Phase 3a: Lamport-stamped refs with CRDT-merge on PutRef #40

Merged
osobh merged 1 commits from phase-3a-stamped-refs into main 2026-07-13 14:34:33 +00:00
4 changed files with 477 additions and 1 deletions
+267
View File
@@ -31,6 +31,85 @@ pub type RefKey = [u8; 32];
/// A raw 32-byte value. /// A raw 32-byte value.
pub type RefValue = [u8; 32]; pub type RefValue = [u8; 32];
/// Phase 3 (2026-07-13): 8-byte node identifier used as a tie-breaker
/// in the CRDT merge order. Convention: `blake3(node_name)[0..8]`.
/// Two nodes will only collide on this if they share the same 8-byte
/// prefix of their names' hashes (birthday-bound at ~4B).
pub type NodeStamp = [u8; 8];
/// A ref value paired with its Lamport clock and originator node
/// stamp — the tuple that Phase 3 CRDT semantics needs to merge
/// concurrent writes deterministically.
///
/// The pair `(clock, node)` forms a total order — `dominates`
/// returns true when `self` is strictly newer than `other`. Equal
/// pairs are treated as "same write, idempotent" (no-op).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct StampedRef {
pub value: RefValue,
pub clock: u64,
pub node: NodeStamp,
}
impl StampedRef {
/// 48-byte on-disk / on-wire representation: `value:32 || clock:u64 LE || node:8`.
pub const ENCODED_LEN: usize = 48;
pub fn to_bytes(&self) -> [u8; Self::ENCODED_LEN] {
let mut out = [0u8; Self::ENCODED_LEN];
out[..32].copy_from_slice(&self.value);
out[32..40].copy_from_slice(&self.clock.to_le_bytes());
out[40..48].copy_from_slice(&self.node);
out
}
pub fn from_bytes(bytes: &[u8]) -> Result<Self> {
if bytes.len() != Self::ENCODED_LEN {
bail!(
"stamped ref wrong length {} (expected {})",
bytes.len(),
Self::ENCODED_LEN
);
}
let mut value = [0u8; 32];
value.copy_from_slice(&bytes[..32]);
let mut clock_bytes = [0u8; 8];
clock_bytes.copy_from_slice(&bytes[32..40]);
let mut node = [0u8; 8];
node.copy_from_slice(&bytes[40..48]);
Ok(Self {
value,
clock: u64::from_le_bytes(clock_bytes),
node,
})
}
/// Strict CRDT merge order: `self` beats `other` iff the pair
/// `(clock, node)` is strictly greater.
pub fn dominates(&self, other: &Self) -> bool {
(self.clock, self.node) > (other.clock, other.node)
}
}
/// Derive a stable 8-byte node stamp from a human node name.
pub fn node_stamp_for(name: &str) -> NodeStamp {
let hash = blake3::hash(name.as_bytes());
let mut out = [0u8; 8];
out.copy_from_slice(&hash.as_bytes()[..8]);
out
}
/// Outcome of a stamped put. Callers can distinguish "we accepted
/// your write" from "someone else already had a newer/equal write".
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum PutOutcome {
/// Incoming write was strictly newer; on-disk value updated.
Merged,
/// A prior write with `(clock, node) >= incoming` already exists.
/// The on-disk value is unchanged; `current` is what's there.
Rejected { current: StampedRef },
}
/// Directory-backed reference store. /// Directory-backed reference store.
#[derive(Debug, Clone)] #[derive(Debug, Clone)]
pub struct RefStore { pub struct RefStore {
@@ -117,6 +196,61 @@ impl RefStore {
.join(format!("{hex}.ref")) .join(format!("{hex}.ref"))
} }
/// Phase 3 (2026-07-13): Lamport-stamped ref lookup. Reads from
/// the `refs-v2/` directory (separate namespace from the raw
/// `refs/` set) so the two lookup surfaces don't interfere.
