feat(m11): data-plane sharding × replication (m11p6 L0-L2)
ClusterNode hosts a BTreeMap<ShardId, Arc<ShardReplica>>: writes hash-route to the owning shard leader, reads scatter over shard groups. In-group shard==region preserved so the engine and tidal-net are untouched; S=1 stays byte-for-byte (today's cluster is a 1-shard × RF=N group). Topology grows shard-group awareness; membership, election, forward, reseed, and join_boot thread ShardId through. Proven by an in-process 2×2 RF=2 gRPC test plus S=1 parity, incl. tier-3 real-OS-process failover. clippy/fmt clean.
This commit is contained in:
jx12n
parent
25ec7630a1
commit
3bfde53b90
259
docs/planning/milestone-11/phase-6.md
Normal file
259
docs/planning/milestone-11/phase-6.md
Normal file
@ -0,0 +1,259 @@
|
||||
# m11p6 — Sharding × Replication + Rebalancing (IN PROGRESS)
|
||||
|
||||
Phase spec and exit gate: [docs/roadmap-to-cluster.md §4/m11p6](../../roadmap-to-cluster.md).
|
||||
Closes the ROADMAP **write-scaling** gap ("EITHER replicated (1 leader for
|
||||
everything) OR sharded (no replication)").
|
||||
Predecessors: p1 (ship queue/perf), p2 (one log), p3 (quorum/commit-index),
|
||||
p4 (election/fencing), p5 (membership/discovery/elasticity).
|
||||
|
||||
**Goal:** writes scale horizontally **without** giving up replication. The
|
||||
"replicated XOR sharded" split ends: `S` shard groups, each a replication group
|
||||
at RF (default 3) with its **own WAL/relay/commit-index and its own elected
|
||||
leader**, leaders balanced across nodes; regions demote to placement labels;
|
||||
any gateway hash-routes to the shard leader (write) or nearest replica (read);
|
||||
rebalancing = snapshot + stream + fenced cutover.
|
||||
|
||||
## The load-bearing realization (why this is tractable, not a rewrite)
|
||||
|
||||
A single `TidalDb` is **already a complete, self-contained shard** (scale-arch
|
||||
spec §1; single-process mode already runs N real `TidalDb` in one process). And
|
||||
**today's cluster is exactly "1 shard × RF=N"** — every region replicates one
|
||||
log under one leader. So m11p6 = **run today's whole single-shard cluster
|
||||
machinery `N` times, once per data-shard group**, where each group spans only
|
||||
its replica nodes, in its own data subdir + gRPC port. The data-shard number
|
||||
`S` is purely a *gateway-routing + directory + port* concern; **inside a group,
|
||||
"shard id == region id" is preserved unchanged**.
|
||||
|
||||
Consequences:
|
||||
- **Engine replication (relay/commit/election/ship/receiver/membership): NO
|
||||
semantic change** — each runs per-group exactly as today's whole cluster.
|
||||
- **`tidal-net`: NO change** — each group owns its own `GrpcTransport` bound to
|
||||
a distinct port, peering only with that group's replica nodes.
|
||||
- **`S=1` is byte-for-byte today's behavior** — every existing tier-3 suite
|
||||
(quorum, membership, election, region, runbook, chaos) runs in `S=1` mode and
|
||||
stays green by construction.
|
||||
- The work concentrates in **`tidal-server`**: decompose the per-shard
|
||||
machinery out of `RegionClusterState` into `ShardReplica`; the node hosts
|
||||
`BTreeMap<ShardId, Arc<ShardReplica>>`; add hash-routing to the right group's
|
||||
leader. Plus topology schema, route unification, rebalancing, stress/harness,
|
||||
the exit gate, docs.
|
||||
|
||||
## Design (as adopted)
|
||||
|
||||
### 1. Two types: process node + shard replica
|
||||
|
||||
> **As-built note (decided during L1):** the §1-era naming below was superseded.
|
||||
> The process node is a NEW type **`ClusterNode`** (not "keep the name
|
||||
> `RegionClusterState`"); the former `RegionClusterState` was **renamed to
|
||||
> `ShardReplica`** (it already WAS one group's machinery). The responsibilities
|
||||
> are exactly as described — only the node's type name changed from the planned
|
||||
> `RegionClusterState` to `ClusterNode`.
