAIM` (in Bitcoin they are respectively `OP_NOP6`, `OP_NOP7`, and `OP_NOP8`). Each op code will push a zero on to the execution stack, and will trigger the claimtrie to perform calculations necessary for each operation. Below are the three supported transactions scripts using these opcodes.
`<pubKey>` can be any valid Bitcoin payout script, so a claimtrie script is also a pay-to-pubkey script to a user-controlled address. Note that the zeros pushed onto the stack by the claimtrie opcodes and vectors are all dropped by `OP_2DROP` and `OP_DROP`. This means that claimtrie transactions exist as prefixes to Bitcoin payout scripts and can be spent just like standard transactions.
For example, a claim transaction setting the name “Fruit” to “Apple” and using a pay-to-pubkey script will have the following payout script:
Like any standard Bitcoin transaction output script, it will be associated with a transaction hash and output index. The transaction hash and output index are concatenated and hashed to create the claimID for this claim. For the example above, let's say the above transaction hash is `7560111513bea7ec38e2ce58a58c1880726b1515497515fd3f470d827669ed43` and the output index is `1`. Then the claimID would be `529357c3422c6046d3fec76be2358004ba22e323`.
A support for this bid will have the following payout script:
And now let's say we want to update the original claim to change the value to “Banana”. An update transaction has a special requirement that it must spend the existing claim that it wishes to update in its redeem script. Otherwise, it will be considered invalid and will not make it into the claimtrie. Thus it will have the following redeem script:
```
<signature><pubKeyForAddressOne>
```
This is identical to the standard way of redeeming a pay-to-pubkey script in Bitcoin.
The address version byte is set to `0x55` for standard (pay-to-public-key-hash) addresses and `0x7a` for multisig (pay-to-script-hash) addresses. P2PKH addresses start with the letter `b`, and P2SH addresses start with `r`.
The proof-of-work target is adjusted every block to better adapt to sudden changes in hash rate. The exact adjustment algorithm can be seen [here](https://github.com/lbryio/lbrycrd/blob/master/src/lbry.cpp).
LBRY uses a combination of SHA256, SHA512, and RIPEMD160. The exact hashing algorithm can be seen [here](https://github.com/lbryio/lbrycrd/blob/master/src/hash.cpp#L18).
The block reward schedule was adjusted to provide an initial testing period, a quick ramp-up to max block rewards, then a logarithmic decay to 0. The source for the algorithm is [here](https://github.com/lbryio/lbrycrd/blob/master/src/main.cpp#L1594).
Claim metadata is stored in a serialized format using [Protocol Buffers](https://developers.google.com/protocol-buffers/). This was chosen for several reasons:
- **Extensibility**. The metadata structure could grow to encompass thousands of fields for dozens of types of content. It should be easy to modify the structure while maintaining backward compatibility. Blockchain data is permanent and cannot be migrated.
- **Compactness**. Blockchain space is expensive. Data should be stored as compactly as possible.
- **Interoperability**. These definitions will be used by many projects written in different languages. Protocol buffers are language-independent and have great support for most popular languages.
No enforcement or validation on metadata happens at the blockchain level. Instead, metadata encoding, decoding, and validation is done by clients. This allows evolution of the metadata without changes to consensus rules.
A useful index of available content must be succinct yet meaningful. It should be machine-readable, comprehensive, and should ideally point you toward the content you’re looking for. LBRY achieves this by defining a set of common properties for streams.
Because the metadata structure can and does change frequently, a complete specification is omitted from this document. Instead, [github.com/lbryio/types](https://github.com/lbryio/types) should be consulted for the precise definition of current metadata structure.
Channels are the unit of identity in the LBRY system. A channel is a claim that start with `@` and contains a metadata structure for identities rather than content. The most important part of channel's metadata is the public key. Claims belonging to a channel are signed with the corresponding private key. A valid signature proves channel membership.
The purpose of channels is to allow content to be clustered under a single pseudonym or identity. This allows publishers to easily list all their content, maintain attribution, and build their brand.
The unit of data in the LBRY network is called a _blob_. A blob is an encrypted chunk of data up to 2MiB in size. Each blob is indexed by its _blob hash_, which is a SHA384 hash of the blob contents. Addressing blobs by their hash protects against naming collisions and ensures that the content you get is what you expect.
Blobs are encrypted using AES-256 in CBC mode and PKCS7 padding. In order to keep each encrypted blob at 2MiB max, a blob can hold at most 2097151 bytes (2MiB minus 1 byte) of plaintext data. The source code for the exact algorithm is available [here](https://github.com/lbryio/lbry.go/blob/master/stream/blob.go). The encryption key and IV for each blob is stored as described below.
Multiple blobs are combined into a _stream_. A stream may be a book, a movie, a CAD file, etc. All content on the network is shared as streams. Every stream begins with the _manifest blob_, followed by one or more _content blobs_. The content blobs hold the actual content of the stream. The manifest blob contains information necessary to find the content blobs and decode them into a file. This includes the hashes of the content blobs, their order in the stream, and cryptographic material for decrypting them.
