Milestone 8 (phases 1-4): - Shard-aware WAL segment naming, BatchHeader v2, ShardRouter - Transport trait, InProcessTransport, WalShipper, FollowerDb - HLC, PNCounter, LWWRegister, CrdtSignalState, ReconciliationEngine - Session replication bridge with SeqNo/HWM, idempotency store Forage application: - Multi-source discovery engine with MAB exploration - Embedding-based label system, server handlers, UI refresh Other: - QUICKSTART.md, README.md, milestone-8 planning docs - Hard negative union semantics, RLHF export enhancements - Recovery benchmark and visibility test expansions - Split 8 oversized source files per CODING_GUIDELINES §9 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
374 lines
12 KiB
Rust
374 lines
12 KiB
Rust
use super::super::error::WalError;
|
|
use crate::replication::{RegionId, ShardId};
|
|
|
|
// ── Signal batch format ─────────────────────────────────────────────────────
|
|
|
|
/// Magic bytes identifying a tidalDB WAL batch frame: "TIDL" in LE byte order.
|
|
///
|
|
/// Stored as `[0x44, 0x4C, 0x49, 0x54]` which is `0x54494C44` as a u32 LE.
|
|
/// This allows `u32::from_le_bytes(magic) == 0x54494C44` to validate.
|
|
pub const MAGIC: [u8; 4] = [0x44, 0x4C, 0x49, 0x54];
|
|
|
|
/// Wire format version 1: original single-node format. Bytes 28-31 are
|
|
/// zero-padded reserved space.
|
|
pub const FORMAT_VERSION_V1: u8 = 1;
|
|
|
|
/// Wire format version 2: adds `shard_id` and `region_id`.
|
|
///
|
|
/// Bytes 28-29 carry `shard_id` (u16 LE) and bytes 30-31 carry `region_id`
|
|
/// (u16 LE) for multi-node replication. Backward compatible with v1
|
|
/// because v1 always wrote zeros at bytes 28-31, which decode as
|
|
/// `ShardId::SINGLE` and `RegionId::SINGLE`.
|
|
pub const FORMAT_VERSION_V2: u8 = 2;
|
|
|
|
/// Current wire format version. All new batches are encoded as v2.
|
|
pub const FORMAT_VERSION: u8 = FORMAT_VERSION_V2;
|
|
|
|
/// Record type discriminant for signal events.
|
|
pub const RECORD_TYPE_SIGNAL: u8 = 0x01;
|
|
|
|
/// Size of the batch header in bytes (one cache line).
|
|
pub const HEADER_SIZE: usize = 64;
|
|
|
|
/// Size of a single event record in bytes.
|
|
pub const EVENT_SIZE: usize = 21;
|
|
|
|
/// Maximum number of events in a single batch.
|
|
pub const MAX_EVENTS_PER_BATCH: u16 = 256;
|
|
|
|
/// Decoded batch header.
|
|
///
|
|
/// # Wire layout (64 bytes, one cache line)
|
|
///
|
|
/// ```text
|
|
/// Bytes 0- 3: MAGIC [0x44, 0x4C, 0x49, 0x54] ("TIDL")
|
|
/// Byte 4: format version (u8)
|
|
/// Byte 5: flags (u8, reserved)
|
|
/// Bytes 6- 7: event_count (u16 LE)
|
|
/// Bytes 8-15: first_seq (u64 LE)
|
|
/// Bytes 16-23: batch_timestamp (u64 LE)
|
|
/// Bytes 24-27: payload_len (u32 LE)
|
|
/// Bytes 28-29: shard_id (u16 LE) -- v2; zero in v1 and single-node
|
|
/// Bytes 30-31: region_id (u16 LE) -- v2; zero in v1 and single-node
|
|
/// Bytes 32-63: BLAKE3 checksum (32 bytes)
|
|
/// ```
|
|
#[derive(Debug, Clone, PartialEq, Eq)]
|
|
pub struct BatchHeader {
|
|
pub version: u8,
|
|
pub flags: u8,
|
|
pub event_count: u16,
|
|
pub first_seq: u64,
|
|
pub batch_timestamp: u64,
|
|
pub payload_len: u32,
|
|
pub checksum: [u8; 32],
|
|
/// Shard that produced this batch. `ShardId::SINGLE` (0) in single-node
|
|
/// deployments and when reading v1 batches.
|
|
pub shard_id: ShardId,
|
|
/// Region that produced this batch. `RegionId::SINGLE` (0) in single-node
|
|
/// deployments and when reading v1 batches.