|
||||
|
||||
- **`ClusterNode` (NEW process node)** = the **process node** and the axum
|
||||
`State`. Owns the node identity (`region`/`region_name`), the gateway
|
||||
`ShardRouter` + `node_http` map, the shared forward client, and
|
||||
`groups: BTreeMap<ShardId, Arc<ShardReplica>>` (the groups this node hosts)
|
||||
plus a `placement` map of every group (to forward writes for groups it does
|
||||
NOT host). The HTTP handlers stay free functions over
|
||||
`State<Arc<ClusterNode>>`; each hash-routes the target group and either calls
|
||||
a local `ShardReplica` method or forwards to that group's leader.
|
||||
- **`ShardReplica` (the renamed `RegionClusterState`)** = one shard group's full
|
||||
replication machinery, essentially the per-shard fields + methods of the old
|
||||
`RegionClusterState`:
|
||||
`db: Arc<TidalDb>`, `ship_feed`, `ship_queue`, `commit` + `commit_watch`
|
||||
bridge, `election_runtime`/`election_store`/`election_boot`,
|
||||
`membership: MembershipView` + `membership_store`, `stream_baseline`,
|
||||
snapshot source + `reseed_marker_store`, `leader: RwLock<Option<RegionId>>`,
|
||||
`activation_prev`, `partitioned`, `deferred_retires`, `ack_default`,
|
||||
`quorum_timeout`, and its **own `GrpcTransport`** bound to this group's port.
|
||||
Construction = today's `RegionClusterState::new` body, parameterized by
|
||||
`(shard, replica nodes, bootstrap leader, group data subdir, group grpc port)`.
|
||||
|
||||
### 2. Shard identity, ports, and data dirs
|
||||
|
||||
- **In-group identity is region-id based** (preserved from today). `ShardReplica`
|
||||
for data-shard `S` on node `N` opens `TidalDb` with
|
||||
`NodeConfig{ shard_id: ShardId(N.region_id), peer_shards: <other group-S
|
||||
nodes' region ids>, .. }`. Two groups on one node use the same `shard_id`
|
||||
value but live in separate data dirs + separate transports, so nothing
|
||||
collides. The data-shard `S` never enters the replication wire.
|
||||
- **gRPC ports:** each replica entry may set an explicit `grpc_addr`/`grpc_bind`
|
||||
per `(node, shard)`; otherwise derive `node.base_port + S`. Legacy `S=0` →
|
||||
offset 0 → the region's declared address verbatim.
|
||||
- **Data dirs:** explicit `shards:` → `<data_dir>/shard-{S:05}/` per hosted
|
||||
group. Legacy (synthesized 1-shard) → `<data_dir>` **verbatim** (existing
|
||||
on-disk clusters restart unchanged).
|
||||
- **`/metrics`:** the node owns ONE listener on `metrics_addr`; the engine's
|
||||
per-`TidalDb` metrics HTTP server is suppressed in cluster mode. The node
|
||||
renders each group's `cluster_metrics` with a `shard="S"` label and the
|
||||
per-shard engine metrics likewise.
|
||||
|
||||
### 3. Topology schema (backward compatible)
|
||||
|
||||
`TopologySpec` gains optional `shards: Option<Vec<ShardSpec>>` and `RegionSpec`
|
||||
gains optional `zone: Option<String>` (placement/read-affinity label), both
|
||||
`#[serde(default)]`.
|
||||
|
||||
```yaml
|
||||
nodes: # `regions:` still accepted as the node list
|
||||
- { name: us-east, grpc_addr: ..., http_addr: ..., zone: az-a }
|
||||
shards: # NEW — optional
|
||||
- id: 0
|
||||
leader: us-east # term-0 / preferred leader (balances placement)
|
||||
replicas: # nodes hosting this group's RF replicas
|
||||
- { node: us-east } # grpc_addr/grpc_bind optional → derive base+S
|
||||
- { node: eu-west }
|
||||
- { node: ap-south }
|
||||
- { id: 1, leader: eu-west, replicas: [...] }
|
||||
- { id: 2, leader: ap-south, replicas: [...] }
|
||||
```
|
||||
|
||||
**Absent `shards:` ⇒ legacy synthesis:** one `ShardSpec{ id: 0,
|
||||
leader: topology.leader, replicas: <every region, grpc verbatim> }`. This is
|
||||
literally today's "1 shard × RF=all-regions," so the model unifies cleanly.
|
||||
|
||||
Validation: shard ids dense & unique; each `leader`/`replica.node` names a
|
||||
declared region; RF ≥ 1; a node's shards do not collide on derived ports.
