Skip to content

Latest commit

 

History

History

lightning-rapid-gossip-sync

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

lightning-rapid-gossip-sync

This crate exposes functionality for rapid gossip graph syncing, aimed primarily at mobile clients. Its server counterpart is the rapid-gossip-sync-server repository.

Mechanism

The (presumed) server sends a compressed gossip response containing gossip data. The gossip data is formatted compactly, omitting signatures and opportunistically incremental where previous channel updates are known.

Essentially, the serialization structure is as follows:

  1. Fixed prefix bytes 76, 68, 75, 1 (the first three bytes are ASCII for LDK)
    • The purpose of this prefix is to identify the serialization format, should other rapid gossip sync formats arise in the future
    • The fourth byte is the protocol version in case our format gets updated
  2. Chain hash (32 bytes)
  3. Latest seen timestamp (u32)
  4. An unsigned int indicating the number of node IDs to follow
  5. An array of compressed node ID pubkeys (all pubkeys are presumed to be standard compressed 33-byte-serializations)
  6. An unsigned int indicating the number of channel announcement messages to follow
  7. An array of significantly stripped down customized channel announcements
  8. An unsigned int indicating the number of channel update messages to follow
  9. A series of default values used for non-incremental channel updates
    • The values are defined as follows:
      1. default_cltv_expiry_delta
      2. default_htlc_minimum_msat
      3. default_fee_base_msat
      4. default_fee_proportional_millionths
      5. default_htlc_maximum_msat (u64, and if the default is no maximum, u64::MAX)
    • The defaults are calculated by the server based on the frequency among non-incremental updates within a given delta set
  10. An array of customized channel updates

You will also notice that NodeAnnouncement messages are omitted altogether as the node IDs are implicitly extracted from the channel announcements and updates.

The data is then applied to the current network graph, artificially dated to the timestamp of the latest seen message less one week, be it an announcement or an update, from the server's perspective. The network graph should not be pruned until the graph sync completes.

Custom Channel Announcement

To achieve compactness and avoid data repetition, we're sending a significantly stripped down version of the channel announcement message, which contains only the following data:

  1. channel_features: u16 + n, where n is the number of bytes indicated by the first u16
  2. short_channel_id: CompactSize (incremental CompactSize deltas starting from 0)
  3. node_id_1_index: CompactSize (index of node id within the previously sent sequence)
  4. node_id_2_index: CompactSize (index of node id within the previously sent sequence)

Custom Channel Update

For the purpose of rapid syncing, we have deviated from the channel update format specified in BOLT 7 significantly. Our custom channel updates are structured as follows:

  1. short_channel_id: CompactSize (incremental CompactSize deltas starting at 0)
  2. custom_channel_flags: u8
  3. update_data

Specifically, our custom channel flags break down like this:

128 64 32 16 8 4 2 1
Incremental update? Disable channel? Direction

If the most significant bit is set to 1, indicating an incremental update, the intermediate bit flags assume the following meaning:

64 32 16 8 4
cltv_expiry_delta has changed htlc_minimum_msat has changed fee_base_msat has changed fee_proportional_millionths has changed htlc_maximum_msat has changed

If the most significant bit is set to 0, the meaning is almost identical, except instead of a change, the flags now represent a deviation from the defaults sent at the beginning of the update sequence.

In both cases, update_data only contains the fields that are indicated by the channel flags to be non-default or to have mutated.

Delta Calculation

The way a server is meant to calculate this rapid gossip sync data is by taking the latest time any change, be it either an announcement or an update, was seen. That timestamp is included in each rapid sync message, so all the client needs to do is cache one variable.

If a particular channel update had never occurred before, the full update is sent. If a channel has had updates prior to the provided timestamp, the latest update prior to the timestamp is taken as a reference, and the delta is calculated against it.

Depending on whether the rapid sync message is calculated on the fly or a snapshotted version is returned, intermediate changes between the latest update seen by the client and the latest update broadcast on the network may be taken into account when calculating the delta.

Performance

Given the primary purpose of this utility is a faster graph sync, we thought it might be helpful to provide some examples of various delta sets. These examples were calculated as of May 19th 2022 with a network graph comprised of 80,000 channel announcements and 160,000 directed channel updates.

Full sync
Message Length 4.7 MB
Gzipped Message Length 2.0 MB
Client-side Processing Time 1.4 s
Week-old sync
Message Length 2.7 MB
Gzipped Message Length 862 kB
Client-side Processing Time 907 ms
Day-old sync
Message Length 191 kB
Gzipped Message Length 92.8 kB
Client-side Processing Time 196 ms