/// Returns `None` when no stamped ref exists.
pub async fn get_stamped(&self, key: &RefKey) -> Result<Option<StampedRef>> {
let path = self.stamped_path(key);
match tokio::fs::read(&path).await {
Ok(bytes) => Ok(Some(StampedRef::from_bytes(&bytes).with_context(|| {
format!("decoding stamped ref at {}", path.display())
})?)),
Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(None),
Err(e) => Err(anyhow::Error::from(e))
.with_context(|| format!("reading stamped ref at {}", path.display())),
}
}
/// Phase 3 (2026-07-13): merge an incoming stamped ref against
/// the local view.
///
/// Merge rule: the pair `(clock, node)` forms a total order.
/// The higher pair wins. Equal pairs are idempotent — the write
/// is treated as a no-op and reports [`PutOutcome::Rejected`]
/// with the current value (so callers can distinguish "already
/// have it" from "someone raced us").
///
/// Never lowers the on-disk value: a stale write from a partitioned
/// peer is simply ignored.
pub async fn put_stamped(
&self,
key: &RefKey,
incoming: StampedRef,
) -> Result<PutOutcome> {
let path = self.stamped_path(key);
if let Some(current) = self.get_stamped(key).await? {
if !incoming.dominates(&current) {
return Ok(PutOutcome::Rejected { current });
}
}
if let Some(parent) = path.parent() {
tokio::fs::create_dir_all(parent).await.with_context(|| {
format!("creating stamped-ref bucket {}", parent.display())
})?;
}
self.atomic_write(&path, &incoming.to_bytes()).await?;
Ok(PutOutcome::Merged)
}
fn stamped_path(&self, key: &RefKey) -> PathBuf {
let hex = hex32(key);
self.root
.join("refs-v2")
.join(&hex[..2])
.join(format!("{hex}.svref"))
}
async fn atomic_write(&self, final_path: &Path, bytes: &[u8]) -> Result<()> { async fn atomic_write(&self, final_path: &Path, bytes: &[u8]) -> Result<()> {
let tmp_dir = self.root.join(".tmp"); let tmp_dir = self.root.join(".tmp");
let tmp_name = format!( let tmp_name = format!(
@@ -240,4 +374,137 @@ mod tests {
let err = store.get(&key).await.unwrap_err().to_string(); let err = store.get(&key).await.unwrap_err().to_string();
assert!(err.contains("wrong length"), "unexpected: {err}"); assert!(err.contains("wrong length"), "unexpected: {err}");
} }
// ── Phase 3: stamped refs ────────────────────────────────────────
fn stamped(value: u8, clock: u64, node: u8) -> StampedRef {
StampedRef {
value: [value; 32],
clock,
node: [node; 8],
}
}
#[test]
fn stamped_ref_encoding_round_trips() {
let r = stamped(7, 42, 3);
let bytes = r.to_bytes();
assert_eq!(bytes.len(), StampedRef::ENCODED_LEN);
let back = StampedRef::from_bytes(&bytes).unwrap();
assert_eq!(back, r);
}
#[test]
fn stamped_ref_from_bytes_rejects_wrong_length() {
let err = StampedRef::from_bytes(&[0u8; 32])
.unwrap_err()
.to_string();
assert!(err.contains("wrong length"), "unexpected: {err}");
}
#[test]
fn dominates_is_a_total_order_on_clock_then_node() {
// Higher clock always dominates.
let a = stamped(1, 5, 1);
let b = stamped(2, 6, 0);
assert!(b.dominates(&a));
assert!(!a.dominates(&b));
// Same clock: higher node stamp wins.
let c = stamped(3, 10, 1);
let d = stamped(4, 10, 2);
assert!(d.dominates(&c));
assert!(!c.dominates(&d));
// Same everything: neither dominates (idempotent).