|
||||
|
||||
### 4. Routing (the new gateway layer)
|
||||
|
||||
Any node, any write to entity `E`:
|
||||
1. `S = ShardRouter.route(E)` (the engine's FNV-1a hash; the same router the
|
||||
`/sharded/*` path already uses).
|
||||
2. If this node hosts group `S` **and leads it** → apply locally on
|
||||
`ShardReplica[S]` (the existing stage→complete→ship→await-quorum path).
|
||||
3. If this node hosts group `S` but follows → forward to group `S`'s leader
|
||||
(resolved from `ShardReplica[S].leader`), to the leader node's `http_addr`.
|
||||
4. If this node does not host group `S` → forward to any replica node of `S`
|
||||
(from topology); that node routes to the leader. Term fencing + the
|
||||
`NotLeader` retry bound the forward chain.
|
||||
|
||||
Reads route to a local replica of `S` if present, else nearest (zone) / any
|
||||
replica. **Scatter-gather merges over shard GROUPS:** one replica per group,
|
||||
totals **summed** (entity-sharded groups are disjoint — no dedup-by-max), the
|
||||
existing degraded/`unavailable_shards`/deadline semantics retained.
|
||||
|
||||
**Route unification:** `/items`, `/embeddings`, `/signals` become the shard-
|
||||
routed surface (they hash-route instead of forwarding to a single global
|
||||
leader); `/sharded/*` is retained as an alias of the same path. `x-tidal-ack`,
|
||||
`x-tidal-seq`, quorum await, and `NotLeader`/`QuorumTimeout` all become
|
||||
per-shard; `NotLeader` names the shard and its leader.
|
||||
|
||||
### 5. Leadership balance
|
||||
|
||||
`ShardSpec.leader` is each group's term-0 / preferred leader; the operator (and
|
||||
the test harness) set shard `i`'s leader to node `i` for balanced placement.
|
||||
Each group boot-classifies its own leader exactly like today's `topology.leader`,
|
||||
but per group. On failover, the group's election picks the surviving max-applied
|
||||
voter (m11p4 protocol, unchanged); only the dead node's groups elect. An
|
||||
optional leadership-transfer-back-to-preferred is the rebalance path, not the
|
||||
availability mechanism.
|
||||
|
||||
### 6. Rebalancing (operator-triggered)
|
||||
|
||||
Shard move / replica change reuses m11p5 verbatim, **per group**: a conf-change
|
||||
on the group's own log adds a Learner replica (a new node, or an existing node
|
||||
gaining a replica of group `S`), which catches up via `FetchSnapshot` +
|
||||
`StreamSegments` (per-group, already per-instance once §1 lands), auto-promotes
|
||||
Learner→Voter, then the operator may transfer leadership; removing a replica is
|
||||
the m11p5 fenced removal on that group's log. Endpoints:
|
||||
`POST /cluster/shards/{id}/replicas` (add/remove) and
|
||||
`POST /cluster/shards/{id}/transfer`. Per the roadmap, **automatic rate-limited
|
||||
rebalancing is explicitly "later"** — m11p6 ships the operator verbs.
|
||||
|
||||
## Exit gate (from the roadmap)
|
||||
|
||||
- 3 shards × RF=3: **≥5,000 quorum signals/s** (≥2.5× single-shard p3).
|
||||
- Kill any node → **only its shard-leaderships move (<10 s), reads never stop**.
|
||||
- `tidal-stress` sharded-vs-replicated comparison **collapses into one path**.
|
||||
|
||||
## Layer plan (each keeps the workspace green)
|
||||
|
||||
- **L0** — this doc + topology schema (`shards:`/`zone:` + validation + legacy
|
||||
synthesis). `S=1` unchanged.
|
||||
- **L1** — extract `ShardReplica`; node hosts a one-entry map; suppress engine
|
||||
metrics server in cluster mode + node aggregates. `S=1` byte-for-byte.
|
||||
- **L2** — multi-shard construction + gateway hash-routing + per-shard forward +
|
||||
route unification + scatter-gather over groups.
|
||||
- **L3** — operator rebalancing verbs.
|
||||
- **L4** — tidal-stress path collapse + nodes×shards tier-3 harness +
|
||||
`cluster_sharding.rs` exit gate (run for real).
|
||||
- **L5** — docs (runbook/monitoring/roadmap/CHANGELOG/k8s/spec) + memory.
|
||||
|
||||
## Status
|
||||
|
||||
- [x] L0 topology schema (`shards:`/`zone:` + resolver + validation; legacy
|
||||
synthesis) + this doc — 20 topology tests green incl. 5 new shard tests.