|
|
pub region_id: RegionId,
|
|
}
|
|
|
|
/// A single signal event record in wire format.
|
|
#[derive(Debug, Clone, PartialEq)]
|
|
pub struct EventRecord {
|
|
pub entity_id: u64,
|
|
pub signal_type: u8,
|
|
pub weight: f32,
|
|
pub timestamp_nanos: u64,
|
|
}
|
|
|
|
impl EventRecord {
|
|
/// Serialize this event into the 21-byte wire format.
|
|
#[must_use]
|
|
pub fn to_bytes(&self) -> [u8; EVENT_SIZE] {
|
|
let mut buf = [0u8; EVENT_SIZE];
|
|
buf[0..8].copy_from_slice(&self.entity_id.to_le_bytes());
|
|
buf[8] = self.signal_type;
|
|
buf[9..13].copy_from_slice(&self.weight.to_le_bytes());
|
|
buf[13..21].copy_from_slice(&self.timestamp_nanos.to_le_bytes());
|
|
buf
|
|
}
|
|
|
|
/// Deserialize an event from 21 bytes of wire format.
|
|
///
|
|
/// # Errors
|
|
///
|
|
/// Returns `WalError::Corruption` if the slice is not exactly 21 bytes.
|
|
pub fn from_bytes(bytes: &[u8]) -> Result<Self, WalError> {
|
|
if bytes.len() != EVENT_SIZE {
|
|
return Err(WalError::Corruption {
|
|
message: format!(
|
|
"event record: expected {EVENT_SIZE} bytes, got {}",
|
|
bytes.len()
|
|
),
|
|
});
|
|
}
|
|
let entity_id =
|
|
u64::from_le_bytes(bytes[0..8].try_into().map_err(|_| WalError::Corruption {
|
|
message: "event record: invalid entity_id bytes".into(),
|
|
})?);
|
|
let signal_type = bytes[8];
|
|
let weight =
|
|
f32::from_le_bytes(bytes[9..13].try_into().map_err(|_| WalError::Corruption {
|
|
message: "event record: invalid weight bytes".into(),
|
|
})?);
|
|
let timestamp_nanos =
|
|
u64::from_le_bytes(bytes[13..21].try_into().map_err(|_| WalError::Corruption {
|
|
message: "event record: invalid timestamp bytes".into(),
|
|
})?);
|
|
Ok(Self {
|
|
entity_id,
|
|
signal_type,
|
|
weight,
|
|
timestamp_nanos,
|
|
})
|
|
}
|
|
}
|
|
|
|
/// Encode a batch of events into the WAL wire format (single-node convenience).
|
|
///
|
|
/// Equivalent to `encode_batch_with_shard(events, first_seq, batch_ts,
|
|
/// ShardId::SINGLE, RegionId::SINGLE)`.
|
|
///
|
|
/// # Errors
|
|
///
|
|
/// Returns `WalError::Corruption` if `events` is empty or exceeds
|
|
/// `MAX_EVENTS_PER_BATCH`.
|
|
pub fn encode_batch(
|
|
events: &[EventRecord],
|
|
first_seq: u64,
|
|
batch_ts: u64,
|
|
) -> Result<Vec<u8>, WalError> {
|
|
encode_batch_with_shard(
|
|
events,
|
|
first_seq,
|
|
batch_ts,
|
|
ShardId::SINGLE,
|
|
RegionId::SINGLE,
|
|
)
|
|
}
|
|
|
|
/// Encode a batch of events with explicit shard and region identity.
|
|
///
|
|
/// Produces a byte vector containing the 64-byte header followed by
|
|
/// tightly packed 21-byte event records. The BLAKE3 checksum covers
|
|
/// `header[0..32] || event_bytes`, which includes the shard/region bytes.
|
|
///
|
|
/// # Errors
|
|
///
|
|
/// Returns `WalError::Corruption` if `events` is empty or exceeds
|
|
/// `MAX_EVENTS_PER_BATCH`.