A manifest blob's contents are encoded using a canonical JSON encoding. The JSON encoding must be canonical to support consistent hashing and validation. The encoding is the same as standard JSON, but adds the following rules:
- Object keys must be quoted and lexicographically sorted.
- All strings are hex-encoded. Hex letters must be lowercase.
- Whitespace before, after, or between tokens is not permitted.
- Floating point numbers, leading zeros, and "minus 0" for integers are not permitted.
- Trailing commas after the last item in an array or object are not permitted.
The `key` field contains the key to decrypt the stream, and is optional. The key may be stored by a third party and made available to a client when presented with proof that the content was purchased. The `version` field is always 1. It is intended to signal structure changes in future versions of this protocol. The `length` field for each blob is the length of the encrypted blob, not the original file chunk.
A file must be encoded into a stream before it can be published. Encoding involves breaking the file into chunks, encrypting the chunks into content blobs, and creating the manifest blob. Here are the steps:
After a [[stream]] is encoded, it must be _announced_ to the network. Announcing is the process of letting other nodes on the network know that you have content available for download. The LBRY network tracks announced content using a distributed hash table.
_Distributed hash tables_ (or DHTs) have proven to be an effective way to build a decentralized content network. Our DHT implementation follows the [Kademlia](https://pdos.csail.mit.edu/~petar/papers/maymounkov-kademlia-lncs.pdf)
A distributed hash table is a key-value store that is spread over multiple nodes in a network. Nodes may join or leave the network anytime, with no central coordination necessary. Nodes communicate with each other using a peer-to-peer protocol to advertise what data they have and what they are best positioned to store.
When a host connects to the DHT, it announces the hash for every [[blob]] it wishes to share. Downloading a blob from the network requires querying the DHT for a list of hosts that announced that blob’s hash (called _peers_), then requesting the blob from the peers directly.
A host announces a hash to the DHT in two steps. First, the host looks for nodes that are closest to the target hash. Then the host asks those nodes to store the fact that the host has the target hash available for download.
Finding the closest nodes is done via iterative `FindNode` DHT requests. The host starts with the closest nodes it knows about and sends a `FindNode(target_hash)` request to each of them. If any of the requests return nodes that are closer to the target hash, the host sends `FindNode` requests to those nodes to try to get even closer. When the `FindNode` requests no longer return nodes that are closer, the search ends.
Once the search is over, the host takes the 8 closest nodes it found and sends a `Store(target_hash)` request to them. The nodes receiving this request store the fact that the host is a peer for the target hash.
A client wishing to download a [[stream]] must first query the [[DHT]] to find [[peers]] hosting the [[blobs]] in that stream, then contact those peers to download the blobs directly.
Querying works almost the same way as [[announcing]]. A client looking for a target hash will start by sending iterative `FindValue(target_hash)` requests to the nodes it knows that are closest to the target hash. If a node receives a `FindValue` request and knows of any peers for the target hash, it will respond with a list of those peers. Otherwise, it will respond with the closest nodes to the target hash that it knows about. The client then queries those closer nodes using the same `FindValue` call. This way, each call either finds the client some peers, or brings it closer to finding those peers. If no peers are found and no closer nodes are being returned, the client will determine that the target hash is not available and will give up.
Downloading a blob from a peer is governed by the _Blob Exchange Protocol_. It is used by hosts and clients to check blob availability, exchange blobs, and negotiate data prices. The protocol is an RPC protocol using Protocol Buffers and the gRPC framework. It has five types of requests.
fixme: protocol does not **negotiate** anything right now. It just checks the price. Should we include negotiation in v1?
##### PriceCheck
PriceCheck gets the price that the server is charging for data transfer. It returns the prices in [[deweys]] per KB.
##### DownloadCheck
DownloadCheck checks whether the server has certain blobs available for download. For each hash in the request, the server returns a true or false to indicate whether the blob is available.
Download requests the blob for a given hash. The response contains the blob, its hash, and the address where to send payment for the data transfer. If the blob is not available on the server, the response will instead contain an error.
UploadCheck asks the server whether blobs can be uploaded to it. For each hash in the request, the server returns a true or false to indicate whether it would accept a given blob for upload. In addition, if any of the hashes in the request is a stream hash and the server has the manifest blob for that stream but is missing some content blobs, it may include the hashes of those content blobs in the response.
Upload sends a blob to the server. If uploading many blobs, the client should use the UploadCheck request to check which blobs the server actually needs. This avoids needlessly uploading blobs that the server already has. If a client tries to upload too many blobs that the server does not want, this may be consider a denial of service attack.
In order for a client to download content, there must be hosts online that have the content the client wants, when the client wants it. To incentivize the continued hosting of data, the blob exchange protocol supports data upload and payment for data. _Reflectors_ are hosts that accept data uploads. They rehost (reflect) the uploaded data and charge for downloads. Using a reflector is optional, but most publishers will probably choose to use them. Doing so obviates the need for the publisher's server to be online and connectable, which can be especially useful for mobile clients or those behind a firewall.
The current version of the protocol does not support sophisticated price negotiation between clients and hosts. The host simply chooses the price it will charge. Clients check this price before downloading, and pay the price after the download is complete. Future protocol versions will include more options for price negotiation, as well as stronger proofs of payment.