|
||||
- [x] L1 `RegionClusterState` → `ShardReplica` rename + group-parameterized
|
||||
`ShardReplica::new(.., group, enable_metrics)` (leader/peers/voters/data-
|
||||
dir/metrics from the resolved group; in-group identity stays region-based).
|
||||
**S=1 byte-for-byte green** (116 lib + cluster_region 11 + cluster_routes 6
|
||||
+ cluster_grpc 2). NB: the `ClusterNode` process wrapper + multi-shard
|
||||
hosting moved to L2 (inseparable from the routing/handler rewiring).
|
||||
- [x] L2 **DONE + green** (the sharding × replication data plane). `ClusterNode`
|
||||
(the axum State + process handle) hosts `BTreeMap<ShardId, Arc<ShardReplica>>`,
|
||||
built by `ClusterNode::new(topology, region, schema, profiles, data_dir,
|
||||
hlc)` (resolves groups; opens one `ShardReplica` per hosted group — own data
|
||||
subdir `shard-<id>/` + derived/explicit gRPC port for `S>1`; node data dir
|
||||
verbatim + metrics owner for `S=1`). `ServeState for ClusterNode` starts
|
||||
every group's election driver and shuts each down via `Arc::try_unwrap`.
|
||||
Entity writes (`/items`,`/embeddings`,`/signals`,`/hardnegs`) hash-route via
|
||||
`route_entity` → local replica (existing leader/forward path) or
|
||||
`forward_to_group_node` (remote group). Reads (`/feed`,`/search`) scatter
|
||||
in-process over hosted groups and merge (sum totals, score-sort — S=1
|
||||
unchanged). `/cluster/status/local` gains a per-shard `shards[]` array;
|
||||
`region_health` aggregates all groups.
|
||||
**Verified:** in-process **2 nodes × 2 shards × RF=2** test
|
||||
(`region_sharded_writes_route_per_shard_and_reads_scatter`) — writes route
|
||||
per-shard (A forwards shard-1 writes to B), each shard replicates to its
|
||||
follower, reads scatter over both groups and return all items from BOTH
|
||||
shards, over real gRPC. **S=1 byte-for-byte:** 116 lib + cluster_region 12 +
|
||||
cluster_routes 6 + cluster_grpc 2 + tier-3 `cluster_multiproc` 5 (real OS
|
||||
processes: replication <2s, leader-crash failover <10s). clippy `-D` + fmt
|
||||
clean. **Deferred to L2-followups (tracked):** per-shard admin verbs
|
||||
(`?shard=` on promote/heal/partition/etc.) and `/sharded/*`→unified-path
|
||||
aliasing — currently the admin verbs + `/sharded/*` target the first hosted
|
||||
group (correct for `S=1`; multi-shard admin needs the selector).
|
||||
|
||||
**Known L2 limitations — `S>1` only, FULL-placement assumed (tracked for L4,
|
||||
NOT silently dropped):** these are correct for `S=1` and for the headline
|
||||
exit-gate shape (3×3 RF=3, where every node hosts every group), and the writes
|
||||
proven by the 2×2 test are unaffected — but a genuinely PARTIAL placement (a
|
||||
node hosting a strict subset of groups) hits them:
|
||||
1. **Unified reads are local-only.** `/feed`//`/search` scatter over the
|
||||
groups THIS node hosts (`hosted_dbs`); a node that does not host every group
|
||||
returns a strict-subset corpus with a 200 and no `degraded` flag. Complete
|
||||
only under full placement; cross-node read fan-out (reuse the `/sharded/*`
|
||||
`HttpShardContext`) is the L4 item. Use `/sharded/*` for partial placements.
|
||||
2. **No cross-shard diversity re-rank.** The scatter merge keeps each group's
|
||||
own diversity pass then score-merges; it does not re-run diversity ACROSS
|
||||
groups, so an `S>1` `/feed` can be less diverse than `S=1`. Summing totals
|
||||
over disjoint groups is exact for cardinality; a cross-shard MMR re-rank
|
||||
(engine entry point) is the L4 follow-up.
|
||||
3. **`S>1` observability is one-group-deep.** Only the metrics-owner group
|
||||
binds the engine `/metrics` listener; the other hosted groups' per-shard
|
||||
`cluster_metrics` (ship/relay/commit/quorum) are not rendered. The
|
||||
kill-node gate uses `/cluster/status/local` (and now `/cluster/status`)
|
||||
`shards[]`, which IS per-group; node-level per-shard `/metrics` rendering
|
||||
(design §2) is the L4/p8 item.
