mirror of
https://github.com/logos-messaging/negentropy.git
synced 2026-01-02 14:03:07 +00:00
docs
This commit is contained in:
Doug Hoyte
parent
ed5f8e7e48
commit
87964be9ea
264
README.md
264
README.md
@ -1,30 +1,30 @@
|
||||
|
||||
|
||||
This repo contains the protocol specification, reference implementations, and tests for the negentropy set-reconciliation protocol.
|
||||
This repo contains the protocol specification, reference implementations, and tests for the negentropy set-reconciliation protocol. See [our article](https://logperiodic.com/rbsr.html) for a detailed description.
|
||||
|
||||
<!-- TOC FOLLOWS -->
|
||||
<!-- START OF TOC -->
|
||||
<!-- DO NOT EDIT! Auto-generated by md-toc: https://github.com/hoytech/md-toc -->
|
||||
|
||||
* [Introduction](#introduction)
|
||||
* [Protocol](#protocol)
|
||||
* [Data Requirements](#data-requirements)
|
||||
* [Setup](#setup)
|
||||
* [Alternating Messages](#alternating-messages)
|
||||
* [Algorithm](#algorithm)
|
||||
* [Frame Size Limits](#frame-size-limits)
|
||||
* [Data Requirements](#data-requirements)
|
||||
* [Setup](#setup)
|
||||
* [Bounds](#bounds)
|
||||
* [Fingerprints](#fingerprints)
|
||||
* [Alternating Messages](#alternating-messages)
|
||||
* [Algorithm](#algorithm)
|
||||
* [Frame Size Limits](#frame-size-limits)
|
||||
* [Definitions](#definitions)
|
||||
* [Varint](#varint)
|
||||
* [Bound](#bound)
|
||||
* [Range](#range)
|
||||
* [Message](#message)
|
||||
* [Analysis](#analysis)
|
||||
* [Reference Implementation APIs](#reference-implementation-apis)
|
||||
* [C++](#c)
|
||||
* [Javascript](#javascript)
|
||||
* [Rust](#rust)
|
||||
* [Adversarial Input](#adversarial-input)
|
||||
* [Pre-computing Fingerprints](#pre-computing-fingerprints)
|
||||
* [Use-Cases](#use-cases)
|
||||
* [Copyright](#copyright)
|
||||
* [Varint](#varint)
|
||||
* [Bound](#bound)
|
||||
* [Range](#range)
|
||||
* [Message](#message)
|
||||
* [Implementations](#implementations)
|
||||
* [Testing](#testing)
|
||||
* [Applications](#applications)
|
||||
* [Author](#author)
|
||||
|
||||
<!-- END OF TOC -->
|
||||
|
||||
|
||||
@ -36,6 +36,7 @@ Suppose two participants on a network each have a set of records that they have
|
||||
|
||||
Negentropy is based on Aljoscha Meyer's work on "Range-Based Set Reconciliation" ([overview](https://github.com/AljoschaMeyer/set-reconciliation) / [paper](https://arxiv.org/abs/2212.13567) / [master's thesis](https://github.com/AljoschaMeyer/master_thesis/blob/main/main.pdf)).
|
||||
|
||||
This page is a technical description of the negentropy wire protocol and the various implementations. Read [our article](https://logperiodic.com/rbsr.html) for a comprehensive introduction to range-based set reconciliation.
|
||||
|
||||
|
||||
## Protocol
|
||||
@ -46,14 +47,14 @@ In order to use negentropy, you need to define some mappings from your data reco
|
||||
|
||||
* `record -> ID`
|
||||
* Typically a cryptographic hash of the entire record
|
||||
* The ID should be 32 bytes in length (although smaller IDs are supported too)
|
||||
* Two different records should not have the same ID (satisfied by using a cryptographic hash)
|
||||
* Two identical records should not have different IDs (records should be canonicalised prior to hashing, if necessary)
|
||||
* The ID must be 32 bytes in length
|
||||
* Different records should not have the same ID (satisfied by using a cryptographic hash)
|
||||
* Equivalent records should not have different IDs (records should be canonicalised prior to hashing, if necessary)
|
||||
* `record -> timestamp`
|
||||
* Although timestamp is the most obvious, any ordering criteria can be used. The protocol will be most efficient if records with similar timestamps are often downloaded/stored/generated together
|
||||
* Units can be anything (seconds, microseconds, etc) as long as they fit in an 64-bit unsigned integer
|
||||
* The largest 64-bit unsigned integer should be reserved as a special "infinity" value
|
||||
* Timestamps do *not* need to be unique (different records can have the same timestamp). If necessary, `0` can be used as the timestamp for every record
|
||||
* Timestamps do **not** need to be unique (different records can have the same timestamp). If necessary, `0` can be used as the timestamp for every record
|
||||
|
||||
Negentropy does not support the concept of updating or changing a record while preserving its ID. This should instead be modelled as deleting the old record and inserting a new one.