assert!(!c.dominates(&c));
}
#[test]
fn node_stamp_is_stable_for_same_name() {
let a = node_stamp_for("tank");
let b = node_stamp_for("tank");
assert_eq!(a, b);
// Different name → different stamp (birthday-bound; tank vs architect definitely differs).
assert_ne!(a, node_stamp_for("architect"));
}
#[tokio::test]
async fn put_stamped_and_get_stamped_round_trip() {
let (_tmp, store) = open();
let key = [9u8; 32];
assert_eq!(store.get_stamped(&key).await.unwrap(), None);
let s = stamped(1, 1, 1);
assert!(matches!(
store.put_stamped(&key, s).await.unwrap(),
PutOutcome::Merged
));
assert_eq!(store.get_stamped(&key).await.unwrap(), Some(s));
}
#[tokio::test]
async fn put_stamped_rejects_older_and_equal_writes() {
// Older by clock: rejected. Equal (clock, node): rejected
// (idempotent). Newer by clock or by node-stamp: merged.
let (_tmp, store) = open();
let key = [0xAB; 32];
let base = stamped(1, 10, 5);
store.put_stamped(&key, base).await.unwrap();
// Older clock → rejected, current returned.
let older = stamped(2, 9, 9);
match store.put_stamped(&key, older).await.unwrap() {
PutOutcome::Rejected { current } => assert_eq!(current, base),
_ => panic!("expected Rejected"),
}
assert_eq!(store.get_stamped(&key).await.unwrap(), Some(base));
// Equal (clock, node) with different value → rejected. Prevents
// "value drift" from a concurrent writer at the same tick.
let dupe = StampedRef {
value: [0xEE; 32],
clock: base.clock,
node: base.node,
};
match store.put_stamped(&key, dupe).await.unwrap() {
PutOutcome::Rejected { current } => assert_eq!(current.value, base.value),
_ => panic!("expected Rejected on equal (clock,node)"),
}
// Same clock but higher node stamp → merged.
let same_clock_higher_node = StampedRef {
value: [0x11; 32],
clock: base.clock,
node: [0x99; 8],
};
assert!(matches!(
store.put_stamped(&key, same_clock_higher_node).await.unwrap(),
PutOutcome::Merged
));
assert_eq!(
store.get_stamped(&key).await.unwrap(),
Some(same_clock_higher_node)
);
// Higher clock always wins regardless of node.
let higher_clock = stamped(0x22, base.clock + 1, 0);
assert!(matches!(
store.put_stamped(&key, higher_clock).await.unwrap(),
PutOutcome::Merged
));
assert_eq!(store.get_stamped(&key).await.unwrap(), Some(higher_clock));
}
#[tokio::test]
async fn stamped_and_unstamped_stores_are_independent() {
// put() writes to refs/, put_stamped() writes to refs-v2/.
// They must not clobber each other.
let (_tmp, store) = open();
let key = [0xCC; 32];
store.put(&key, &[0xAA; 32]).await.unwrap();
let s = stamped(0xBB, 1, 1);
store.put_stamped(&key, s).await.unwrap();
assert_eq!(store.get(&key).await.unwrap(), Some([0xAA; 32]));
assert_eq!(store.get_stamped(&key).await.unwrap(), Some(s));
}
} }
+63 -1
View File
@@ -28,7 +28,7 @@
use crate::cluster::blob::{BlobId, BlobManifest, BlobStat, BlobStore, ChunkHash}; use crate::cluster::blob::{BlobId, BlobManifest, BlobStat, BlobStore, ChunkHash};
use crate::cluster::gossip::{ClusterGossip, PeerView}; use crate::cluster::gossip::{ClusterGossip, PeerView};
use crate::cluster::metrics::{CacheMetrics, MetricsReply}; use crate::cluster::metrics::{CacheMetrics, MetricsReply};
use crate::cluster::refs::{RefKey, RefStore, RefValue}; use crate::cluster::refs::{PutOutcome, RefKey, RefStore, RefValue, StampedRef};
use crate::cluster::tags::{TagEntry, TagStore}; use crate::cluster::tags::{TagEntry, TagStore};
use anyhow::{bail, Context, Result}; use anyhow::{bail, Context, Result};
use quinn::{Connection, ConnectionError}; use quinn::{Connection, ConnectionError};
@@ -137,6 +137,22 @@ pub enum Method {
/// `payload`: 32-byte `RefKey`. Reply: 32 bytes on hit or /// `payload`: 32-byte `RefKey`. Reply: 32 bytes on hit or
/// single-byte [`ErrorCode::NotFound`]. /// single-byte [`ErrorCode::NotFound`].