|
||||
- [ ] L3 rebalancing verbs (operator shard move; reuse m11p5 per-group)
|
||||
- [ ] L4 tidal-stress path collapse + nodes×shards tier-3 harness +
|
||||
`cluster_sharding.rs` exit gate (3×3 kill-node + per-shard ledger; run for
|
||||
real). NB: the ≥5,000/s throughput sub-gate is a Ref-A line item — k3s
|
||||
access has blocked the Ref-A runs since p1 (same standing caveat).
|
||||
- [ ] L5 docs (runbook/monitoring/roadmap/CHANGELOG/k8s/spec) + memory
|
||||
|
||||
## Exit-gate evidence
|
||||
|
||||
**Data-plane correctness (local, in-process, real gRPC) — DONE.** The 2×2 test
|
||||
proves the replicated-XOR-sharded split is gone: a single gateway routes writes
|
||||
to the owning shard's leader (forwarding cross-node), each shard group replicates
|
||||
independently, and corpus reads merge across groups. The per-shard election +
|
||||
quorum + commit machinery is the unchanged m11p4/m11p3 code instantiated per
|
||||
group, and S=1 failover is re-proven over real OS processes (`cluster_multiproc`).
|
||||
|
||||
**Remaining for the headline exit gate:** the 3 shards × RF=3 tier-3 run over real
|
||||
OS processes (kill any node → only its shard-leaderships move <10 s, reads never
|
||||
stop; per-shard zero-acked-loss ledger) and the ≥5,000/s `tidal-stress` figure
|
||||
(Ref-A-pending). These need the L4 harness extension (per-(node,shard) ports +
|
||||
`shards:` emission) + the rebalance verbs (L3).
|
||||
@ -3,7 +3,7 @@
|
||||
//! The pure state machine lives in the engine
|
||||
//! ([`tidaldb::replication::ElectionState`]); this module owns everything the
|
||||
//! machine deliberately does not — timers, RPC fan-outs, durable persistence,
|
||||
//! and the leadership transitions on [`RegionClusterState`]:
|
||||
//! and the leadership transitions on [`ShardReplica`]:
|
||||
//!
|
||||
//! * **One driver thread per node** ticks the machine every
|
||||
//! [`TICK_INTERVAL`] and drains the [`ElectionNetEvent`] inbox (vote
|
||||
@ -39,7 +39,7 @@ use tidaldb::replication::{
|
||||
shard::{RegionId, ShardId},
|
||||
};
|
||||
|
||||
use super::{node::RegionClusterState, topology::shard_of_region};
|
||||
use super::{node::ShardReplica, topology::shard_of_region};
|
||||
|
||||
/// Driver tick cadence: the machine's own deadlines (election timeout,
|
||||
/// heartbeat due, lease) are all ≥ 100ms-scale, so 50ms keeps every deadline
|
||||
@ -62,7 +62,7 @@ pub struct ElectionRuntime {
|
||||
quarantined: AtomicBool,
|
||||
/// The node, for transitions and log-position reads. Weak: the runtime
|
||||
/// is owned BY the node; a strong ref would leak both.
|
||||
node: Weak<RegionClusterState>,
|
||||
node: Weak<ShardReplica>,
|
||||
net: ElectionNet,
|
||||
/// The driver's inbox for async RPC outcomes.
|
||||
inbox_tx: mpsc::Sender<ElectionNetEvent>,
|
||||
@ -697,7 +697,7 @@ fn decide_join(term: u64, own: LogPosition, prev_log: LogPosition) -> JoinDecisi
|
||||
/// spawn the driver thread. Returns the runtime handle the node stores (for
|
||||
/// status reads and shutdown).
|
||||
pub fn start(
|
||||
node: &Arc<RegionClusterState>,
|
||||
node: &Arc<ShardReplica>,
|
||||
machine: ElectionState,
|
||||
store: tidaldb::replication::ElectionStore,
|
||||
hooks_cell: &Arc<OnceLock<Arc<dyn tidal_net::ElectionHooks>>>,
|
||||
|
||||
@ -21,7 +21,7 @@
|
||||
//! concurrently with the marker set, collecting per-peer success/failure so the
|
||||
//! handler can report `{"replicated_to": n, "failed": [..]}` honestly.
|
||||
//!