|
||||
|
||||
@ -61,27 +62,36 @@ Negentropy does not support the concept of updating or changing a record while p
|
||||
|
||||
The two parties engaged in the protocol are called the client and the server. The client is also called the *initiator*, because it creates and sends the first message in the protocol.
|
||||
|
||||
Each party should begin by sorting their records in ascending order by timestamp. If the timestamps are equivalent, records should be sorted lexically by their IDs. This sorted array and contiguous slices of it are called *ranges*. The *fingerprint* of a range is equal to the SHA-256 hash of the IDs of all records within this range (sorted as described above).
|
||||
Each party should begin by sorting their records in ascending order by timestamp. If the timestamps are equivalent, records should be sorted lexically by their IDs. This sorted array and contiguous slices of it are called *ranges*.
|
||||
|
||||
For the purpose of this specification, we will assume that records are always stored in arrays. However, implementations may provide more advanced storage data-structures such as trees.
|
||||
|
||||
### Bounds
|
||||
|
||||
Because each side potentially has a different set of records, ranges cannot be referred to by their indices in one side's sorted array. Instead, they are specified by lower and upper *bounds*. A bound is a timestamp and a variable-length ID prefix. In order to reduce the sizes of reconciliation messages, ID prefixes are as short as possible while still being able to separate records from their predecessors in the sorted array. If two adjacent records have different timestamps, then the prefix for a bound between them is empty.
|
||||
|
||||
Lower bounds are *inclusive* and upper bounds are *exclusive*, as is [typical in computer science](https://www.cs.utexas.edu/users/EWD/transcriptions/EWD08xx/EWD831.html). This means that given two adjacent ranges, the upper bound of the first is equal to the lower bound of the second. In order for a range to have full coverage over the universe of possible timestamps/IDs, the lower bound would have a 0 timestamp and all-0s ID, and the upper-bound would be the specially reserved "infinity" timestamp (max u64), and the ID doesn't matter.
|
||||
|
||||
When negotiating a reconciliation, the client and server should decide on a special `idSize` value. It must satisfy `8 <= idSize <= 32`. Using values less than the full 32 bytes will save bandwidth, at the expense of making collisions more likely.
|
||||
### Fingerprints
|
||||
|
||||
The *fingerprint* of a range is computed with the following algorithm:
|
||||
|
||||
* Compute the addition mod `2**256` of the element IDs (interpreted as 32-byte little-endian unsigned integers)
|
||||
* Concatenate with the number of elements in the range, encoded as a [variable-length integer](#varint)
|
||||
* Hash with SHA-256
|
||||
* Take the first 16 bytes
|
||||
|
||||
### Alternating Messages
|
||||
|
||||
After both sides have setup their sorted arrays, the client creates an initial message and sends it to the server. The server will then reply with another message, and the two parties continue exchanging messages until the protocol terminates (see below). After the protocol terminates, the client will have determined what IDs it has (and the server needs) and which it needs (and the server has).
|
||||
After both sides have setup their sorted arrays, the client creates an initial message and sends it to the server. The server will then reply with another message, and the two parties continue exchanging messages until the protocol terminates (see below). After the protocol terminates, the *client* will have determined what IDs it has (and the server needs) and which it needs (and the server has). If desired, it can then respectively upload and/or download the missing records.