|
|
pub fn encode_batch_with_shard(
|
|
events: &[EventRecord],
|
|
first_seq: u64,
|
|
batch_ts: u64,
|
|
shard_id: ShardId,
|
|
region_id: RegionId,
|
|
) -> Result<Vec<u8>, WalError> {
|
|
let event_count = events.len();
|
|
if event_count == 0 || event_count > usize::from(MAX_EVENTS_PER_BATCH) {
|
|
return Err(WalError::Corruption {
|
|
message: format!(
|
|
"batch event count {event_count} out of range [1, {MAX_EVENTS_PER_BATCH}]"
|
|
),
|
|
});
|
|
}
|
|
|
|
let payload_len = event_count * EVENT_SIZE;
|
|
let total_len = HEADER_SIZE + payload_len;
|
|
let mut buf = vec![0u8; total_len];
|
|
|
|
// Write header fields [0..32]
|
|
buf[0..4].copy_from_slice(&MAGIC);
|
|
buf[4] = FORMAT_VERSION;
|
|
buf[5] = 0; // flags: reserved
|
|
#[allow(clippy::cast_possible_truncation)]
|
|
let count_u16 = event_count as u16;
|
|
buf[6..8].copy_from_slice(&count_u16.to_le_bytes());
|
|
buf[8..16].copy_from_slice(&first_seq.to_le_bytes());
|
|
buf[16..24].copy_from_slice(&batch_ts.to_le_bytes());
|
|
#[allow(clippy::cast_possible_truncation)]
|
|
let payload_len_u32 = payload_len as u32;
|
|
buf[24..28].copy_from_slice(&payload_len_u32.to_le_bytes());
|
|
// v2: shard_id and region_id at bytes 28-31.
|
|
buf[28..30].copy_from_slice(&shard_id.0.to_le_bytes());
|
|
buf[30..32].copy_from_slice(®ion_id.0.to_le_bytes());
|
|
|
|
// Write event records starting at offset 64
|
|
for (i, event) in events.iter().enumerate() {
|
|
let offset = HEADER_SIZE + i * EVENT_SIZE;
|
|
buf[offset..offset + EVENT_SIZE].copy_from_slice(&event.to_bytes());
|
|
}
|
|
|
|
// Compute BLAKE3 over header[0..32] || event_bytes
|
|
let checksum = compute_checksum(&buf[0..32], &buf[HEADER_SIZE..]);
|
|
buf[32..64].copy_from_slice(checksum.as_bytes());
|
|
|
|
Ok(buf)
|
|
}
|
|
|
|
/// Decode a batch from raw bytes.
|
|
///
|
|
/// Two-phase validation:
|
|
/// - Phase 1: magic bytes, version, payload length bounds
|
|
/// - Phase 2: BLAKE3 checksum verification
|
|
///
|
|
/// # Errors
|
|
///
|
|
/// Returns `WalError::Corruption` on any validation failure.
|
|
pub fn decode_batch(bytes: &[u8]) -> Result<(BatchHeader, Vec<EventRecord>), WalError> {
|
|
if bytes.len() < HEADER_SIZE {
|
|
return Err(WalError::Corruption {
|
|
message: format!(
|
|
"batch too short for header: {} bytes, need {HEADER_SIZE}",
|
|
bytes.len()
|
|
),
|
|
});
|
|
}
|
|
|
|
// Phase 1: structural validation
|
|
if bytes[0..4] != MAGIC {
|
|
return Err(WalError::Corruption {
|
|
message: "invalid magic bytes".into(),
|
|
});
|
|
}
|
|
|
|
let version = bytes[4];
|
|
if version != FORMAT_VERSION_V1 && version != FORMAT_VERSION_V2 {
|
|
return Err(WalError::Corruption {
|
|
message: format!("unsupported format version: {version}"),
|
|
});
|
|
}
|
|
|
|
let flags = bytes[5];
|
|
let event_count =
|
|
u16::from_le_bytes(bytes[6..8].try_into().map_err(|_| WalError::Corruption {
|
|
message: "invalid event_count bytes".into(),
|
|
})?);
|
|
|
|
if event_count == 0 || event_count > MAX_EVENTS_PER_BATCH {
|
|
return Err(WalError::Corruption {
|
|
message: format!("event count {event_count} out of range [1, {MAX_EVENTS_PER_BATCH}]"),
|
|
});
|
|
}
|
|
|
|
let first_seq =
|
|
u64::from_le_bytes(bytes[8..16].try_into().map_err(|_| WalError::Corruption {
|
|
message: "invalid first_seq bytes".into(),
|
|
})?);
|
|
|
|
let batch_timestamp =
|
|
u64::from_le_bytes(bytes[16..24].try_into().map_err(|_| WalError::Corruption {
|
|
message: "invalid batch_timestamp bytes".into(),
|
|
})?);
|
|
|
|
let payload_len =
|
|
u32::from_le_bytes(bytes[24..28].try_into().map_err(|_| WalError::Corruption {
|
|
message: "invalid payload_len bytes".into(),
|
|
})?);
|
|
|
|
// v2 fields at bytes 28-31. For v1 batches these bytes are always zero,
|
|
// which correctly decodes as ShardId::SINGLE and RegionId::SINGLE.