GetRefLocal = 0x14, GetRefLocal = 0x14,
/// Phase 3 (2026-07-13): stamped/versioned PutRef. Merges under a
/// Lamport-clock + node-stamp total order — concurrent writers
/// can no longer clobber each other silently.
/// `payload`: 32-byte `RefKey` || 32-byte `RefValue` ||
/// 8-byte little-endian `u64` clock || 8-byte node stamp
/// (`blake3(node_name)[0..8]`).
/// Reply: single-byte status —
/// `STREAM_STATUS_OK` = accepted (Merged);
/// [`ErrorCode::AlreadyExists`] = rejected (older/equal).
PutRefVersioned = 0x15,
/// Phase 3: stamped GetRef. Returns the current
/// (value, clock, node) triple as 48 bytes.
/// `payload`: 32-byte `RefKey`.
/// Reply: 48 bytes on hit, single-byte
/// [`ErrorCode::NotFound`] on miss.
GetRefVersioned = 0x16,
} }
impl Method { impl Method {
@@ -164,6 +180,8 @@ impl Method {
0x12 => Some(Method::ListTags), 0x12 => Some(Method::ListTags),
0x13 => Some(Method::GetMetrics), 0x13 => Some(Method::GetMetrics),
0x14 => Some(Method::GetRefLocal), 0x14 => Some(Method::GetRefLocal),
0x15 => Some(Method::PutRefVersioned),
0x16 => Some(Method::GetRefVersioned),
_ => None, _ => None,
} }
} }
@@ -199,6 +217,11 @@ pub enum ErrorCode {
/// configured (e.g. Blob RPCs called on a node with no local /// configured (e.g. Blob RPCs called on a node with no local
/// blob store). /// blob store).
NotConfigured = 0xf5, NotConfigured = 0xf5,
/// Phase 3 (2026-07-13): the write was rejected because a prior
/// value with an equal or higher `(clock, node)` already exists.
/// The current value stayed on disk. Used by
/// [`Method::PutRefVersioned`] to signal a "not merged" outcome.
AlreadyExists = 0xf6,
} }
impl ErrorCode { impl ErrorCode {
@@ -213,6 +236,9 @@ impl ErrorCode {
ErrorCode::NotFound => "not found", ErrorCode::NotFound => "not found",
ErrorCode::InvalidRequest => "invalid request", ErrorCode::InvalidRequest => "invalid request",
ErrorCode::NotConfigured => "server subsystem not configured", ErrorCode::NotConfigured => "server subsystem not configured",
ErrorCode::AlreadyExists => {
"rejected: existing value dominates incoming write"
}
} }
} }
} }
@@ -609,6 +635,42 @@ impl RpcRouter {
store.put(&key, &value).await?; store.put(&key, &value).await?;
Ok(HandlerOutcome::Reply(vec![STREAM_STATUS_OK])) Ok(HandlerOutcome::Reply(vec![STREAM_STATUS_OK]))
} }
Method::PutRefVersioned => {
let store = match &self.ref_store {
Some(s) => s,
None => return Ok(HandlerOutcome::Error(ErrorCode::NotConfigured)),
};
// Wire: 32-key || 32-value || 8-clock (LE) || 8-node = 80.