|
||||
|
||||
The initial message consists of a protocol version byte followed by an ordered sequence of ranges. Subsequent messages are the same, except without the protocol version byte. Each range contains an upper bound, a mode, and a payload. The range's lower bound is the same as the previous range's upper bound (or 0, if it is the first range). The mode indicates what type of processing is needed for this range, and therefore how the payload should be parsed.
|
||||
Each message consists of a protocol version byte followed by an ordered sequence of ranges. Each range contains an upper bound, a mode, and a payload. The range's implied lower bound is the same as the previous range's upper bound (or 0, if it is the first range). The mode indicates what type of processing is needed for this range, and therefore how the payload should be parsed.
|
||||
|
||||
The modes supported are:
|
||||
|
||||
* `Skip`: No further processing is needed for this range. Payload is empty.
|
||||
* `Fingerprint`: Payload contains the fingerprint for this range.
|
||||
* `IdList`: Payload contains a complete list of IDs for this range.
|
||||
* `Continuation`: Indicates that the sender has more to send in the next reconciliation round, but is not sending now because of a configured [frame size limit](#frame-size-limits).
|
||||
* `UnsupportedProtocolVersion`: The server is unable to handle the protocol version suggested by the client.
|
||||
|
||||
If a message does not end in a range with an "infinity" upper bound, an implicit range with upper bound of "infinity" and mode `Skip` is appended. This means that an empty message indicates that all ranges have been processed and the sender believes the protocol can now terminate.
|
||||
|
||||
@ -91,9 +101,9 @@ Upon receiving a message, the recipient should loop over the message's ranges in
|
||||
|
||||
`IdList` ranges represent a complete list of IDs held by the sender. Because the receiver obviously knows the items it has, this information is enough to fully reconcile the range. Therefore, when the client receives an `IdList` range, it should reply with a `Skip` range. However, since the goal of the protocol is to ensure the *client* has this information, when a server receives an `IdList` range it should reply with its own ranges (typically `IdList` and/or skip ranges).
|
||||
|
||||
`Fingerprint` ranges contain enough information to determine whether or not the data items within a range are equivalent on each side, however determining the actual difference in IDs requires further recursive processing.
|
||||
`Fingerprint` ranges contain enough information to determine whether or not the set of data items within the range are equal on each side, however determining the actual difference in IDs requires further recursive processing.
|
||||
* Since `IdList` or `Skip` messages will always cause the client to terminate processing for the given ranges, these messages are considered *base cases*.
|
||||
* When the fingerprints on each side differ, the reciever should *split* its own range and send the results back in the next message. When splitting, the number of records within each sub-range should be considered. When small, an `IdList` range should be sent. When large, the sub-ranges should themselves be sent as `Fingerprint`s (this is the recursion).
|
||||
* When the fingerprints on each side differ, the reciever should *split* its own range and send the results back in the next message. When splitting, the number of records within each sub-range should be considered. If small, an `IdList` range should be sent. If large, the sub-ranges should themselves be sent as `Fingerprint`s (this is the recursion).
|
||||
* When a range is split, the sub-ranges should completely cover the original range's lower and upper bounds.
|
||||
* Unlike in Meyer's designs, an "empty" fingerprint is never sent to indicate the absence of items within a range. Instead, an `IdList` of length 0 is sent because it is smaller.
|
||||
* How to split the range is implementation-defined. The simplest way is to divide the records that fall within the range into N equal-sized buckets, and emit a `Fingerprint` sub-range for each of these buckets. However, an implementation could choose different grouping criteria. For example, events with similar timestamps could be grouped into a single bucket. If the implementation believes recent events are less likely to be reconciled, it could make the most recent bucket an `IdList` instead of `Fingerprint`.
|
||||
@ -109,7 +119,7 @@ If there are too many differences and/or they are too evenly distributed through
|
||||
|
||||
Because of this, implementations may support a *frame size limit* parameter. If configured, all messages created by this instance will be of length equal to or smaller than this number of bytes. After processing each message, any discovered differences will be included in the `have`/`need` arrays on the client.
|
||||
|
||||
To implement this, instead of sending all the ranges it has found that need syncing, the instance will send a smaller number of them to stay under the size limit. Remaining ranges will be transmitted in subsequent message rounds. This means that a frame size limit can increase the number of messaging round-trips.