|
|
let shard_id = ShardId(u16::from_le_bytes(bytes[28..30].try_into().map_err(
|
|
|_| WalError::Corruption {
|
|
message: "invalid shard_id bytes".into(),
|
|
},
|
|
)?));
|
|
let region_id = RegionId(u16::from_le_bytes(bytes[30..32].try_into().map_err(
|
|
|_| WalError::Corruption {
|
|
message: "invalid region_id bytes".into(),
|
|
},
|
|
)?));
|
|
|
|
let expected_payload = u32::from(event_count) * EVENT_SIZE as u32;
|
|
if payload_len != expected_payload {
|
|
return Err(WalError::Corruption {
|
|
message: format!(
|
|
"payload_len {payload_len} != event_count {event_count} * {EVENT_SIZE}"
|
|
),
|
|
});
|
|
}
|
|
|
|
let total_len = HEADER_SIZE + payload_len as usize;
|
|
if bytes.len() < total_len {
|
|
return Err(WalError::Corruption {
|
|
message: format!(
|
|
"batch truncated: have {} bytes, need {total_len}",
|
|
bytes.len()
|
|
),
|
|
});
|
|
}
|
|
|
|
// Extract stored checksum
|
|
let mut checksum = [0u8; 32];
|
|
checksum.copy_from_slice(&bytes[32..64]);
|
|
|
|
// Phase 2: BLAKE3 verification
|
|
let event_bytes = &bytes[HEADER_SIZE..total_len];
|
|
let computed = compute_checksum(&bytes[0..32], event_bytes);
|
|
if computed.as_bytes() != &checksum {
|
|
return Err(WalError::Corruption {
|
|
message: "BLAKE3 checksum mismatch".into(),
|
|
});
|
|
}
|
|
|
|
// Parse event records
|
|
let mut events = Vec::with_capacity(usize::from(event_count));
|
|
for i in 0..usize::from(event_count) {
|
|
let offset = i * EVENT_SIZE;
|
|
let event = EventRecord::from_bytes(&event_bytes[offset..offset + EVENT_SIZE])?;
|
|
events.push(event);
|
|
}
|
|
|
|
let header = BatchHeader {
|
|
version,
|
|
flags,
|
|
event_count,
|
|
first_seq,
|
|
batch_timestamp,
|
|
payload_len,
|
|
checksum,
|
|
shard_id,
|
|
region_id,
|
|
};
|
|
|
|
Ok((header, events))
|
|
}
|
|
|
|
/// Compute the BLAKE3 checksum for a batch.
|
|
///
|
|
/// Input: `header_prefix[0..32] || event_bytes`.
|
|
/// The hash field at `[32..64]` is NOT part of the hash input.
|
|
fn compute_checksum(header_prefix: &[u8], event_bytes: &[u8]) -> blake3::Hash {
|
|
let mut hasher = blake3::Hasher::new();
|
|
hasher.update(header_prefix);
|
|
hasher.update(event_bytes);
|
|
hasher.finalize()
|
|
}
|
|
|
|
/// Compute the per-event content hash used for deduplication.
|
|
///
|
|
/// Returns the first 128 bits of the BLAKE3 hash of the 21-byte event record.
|
|
///
|
|
/// # Panics
|
|
///
|
|
/// Cannot panic. The `expect` is on a `try_into` converting a 16-byte slice
|
|
/// (from a 32-byte BLAKE3 hash) into `[u8; 16]`, which is infallible.
|
|
#[must_use]
|
|
pub fn event_content_hash(event: &EventRecord) -> u128 {
|
|
let bytes = event.to_bytes();
|
|
let hash = blake3::hash(&bytes);
|
|
let hash_bytes: &[u8; 32] = hash.as_bytes();
|
|
u128::from_le_bytes(
|
|
hash_bytes[..16]
|
|
.try_into()
|
|
.expect("BLAKE3 hash is always 32 bytes; first 16 is infallible"),
|
|
)
|
|
}
|
|
|
|
#[cfg(test)]
|
|
#[allow(clippy::unwrap_used)]
|
|
#[path = "batch_tests.rs"]
|
|
mod tests;
|