|
||||
//! The shared [`reqwest::Client`] lives on [`RegionClusterState`] (connection
|
||||
//! The shared [`reqwest::Client`] lives on [`ShardReplica`] (connection
|
||||
//! pooling across requests); its timeouts are tight (connect 1s, request 5s) so
|
||||
//! a dead peer degrades a forward into a clean 503/partial result instead of
|
||||
//! hanging an axum worker. Operation-specific budgets override the client
|
||||
@ -32,7 +32,11 @@
|
||||
|
||||
use std::time::Duration;
|
||||
|
||||
use axum::http::{HeaderMap, StatusCode, header::AUTHORIZATION};
|
||||
use axum::{
|
||||
Json,
|
||||
http::{HeaderMap, HeaderName, HeaderValue, StatusCode, header::AUTHORIZATION},
|
||||
response::{IntoResponse, Response},
|
||||
};
|
||||
use serde::Serialize;
|
||||
|
||||
/// Header name for the internal-propagation marker.
|
||||
@ -97,7 +101,7 @@ pub const STATUS_PEER_TIMEOUT: Duration = Duration::from_millis(500);
|
||||
pub const BROADCAST_PEER_TIMEOUT: Duration = Duration::from_secs(2);
|
||||
|
||||
/// Build the shared forwarding client with the standard connect/request
|
||||
/// timeouts. One per node (held on [`RegionClusterState`]) so connections pool
|
||||
/// timeouts. One per node (held on [`ShardReplica`]) so connections pool
|
||||
/// across requests.
|
||||
///
|
||||
/// # Errors
|
||||
@ -151,6 +155,43 @@ pub struct ForwardedResponse {
|
||||
pub body: serde_json::Value,
|
||||
}
|
||||
|
||||
/// Rebuild an axum [`Response`] from a [`ForwardedResponse`], relaying the
|
||||
/// replicated-log verdict headers (`x-tidal-seq` / `x-tidal-deduplicated`) the
|
||||
/// peer set. This is the ONE place a forwarded write's quorum/dedup verdict is
|
||||
/// re-attached to the client-facing response, so the follower→leader hop
|
||||
/// (`forward_write`) and the cross-shard gateway hop (`forward_to_group_node`)
|
||||
/// cannot drift on the relay contract.
|
||||
#[must_use]
|
||||
pub fn relay_forwarded(resp: ForwardedResponse) -> Response {
|
||||
let mut response = (resp.status, Json(resp.body)).into_response();
|
||||
if let Some(seq) = resp.seq
|
||||
&& let Ok(value) = HeaderValue::from_str(&seq)
|
||||
{
|
||||
response
|
||||
.headers_mut()
|
||||
.insert(HeaderName::from_static(SEQ_HEADER), value);
|
||||
}
|
||||
if resp.deduplicated {
|
||||
response.headers_mut().insert(
|
||||
HeaderName::from_static(DEDUP_HEADER),
|
||||
HeaderValue::from_static(DEDUP_HEADER_VALUE),
|
||||
);
|
||||
}
|
||||
response
|
||||
}
|
||||
|
||||
/// The `x-tidal-ack` override (m11p3) as a forward passthrough entry, if the
|
||||
/// caller set one — so a forwarded write reaches the leader with the CALLER's
|
||||
/// ack mode, not the gateway's default. Empty when absent.
|
||||
#[must_use]
|
||||
pub fn ack_passthrough(headers: &HeaderMap) -> Vec<(&'static str, String)> {
|
||||
headers
|
||||
.get(ACK_HEADER)
|
||||
.and_then(|v| v.to_str().ok())
|
||||
.map(|v| vec![(ACK_HEADER, v.to_owned())])
|
||||
.unwrap_or_default()
|
||||
}
|
||||
|
||||
/// Forward a JSON request to one peer and relay its status + body back.
|
||||
///
|
||||
/// `auth` is the caller's verbatim `Authorization` header (passed through so the
|
||||
|
||||
@ -20,21 +20,21 @@
|
||||
//! serves from 1).
|
||||
//! 5. **Synthesize a topology** from the JOIN-RESPONSE roster (this node's
|
||||
//! `--region` = its assigned id), carrying the local file's behavioral knob
|
||||
//! blocks (§3.5). `RegionClusterState::new` builds the transport/ship/election
|
||||
//! blocks (§3.5). `ShardReplica::new` builds the transport/ship/election
|
||||
//! peer tables from THIS, not the local topology's `regions:` list.
|
||||
//!