|
||||
To implement this, instead of sending all the ranges it has found that need syncing, the instance will send a smaller number of them to stay under the size limit. Any other ranges that were sent are replied to with a single Fingerprint range so that they will be processed in subsequent message rounds. Frame size limits can increase the number of messaging round-trips and bandwidth consumed.
|
||||
|
||||
|
||||
## Definitions
|
||||
@ -132,207 +142,97 @@ Each encoded bound consists of an encoded timestamp and a variable-length disamb
|
||||
|
||||
Offsets are always non-negative since the upper bound's timestamp is always `>=` to the lower bound's timestamp, ranges in a message are always encoded in ascending order, and ranges never overlap.
|
||||
|
||||
* The `idPrefix` parameter's size is encoded in `length`, and can be between `0` and `idSize` bytes, inclusive. Efficient implementations will use the shortest possible prefix needed to separate the first record of this range from the last record of the previous range. If these records' timestamps differ, then the length should be 0, otherwise it should be the byte-length of their common prefix plus 1.
|
||||
* The size of `idPrefix` is encoded in `length`, and can be between `0` and `32` bytes, inclusive. Efficient implementations will use the shortest possible prefix needed to separate the first record of this range from the last record of the previous range. If these records' timestamps differ, then the length should be 0, otherwise it should be the byte-length of their common ID-prefix plus 1.
|
||||
|
||||
If the `idPrefix` length is less than `idSize` then the omitted trailing bytes are filled with 0 bytes.
|
||||
If the `idPrefix` length is less than `32` then the omitted trailing bytes are considered to be 0 bytes.
|
||||
|
||||
### Range
|
||||
|
||||
IDs are represented as byte-strings truncated to length `idSize`:
|
||||
IDs are represented as byte-strings of length `32`:
|
||||
|
||||
Id := Byte{idSize}
|
||||
Id := Byte{32}
|
||||
|
||||
A range consists of an upper bound, a mode, and a payload (determined by mode):
|
||||
|
||||
Range := <upperBound (Bound)> <mode (Varint)> <Skip | Fingerprint | IdList | Continuation | UnsupportedProtocolVersion>
|
||||
Range := <upperBound (Bound)> <mode (Varint)> <payload (Skip | Fingerprint | IdList)>
|
||||
|
||||
* If `mode = 0`, then payload is `Skip`, which is simply empty:
|
||||
|
||||
Skip :=
|
||||
|
||||
* If `mode = 1`, then payload is `Fingerprint`, the SHA-256 hash of all the IDs in this range sorted ascending by `(timestamp,id)` and truncated to `idSize`:
|
||||
* If `mode = 1`, then payload is `Fingerprint`, the [fingerprint](#fingerprints) of all the IDs in this range:
|
||||
|
||||
Fingerprint := <Id>
|
||||
Fingerprint := Byte{16}
|
||||
|
||||
* If `mode = 2`, the payload is `IdList`, a variable-length list of all IDs within this range:
|
||||
|
||||
IdList := <length (Varint)> <ids (Id)>*
|
||||
|
||||
* If `mode = 3`, then payload is `Continuation`, which is simply empty. If present, a continuation range should always be the last range in a message, and its upperBound should be "infinity".
|
||||
|
||||
Continuation :=
|
||||
|
||||
* If `mode = 4`, then payload is `UnsupportedProtocolVersion`. This is created by the server when it is unable to handle the protocol version indicated by the protocol version byte in the initial message. The server's preferred protocol version is stored in the timestamp of the `upperBound` of this range (offset by `0x60`). If the client can support this version then its reply should be another initial message using this protocol version.
|
||||
|
||||
UnsupportedProtocolVersion :=
|
||||
|
||||
### Message
|
||||
|
||||
A reconciliation message is an ordered list of ranges:
|
||||
A reconciliation message is a protocol version byte followed by an ordered list of ranges:
|
||||
|
||||
Message := <Range>*
|
||||
Message := <protocolVersion (Byte)> <Range>*
|
||||
|
||||
* Zero ranges represents an implicit `Skip` over the full universe of IDs.
|
||||
* The current protocol version is 1, represented by the byte `0x61`. Protocol version 2 will be `0x62`, and so forth.