if payload.len() != 32 + StampedRef::ENCODED_LEN {
return Ok(HandlerOutcome::Error(ErrorCode::InvalidRequest));
}
let mut key = [0u8; 32];
key.copy_from_slice(&payload[..32]);
let incoming = match StampedRef::from_bytes(&payload[32..]) {
Ok(v) => v,
Err(_) => return Ok(HandlerOutcome::Error(ErrorCode::InvalidRequest)),
};
match store.put_stamped(&key, incoming).await? {
PutOutcome::Merged => Ok(HandlerOutcome::Reply(vec![STREAM_STATUS_OK])),
PutOutcome::Rejected { .. } => {
Ok(HandlerOutcome::Error(ErrorCode::AlreadyExists))
}
}
}
Method::GetRefVersioned => {
let store = match &self.ref_store {
Some(s) => s,
None => return Ok(HandlerOutcome::Error(ErrorCode::NotConfigured)),
};
let key = match decode_32(payload) {
Some(k) => k,
None => return Ok(HandlerOutcome::Error(ErrorCode::InvalidRequest)),
};
match store.get_stamped(&key).await? {
Some(s) => Ok(HandlerOutcome::Reply(s.to_bytes().to_vec())),
None => Ok(HandlerOutcome::Error(ErrorCode::NotFound)),
}
}
Method::PutTag => { Method::PutTag => {
let store = match &self.tag_store { let store = match &self.tag_store {
Some(s) => s, Some(s) => s,
+56
View File
@@ -180,6 +180,7 @@ pub fn decode_error(b: u8) -> Option<ErrorCode> {
0xf3 => Some(ErrorCode::NotFound), 0xf3 => Some(ErrorCode::NotFound),
0xf4 => Some(ErrorCode::InvalidRequest), 0xf4 => Some(ErrorCode::InvalidRequest),
0xf5 => Some(ErrorCode::NotConfigured), 0xf5 => Some(ErrorCode::NotConfigured),
0xf6 => Some(ErrorCode::AlreadyExists),
_ => None, _ => None,
} }
} }
@@ -472,6 +473,61 @@ pub async fn call_put_ref(
} }
} }
// ── Phase 3: stamped-ref client helpers ──────────────────────────────
/// Phase 3 (2026-07-13): submit a stamped (CRDT-merge) PutRef.
///
/// Returns `Ok(true)` when the peer merged the write (Merged),
/// `Ok(false)` when the peer rejected it because an equal or
/// higher `(clock, node)` already exists (AlreadyExists). Any
/// other reply is an error.
pub async fn call_put_ref_versioned(
conn: &Connection,
key: &crate::cluster::refs::RefKey,
incoming: &crate::cluster::refs::StampedRef,
) -> Result<bool> {
let mut payload = Vec::with_capacity(32 + crate::cluster::refs::StampedRef::ENCODED_LEN);
payload.extend_from_slice(key);
payload.extend_from_slice(&incoming.to_bytes());
let reply = rpc_call(conn, Method::PutRefVersioned, &payload).await?;
if reply.len() != 1 {
bail!(
"expected single-byte PutRefVersioned reply, got {} bytes",
reply.len()
);
}
match reply[0] {
STREAM_STATUS_OK => Ok(true),
code => match decode_error(code) {
Some(ErrorCode::AlreadyExists) => Ok(false),
Some(err) => bail!("peer rejected PutRefVersioned: {}", err.describe()),
None => bail!(
"peer replied with unknown byte 0x{:02x} for PutRefVersioned",
code
),
},
}
}
/// Phase 3: fetch a stamped (CRDT-merge) ref.
pub async fn call_get_ref_versioned(
conn: &Connection,
key: &crate::cluster::refs::RefKey,
) -> Result<Option<crate::cluster::refs::StampedRef>> {
let reply = rpc_call(conn, Method::GetRefVersioned, key).await?;
if reply.len() == 1 {
match decode_error(reply[0]) {
Some(ErrorCode::NotFound) => return Ok(None),
Some(err) => bail!("peer replied with error: {}", err.describe()),
None => {}
}
}
Ok(Some(
crate::cluster::refs::StampedRef::from_bytes(&reply)
.context("decoding stamped ref reply")?,
))
}
// ── Phase 5g: cache metrics client helper ──────────────────────────── // ── Phase 5g: cache metrics client helper ────────────────────────────
/// Fetch the peer's current cache-metrics snapshot. /// Fetch the peer's current cache-metrics snapshot.