|
||||
//! A **restart of a joiner** (the data dir already has a WAL + a membership
|
||||
//! cache) skips the seed loop: it boots from the cache roster (no leader
|
||||
//! discovery, no fetch) and the runtime catch-up self-heal converges it. The
|
||||
//! authoritative roster once the WAL opens is the recovered `ClusterMembership`
|
||||
//! cell (§3.6 precedence, resolved inside `RegionClusterState::new`); the cache
|
||||
//! cell (§3.6 precedence, resolved inside `ShardReplica::new`); the cache
|
||||
//! and the synthesized topology only feed the pre-open peer-table construction,
|
||||
//! and the cache is rewritten from the cell after every applied record.
|
||||
//!
|
||||
//! # A joiner is a LEARNER
|
||||
//!
|
||||
//! The join response role is `Learner`. The synthesized topology marks the
|
||||
//! joiner's role so `RegionClusterState`'s era-0 view (when no cell exists yet)
|
||||
//! joiner's role so `ShardReplica`'s era-0 view (when no cell exists yet)
|
||||
//! does not mistake it for a voter — but in practice a snapshot install always
|
||||
//! carries the leader's WAL with the kind-4 Learner record, so the view boots
|
||||
//! `from_record` with the correct role. `ElectionState` boots with `self` NOT in
|
||||
@ -69,14 +69,14 @@ const BACKOFF_MAX: Duration = Duration::from_secs(5);
|
||||
const DEFAULT_JOIN_WINDOW_MS: u64 = 120_000;
|
||||
|
||||
/// The synthesized topology a seed-join boot constructs for
|
||||
/// `RegionClusterState::new` (§3.4).
|
||||
/// `ShardReplica::new` (§3.4).
|
||||
///
|
||||
/// The join-response roster becomes the `regions:` list, the local file's knob
|
||||
/// blocks are carried verbatim, and this node is included at its assigned id.
|
||||
pub struct SeedJoinBoot {
|
||||
/// The synthesized topology (roster from the join, knobs from the local file).
|
||||
pub topology: TopologySpec,
|
||||
/// This node's region name (echoed for `RegionClusterState::new`).
|
||||
/// This node's region name (echoed for `ShardReplica::new`).
|
||||
pub region: String,
|
||||
}
|
||||
|
||||
@ -117,7 +117,7 @@ pub fn seed_join_boot(input: &SeedJoinInput<'_>) -> Result<SeedJoinBoot> {
|
||||
// §3.6 precedence for the PRE-OPEN loop: a restart of a joiner whose data dir
|
||||
// already carries a WAL boots from the cache WITHOUT the seed. The
|
||||
// authoritative roster once the WAL opens is the recovered cell (resolved
|
||||
// inside `RegionClusterState::new`); this cache only seeds the peer tables.
|
||||
// inside `ShardReplica::new`); this cache only seeds the peer tables.
|
||||
let wal_exists = input
|
||||
.data_dir
|
||||
.join("wal")
|
||||
@ -589,7 +589,7 @@ fn adopt_term(data_dir: &Path, term: u64) -> Result<()> {
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Synthesize the topology `RegionClusterState::new` consumes (§3.4): the
|
||||
/// Synthesize the topology `ShardReplica::new` consumes (§3.4): the
|
||||
/// join-response roster as `regions:` (in id order so the positional `RegionId`
|
||||
/// MATCHES the assigned member id — the seam every durable id depends on), this
|
||||
/// node's `--region` carried, the local file's knob blocks copied verbatim
|
||||
@ -644,6 +644,7 @@ fn synthesize_topology(
|
||||
} else {
|
||||
None
|
||||
},
|
||||
zone: None,
|
||||
});
|
||||
} else {
|
||||
// A burned id with no member entry at all (a gap that should not
|
||||
@ -656,6 +657,7 @@ fn synthesize_topology(
|
||||
http_addr: Some(format!("http://0.0.0.0:{}", 1u16.wrapping_add(id))),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
});
|
||||
}
|
||||
}
|
||||
@ -694,6 +696,9 @@ fn synthesize_topology(
|
||||
replication: clone_replication(&input.knobs.replication),
|
||||
wal: clone_wal(&input.knobs.wal),
|
||||
election: clone_election(&input.knobs.election),
|
||||
// Seed-join synthesizes the legacy single group (the joiner learns its
|
||||
// shard assignment from the membership log, not the synthesized file).