|
||||
* If a server (non-initiator) receives a messages with a protocol version that it cannot handle, it should reply with a single byte containing the highest protocol version it can.
|
||||
|
||||
The initial message is prefixed by a protocol version byte:
|
||||
|
||||
InitialMessage := <protocolVersion (Byte)> <Message>
|
||||
|
||||
* The current protocol version is 0, represented by the byte `0x60`. Protocol version 1 will be `0x61`, and so forth.
|
||||
## Implementations
|
||||
|
||||
This section lists all the currently-known negentropy implementations. If you know of a new one, please let us know by [opening an issue](https://github.com/hoytech/negentropy/issues/new).
|
||||
|
||||
|
||||
## Analysis
|
||||
| **Language** | **Author** | **Status** | **Storage** |
|
||||
| ---- | ---- | ---- | ---- |
|
||||
| [C++](cpp/README.md) | reference | Stable | Vector, BTreeMem, BTreeLMDB, SubRange |
|
||||
| [Javascript](js/README.md) | reference | Stable | Vector |
|
||||
| [Rust](https://github.com/yukibtc/rust-negentropy) | Yuki Kishimoto | Stable | Vector |
|
||||
|
||||
If you are searching for a single record in an ordered array, binary search allows you to find the record with a logarithmic number of operations. This is because each operation cuts the search space in half. So, searching a list of 1 million items will take about 20 operations:
|
||||
|
||||
log(1e6)/log(2) = 19.9316
|
||||
|
||||
Range-based reconciliation uses a similar principle. Each communication divides items into their own buckets and compares the fingerprints of the buckets. If we always split into 2 buckets, and there was exactly 1 difference, we would cut the search-space in half on each operation.
|
||||
## Testing
|
||||
|
||||
For effective performance, negentropy requires minimising the number of "round-trips" between the two sides. A sync that takes 20 back-and-forth communications to determine a single difference would take unacceptably long. Fortunately we can expend a small amount of extra bandwidth by splitting our ranges into more than 2 ranges. This has the effect of increasing the base of the logarithm. For example, if we split it into 16 pieces instead:
|
||||
There is a conformance test-suite available in the `testing` directory.
|
||||
|
||||
log(1e6)/log(16) = 4.98289
|
||||
In order to test a new language you should create a "harness", which is a basic stdio line-based adapter for your implementation. See the [test/cpp/harness.cpp](test/cpp/harness.cpp) and [test/js/harness.js](test/js/harness.js) files for examples. Next, edit the file `test/Utils.pm` and configure how your harness should be invoked.
|
||||
|
||||
Additionally, each direction of the protocol's communication can result in a split, so since we are measuring round-trips, we divide this by two:
|
||||
Harnesses may require some setup before they are usable. For example, to use the C++ harness you must first run `make` within the `test/cpp/` directory.
|
||||
|
||||
log(1e6)/log(16)/2 = 2.49145
|
||||
Once setup, you should be able to run something like `perl test.pl cpp,js` from the `test/` directory. This will perform the following:
|
||||
|
||||
This means that in general, three round-trip communications will be required to synchronise two sets of 1 million records that differ by 1 record. With an `idSize` of 16, each communication will consume `16*16 + overhead` -- roughly 300 bytes. So total bandwidth in one direction would be about 900 bytes and the other direction about 600 bytes.
|
||||
* For each combination of language run the following fuzz tests:
|
||||
* Client has all records
|
||||
* Server has all records
|
||||
* Both have all records
|
||||
* Client is missing some and server is missing some
|
||||
|
||||
What if they differ by multiple records? Because communication is batched, the splitting of multiple differing ranges can happen in parallel. So, the number of round-trips will not be affected (assuming that every message can be delivered in exactly one packet transmission, independent of size, which is of course not entirely true on real networks).
|
||||
The test is repeated using each language as both the client and the server.
|
||||
|
||||
The amount of bandwidth consumed will grow linearly with the number of differences, but this would of course be true even assuming a perfect synchronisation method that had no overhead other than transmitting the differing records.
|
||||
Afterwards, a different fuzz test is run for each language in isolation, and the exact protocol output is stored for each language. These are compared to ensure they are byte-wise identical.