@@ -1273,3 +1273,94 @@ async fn parallel_blob_get_reassembles_multi_chunk_blob_byte_equal() {
tokio::time::sleep(Duration::from_millis(50)).await; tokio::time::sleep(Duration::from_millis(50)).await;
acc_a.abort(); acc_a.abort();
} }
// ── Phase 3: stamped-ref RPC end-to-end ──────────────────────────────
#[tokio::test]
async fn end_to_end_put_ref_versioned_merges_and_rejects() {
// Two writers publish stamped refs for the same key. Higher
// (clock, node) wins; lower is rejected; equal is idempotent.
use crate::cluster::refs::{node_stamp_for, StampedRef};
let (id_a, id_b) = NodeIdentity::generate_test_pair("a", "b").unwrap();
let (_tmp_a, router_a) = router_with_blobs_and_refs("a", next_port()).await;
let server_a = QuicServer::bind(loopback(0), id_a).unwrap();
let addr = server_a.local_addr().unwrap();
let ra = router_a.clone();
let acc = tokio::spawn(async move {
while let Some(Ok(conn)) = server_a.accept().await {
let r = ra.clone();
tokio::spawn(async move {
let _ = serve_connection(conn, r).await;
});
}
});
let client = QuicClient::new(loopback(0), id_b).unwrap();
let conn = client.connect(addr, "a").await.unwrap();
let key = [0x77; 32];
let node_x = node_stamp_for("runner-x");
let node_y = node_stamp_for("runner-y");
// Empty: nothing to return.
assert_eq!(call_get_ref_versioned(&conn, &key).await.unwrap(), None);
// First write: merged.
let first = StampedRef {
value: [0xA1; 32],
clock: 10,
node: node_x,
};
assert!(call_put_ref_versioned(&conn, &key, &first).await.unwrap());
assert_eq!(
call_get_ref_versioned(&conn, &key).await.unwrap(),
Some(first)
);
// Older clock from a different writer: rejected.
let older = StampedRef {
value: [0xA2; 32],
clock: 9,
node: node_y,
};
assert!(!call_put_ref_versioned(&conn, &key, &older).await.unwrap());
assert_eq!(
call_get_ref_versioned(&conn, &key).await.unwrap(),
Some(first),
"older write must not overwrite"
);
// Same clock, higher node stamp: merged if node_y > node_x, else rejected.
let tied = StampedRef {
value: [0xA3; 32],
clock: 10,
node: node_y,
};
let merged = call_put_ref_versioned(&conn, &key, &tied).await.unwrap();
let after_tie = call_get_ref_versioned(&conn, &key).await.unwrap().unwrap();
if node_y > node_x {
assert!(merged, "higher node stamp should merge");
assert_eq!(after_tie, tied);
} else {
assert!(!merged, "lower node stamp should be rejected");
assert_eq!(after_tie, first);
}
// Higher clock always wins.
let latest = StampedRef {
value: [0xA4; 32],
clock: 100,
node: node_x,
};
assert!(call_put_ref_versioned(&conn, &key, &latest).await.unwrap());
assert_eq!(
call_get_ref_versioned(&conn, &key).await.unwrap(),
Some(latest)
);
conn.close(quinn::VarInt::from_u32(0), b"done");
client.shutdown().await;
tokio::time::sleep(Duration::from_millis(50)).await;
acc.abort();
}