[bitcoin-dev] Compressed block headers

[Will Clark <will8clark@gmail.com>]

Hello list,

I would like to propose a compressed block header scheme for IBD and block announcements. This proposal is derivative of previous proposals found on this list (see links in spec below) with some modifications and clarifications.

The below specification (also found at https://github.com/willcl-ark/compressed-block-headers/blob/v1.0/compressed-block-headers.adoc ) details the compression recommended along with the generated bandwidth savings in the best-case scenario.

I look forward to any feedback anyone has to offer on the specification itself, as well as any additions or objections to the motivation.

Cheers,
Will


= Compressed block headers
Will Clark <will8clark@gmail.com>
v1.0, May 2020:
:toc: preamble
:toclevels: 4


This work is a derivation of these mailing list posts:

1. https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2017-August/014876.html[bitcoin-dev: "Compressed" headers stream - 2017] (with resurrection https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2017-December/015385.html[here])

2. https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2018-March/015851.html[bitcoin-dev: Optimized Header Sync]

'''

== Motivation

Block headers as exchanged by nodes over the p2p network are currently 81 bytes each.

For low bandwidth nodes who are doing a headers-only sync, reducing the size of the headers can provide a significant bandwidth saving. Also, nodes can support more header-only peers for IBD and protection against eclipse attacks if header bandwidth is reduced.

=== Background

Currently headers are sent over the p2p network as a vector of `block_headers`, which are composed of the following sized fields:

[cols="<,>"]
|===
|Field               |Size

|Version             |4 bytes
|Previous block hash |32 bytes
|Merkle root hash    |32 bytes
|Time                |4 bytes
|nBits               |4 bytes
|nonce               |4 bytes
|txn_count           |1 byte
|*Total*             |81 bytes
|===

Some fields can be removed completely, others can be compressed under certain conditions.

== Proposed specification

=== block_header2 data type

The following table illustrates the proposed `block_header2` data type specification.

[cols="<,>,>"]
|===
|Field               |Size     |Compressed

|Bitfield            |1 byte   | 1 byte
|Version             |4 bytes  |0 \| 4 bytes
|Previous block hash |32 bytes |0 \| 32 bytes
|Merkle root hash    |32 bytes |32 bytes
|Time                |4 bytes  |2 \| 4 bytes
|nBits               |4 bytes  |0 \| 4 bytes
|nonce               |4 bytes  |4 bytes
|*Total*             |81 bytes |range: 39 - 81 bytes
|===

This compression results in a maximum reduction from an 81 byte header to best-case 39 byte header. With 629,474 blocks in the current blockchain, a continuous header sync from genesis (requiring a single full 81 byte header followed by only compressed `block_header2`) has been tested to have its required bandwidth reduced from 50.98MB down to 25.86MB, a saving of 49%.

==== Bitfield

To make parsing of header messages easier and further increase header compression, a single byte bitfield was suggested by gmaxwell footnote:[https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2017-December/015397.html]. We propose the following amended bitfield meanings (bits re-ordered to match `headers2` field order):

[cols="<,<"]
|===
|Bit |Meaning + field size to read

|0 +
1 +
2    |version: same as the last *distinct* value 1st ... 7th (0 byte field) or a new 32bit distinct value (4 byte field).
|3   |prev_block_hash: is omitted (0 byte field) or included (32 byte field)
|4   |timestamp: as small offset (2 byte field) or full (4 byte field).
|5   |nbits: same as last header (0 byte field) or new (4 byte field).
|6   |possibly to signal "more headers follow" to make the encoding self-delimiting.
|7   |currently undefined
|===

This bitfield adds 1 byte for every block in the chain, for a current total increase of 629,474B.

==== Version

In most cases the Version field will be identical to one referenced in one of the previous 7 unique versions, as indicated by bits 0,1,2 of the Bitfield.

To block 629,474 there were 616,137 blocks whose version was in the previous 7 distinct versions, and only 13,338 blocks whose version was not, this includes any version bit manipulation done via overt ASIC boost.

[cols=">,>,>,>"]
|===
|Genesis to block |Current (B) |Compressed (B) |Saving (%)

|629,474          |2,517,896   |53,352         |98
|===

==== Previous block hash

The previous block hash will always be the
`SHA256(SHA256(<previous_header>))` so is redundant, presuming you have the previous header in the chain.