|
||||
|
||||
Finally, a protocol upgrade test is run for each language to ensure that when run as a server it correctly indicates to the client when it cannot handle a specific protocol version.
|
||||
|
||||
* For the Rust implementation, check out its repo in the same directory as the `negentropy` repo, build the `harness` commands for both C++ and Rust, and then inside `negentropy/test/` directory running `perl test.pl cpp,rust`
|
||||
|
||||
|
||||
## Reference Implementation APIs
|
||||
|
||||
### C++
|
||||
## Applications
|
||||
|
||||
The library is contained in a single-header with the only dependency being OpenSSL (for SHA-256):
|
||||
|
||||
#include "Negentropy.h"
|
||||
|
||||
First, create a `Negentropy` object. The `16` argument is `idSize` and which is followed by an optional `frameSizeLimit`:
|
||||
|
||||
Negentropy ne(16, 50'000);
|
||||
|
||||
Next, add all the items in your collection, and `seal()`:
|
||||
|
||||
for (const auto &item : myItems) {
|
||||
ne.addItem(item.timestamp(), item.id());
|
||||
}
|
||||
|
||||
ne.seal();
|
||||
|
||||
On the client-side, create an initial message, and then transmit it to the server, receive the response, and `reconcile` until complete:
|
||||
|
||||
std::string msg = ne.initiate();
|
||||
|
||||
while (true) {
|
||||
std::string response = queryServer(msg);
|
||||
|
||||
std::vector<std::string> have, need;
|
||||
std::optional<std::string> newMsg = ne.reconcile(response, have, need);
|
||||
|
||||
// handle have/need
|
||||
|
||||
if (!newMsg) break; // done
|
||||
else std::swap(msg, *newMsg);
|
||||
}
|
||||
|
||||
In each loop iteration, `have` contains IDs that the client has that the server doesn't, and `need` contains IDs that the server has that the client doesn't.
|
||||
|
||||
The server-side is similar, except it doesn't create an initial message, there are no `have`/`need` arrays, and it doesn't return an optional (servers must always reply to a request):
|
||||
|
||||
while (true) {
|
||||
std::string msg = receiveMsgFromClient();
|
||||
std::string response = ne.reconcile(msg);
|
||||
respondToClient(response);
|
||||
}
|
||||
|
||||
### Javascript
|
||||
|
||||
The library is contained in a single javascript file. It shouldn't need any dependencies, in either a browser or node.js:
|
||||
|
||||
const Negentropy = require('Negentropy.js');
|
||||
|
||||
First, create a `Negentropy` object. The `16` argument is `idSize` and which is followed by an optional `frameSizeLimit`:
|
||||
|
||||
let ne = new Negentropy(16, 50000);
|
||||
|
||||
Next, add all the items in your collection, and `seal()`:
|
||||
|
||||
for (let item of myItems) {
|
||||
ne.addItem(item.timestamp(), item.id());
|
||||
}
|
||||
|
||||
ne.seal();
|
||||
|
||||
* `timestamp` should be a JS `Number`
|
||||
* `id` should be a hex string, `Uint8Array`, or node.js `Buffer`
|
||||
|
||||
On the client-side, create an initial message, and then transmit it to the server, receive the response, and `reconcile` until complete (signified by returning `null` for `newMsg`):
|
||||
|
||||
let msg = await ne.initiate();
|
||||
|
||||
while (msg.length !== null) {
|
||||
let response = queryServer(msg);
|
||||
let [newMsg, have, need] = await ne.reconcile(msg);
|
||||
msg = newMsg;
|
||||
// handle have/need
|
||||
}
|
||||
|
||||
* The output `msg`s and the IDs in `have`/`need` are hex strings, but you can set `ne.wantUint8ArrayOutput = true` if you want `Uint8Array`s instead.
|
||||
|
||||
The server-side is similar, except it doesn't create an initial message, there are no `have`/`need` arrays, and `newMsg` will never be `null`:
|
||||
|
||||
while (1) {
|
||||
let msg = receiveMsgFromClient();
|
||||
let [newMsg] = await ne.reconcile(msg);
|
||||
respondToClient(newMsg);
|
||||
}
|
||||
|
||||
* The `initiate()` and `reconcile()` methods are async because the `crypto.subtle.digest()` browser API is async.
|
||||
* Timestamp values greater than `Number.MAX_VALUE` will currently cause failures.
|
||||
|
||||
### Rust
|
||||
|
||||
A third-party Rust implementation by Yuki Kishimoto is available in the [rust-negentropy](https://github.com/yukibtc/rust-negentropy) repository.
|
||||
|
||||
The reference test suite can be run against this implementation by checking it out in the same directory as the `negentropy` repo, building the `harness` commands for both C++ and Rust, and then inside `negentropy/test/` directory running `perl test.pl cpp,rust`
|
||||
|
||||
|
||||
|
||||
## Adversarial Input
|
||||
|
||||
Negentropy is intended to be secure against a user who can insert arbitrary records into the DBs of one or both parties of the protocol. Secure in this sense means that the reconciliation will succeed without missing any records. A previous version of negentropy used xor to combine hashes, meaning it was only suitable for trusted input, or application-level countermeasures should have been taken (as described in Meyer's master's thesis).
|
||||
|
||||
Note that the threat model does not include one of the sides of the parties in the protocol themselves being malicious (except that implementations should obviously be secure against implementation defects such as buffer overflows, etc). In this case, a malicious party could always pretend to have or not have records in the first place, meaning that corrupting the protocol doesn't gain them anything additional, except possibly wasting the other side's CPU/bandwidth/memory resources.
|
||||
|
||||
|
||||
## Pre-computing Fingerprints
|
||||
|
||||
The reference implementations work by storing the IDs (sorted by timestamp and then id) in a contiguous memory buffer. This means that a single call to the SHA-256 hash function is sufficient for computing the fingerprint of any sub-range. The result is that fingerprints can be computed rapidly, especially when using dedicated CPU instructions. On my laptop, computing a fingerprint for 1 million records (with ID size 16) takes approximately 1 millisecond.
|
||||
|
||||
However, for extremely large DBs, and also in cases where a significant amount of IO would be required to load all the IDs into a contiguous memory buffer, it would be helpful to be able to pre-compute fingerprints. As Meyer points out, storing a fingerprint for every possible sub-range would be infeasible, and SHA-256 cannot be computed incrementally, which reduces our options.
|
||||
|
||||
Fortunately, because of the flexibility of the range-based approach, we can still pre-compute fingerprints for various common ranges, and then direct the protocol to prefer using those ranges. For example, suppose we wanted to send a large range that we haven't pre-computed, but we have several pre-computed sub-ranges within this range. The protocol functions perfectly well if we instead sent multiple adjacent range fingerprints instead of re-hashing the entire range, at the expense of some extra bandwidth usage.
|
||||
|
||||
For example, suppose we pre-computed the fingerprints for all records within each day. In this case, when performing a reconciliation, these daily ranges would always be preferred. Depending on how many records are stored each day, perhaps hourly fingerprints would be pre-computed in addition. When reconciling differing hours, it would fall back to loading the IDs into contiguous memory regions.
|
||||
|
||||
This approach will work best if all participants in the system agree on globally-consistent daily/hourly boundaries. Because most often records that are being inserted or modified are for the current day, it may not be beneficial to pre-compute the current day's fingerprint as it would be changing too frequently. But a corollary to this is that re-computing the fingerprints of old/archival ranges would be infrequent.
|
||||
|
||||
|
||||
## Use-Cases
|
||||
This section lists the currently-known production applications that use negentropy. If you know of a new one, please let us know by [opening an issue](https://github.com/hoytech/negentropy/issues/new).
|
||||
|
||||
* [Bandwidth-efficient Nostr event syncing](https://github.com/hoytech/strfry/blob/next/docs/negentropy.md)
|
||||
|
||||
|
||||
## Copyright
|
||||
|
||||
## Author
|
||||
|
||||
(C) 2023 Doug Hoyte
|
||||
|
||||
Protocol specification, reference implementations, and tests are MIT licensed.
|
||||
|
||||
Negentropy is a [Log Periodic](https://logperiodic.com) project.
|
||||
|
||||
See [our article](https://logperiodic.com/rbsr.html) for more information.
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user