[cols=">,>,>,>"]
|===
|Genesis to block |Current (B) |Compressed (B) |Saving (%)

|629,474          |20,143,168  |0              |100
|===

==== Time

The timestamp (in seconds) is consensus bound, based both on the time in the previous
header: `MAX_FUTURE_BLOCK_TIME = 2 * 60 * 60 = 7200`, and being greater than the `MedianTimePast` of the previous 11 blocks. Therefore this can be safely represented as an offset from the previous headers' timestamp using a 2 byte `signed short int`.

[cols=">,>,>,>"]
|===
|Genesis to block |Current (B) |Compressed (B) |Saving (%)

|629,474          |2,517,896   |1,258,952      |50
|===

==== nBits

nBits currently changes once every 2016 blocks. It could be entirely calculated by the client from the timestamps of the previous 2015 blocks footnote:[2015 blocks are used in the adjustment calculation due to an off-by-one error: https://bitcointalk.org/index.php?topic=43692.msg521772#msg521772"].

To simplify 'light' client implementations which would otherwise require consensus-valid calculation of the adjustments, we propose to transmit this according to the <<Bitfield>> specification above.

To block 629,474 there have been 298 nBits adjustments (vs an expected 311 -- there was none before block 32,256).

[cols=">,>,>,>"]
|===
|Genesis to block |Current (B) |Compressed (B) |Saving (%)

|629,474          |2,517,896   |1,196          |99.6
|===

==== txn_count

txn_count is included to make parsing of these messages compatible with parsing of `block` messages footnote:[https://bitcoin.stackexchange.com/questions/2104/why-is-the-block-header-txn-count-field-always-zero]. Therefore this field and its associated byte can be removed for transmission of compact headers.

[cols=">,>,>,>"]
|===
|Genesis to block |Current (B) |Compressed (B) |Saving (%)

|629,474          |629,474     |0              |100
|===

=== Service Bit

A new service bit would be required so that the nodes can advertise their ability to supply compact headers.

=== P2P Messages

Three new messages would be used by nodes that enable compact block header support, two query messages: `getheaders2` and `sendheaders2` and one response: `headers2`.

==== `getheaders2` -- Requesting compact headers

The new p2p message required to request compact block headers would require the same fields as the current `getheaders` message:

[cols=">,<,<,<"]
|===
|Field Size |Description          |Data type |Comments

|4          |version              |uint32_t  |the protocol version
|1+         |hash count           |var_int   |number of block locator hash entries
|32+        |block locator hashes |char[32]  |block locator object; newest back to genesis block (dense to start, but then sparse)
|32         |hash_stop            |char[32]  |hash of the last desired block header; set to zero to get as many blocks as possible (2000)
|===

==== `sendheaders2` -- Request compact header announcements

Since https://github.com/bitcoin/bips/blob/master/bip-0130.mediawiki[BIP-130], nodes have been able to request to receive new headers directly in `headers` messages, rather than via an `inv` of the new block hash and subsequent `getheader` request and `headers` response (followed by a final `getdata` to get the tip block itself, if desired). This is requested by transmitting an empty `sendheaders` message after the version handshake is complete.]

Upon receipt of this message, the node is permitted, but not required, to preemptively announce new headers with the `headers2` message (instead of `inv`). Preemptive header announcement is supported by the protocol version ≥ 70012 | Bitcoin Core version ≥ 0.12.0.

For the motivational use-case it makes sense to also update this mechanism to support sending header updates using compact headers using a new message.

==== `headers2` -- Receiving compact headers

A `headers2` message is returned in response to `getheaders2` or at new header announcement following a `sendheaders2` request. It contains both `length` and `headers` fields. The `headers` field contains a variable length vector of `block_header2`:

|===
|Field Size |Description |Data type       |Comments

|1+         |length      |var_int         |Length of `headers`
|39-81x?    |headers     |block_header2[] |Compressed block headers in <<block_header2 data type>> format
|===

=== Implementation

* The first header in the first `block_header2[]` vector to a newly-connected client MUST contain the full nBits`, `timestamp`, `version` and `prev_block_hash` fields, along with a correctly populated `bitfield` byte.
* Subsequent headers in a contiguous vector SHOULD follow the compressed <<block_header2 data type>> format.
* Subsequent compressed headers supplied to an already-connected client (requesting compressed headers), SHOULD follow the compressed <<block_header2 data type>> format.