|
||||
shards: None,
|
||||
})
|
||||
}
|
||||
|
||||
@ -770,6 +775,7 @@ mod tests {
|
||||
http_addr: Some("127.0.0.1:9001".into()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
}],
|
||||
leader: "us-east".into(),
|
||||
write_workers: None,
|
||||
@ -781,6 +787,7 @@ mod tests {
|
||||
},
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@ -42,7 +42,7 @@ use std::sync::RwLock;
|
||||
use tidaldb::replication::shard::{RegionId, ShardId};
|
||||
use tidaldb::wal::format::{MemberEntry, MemberRole, MembershipRecord};
|
||||
|
||||
use super::topology::TopologySpec;
|
||||
use super::topology::{ResolvedShardGroup, TopologySpec};
|
||||
|
||||
/// A coherent snapshot of the roster the view derived (era 0 or a kind-4
|
||||
/// record). All the lookup maps are built from this so they can never drift.
|
||||
@ -158,35 +158,44 @@ pub struct ApplyPlan {
|
||||
}
|
||||
|
||||
impl MembershipView {
|
||||
/// Build the era-0 view from the topology positional tables — byte-for-byte
|
||||
/// today's behavior. `name_to_id` is the topology declaration-order map (the
|
||||
/// node already computed it); `peer_http`/`peer_grpc` are the
|
||||
/// `build_peer_tables` outputs. The roster is reconstructed so the view's
|
||||
/// `voter_ids`/`learner_ids`/status surfaces match the implicit topology
|
||||
/// roster (every declared region is a Voter, ids = positional).
|
||||
/// Build the era-0 view for one shard `group` from its resolved replicas.
|
||||
/// `name_to_id`/`id_to_name` stay CLUSTER-WIDE (the node's declaration-order
|
||||
/// maps) so a stale cross-shard forward still resolves a non-replica region's
|
||||
/// NAME; `peer_http`/`peer_grpc` are the group-scoped `build_group_peer_tables`
|
||||
/// outputs. The roster lists exactly the GROUP's replicas (every one a Voter,
|
||||
/// id = positional `RegionId`, gRPC from the resolved replica address, HTTP
|
||||
/// from the hosting node's spec) so `voter_ids`/`voter_count`/`live_ids` and
|
||||
/// the status surfaces agree with the (group-scoped) peer/quorum/election sets
|
||||
/// that the rest of `ShardReplica::new` built. For the legacy single group
|
||||
/// (`resolve_shard_groups` synthesizes every region as a replica, addresses
|
||||
/// verbatim) this is byte-for-byte the pre-m11p6 all-regions-Voter roster.
|
||||
#[must_use]
|
||||
pub fn era0(
|
||||
self_region: RegionId,
|
||||
group: &ResolvedShardGroup,
|
||||
topology: &TopologySpec,
|
||||
name_to_id: &HashMap<String, RegionId>,
|
||||
id_to_name: &HashMap<RegionId, String>,
|
||||
peer_http: &HashMap<RegionId, String>,
|
||||
peer_grpc: &HashMap<ShardId, String>,
|
||||
) -> Self {
|
||||
// Reconstruct the implicit era-0 roster from the topology: every region
|
||||
// is a Voter, id = its positional RegionId, addresses from the spec.
|
||||
let mut members: Vec<MemberEntry> = topology
|
||||
.regions
|
||||
// Reconstruct the era-0 roster from THIS GROUP's replicas (NOT every
|
||||
// topology region): each replica is a Voter, id = positional RegionId,
|
||||
// gRPC from the resolved replica address, HTTP from the hosting node's
|
||||
// RegionSpec (ResolvedReplica carries no HTTP — one HTTP server per node).
|
||||
let mut members: Vec<MemberEntry> = group
|
||||
.replicas
|
||||
.iter()
|
||||
.map(|spec| {
|
||||
let id = name_to_id[&spec.name].0;
|
||||
MemberEntry {
|
||||
id,
|
||||
name: spec.name.clone(),
|
||||
grpc_addr: spec.grpc_addr.clone().unwrap_or_default(),
|
||||
http_addr: spec.http_addr.clone().unwrap_or_default(),
|
||||
role: MemberRole::Voter,
|
||||
}
|
||||
.map(|r| MemberEntry {
|
||||
id: r.region.0,
|
||||
name: r.name.clone(),
|
||||
grpc_addr: r.grpc_addr.clone(),
|
||||
http_addr: topology
|
||||
.regions
|
||||
.get(usize::from(r.region.0))
|
||||
.and_then(|n| n.http_addr.clone())
|
||||
.unwrap_or_default(),
|
||||
role: MemberRole::Voter,
|
||||
})
|
||||
.collect();
|
||||
members.sort_unstable_by_key(|m| m.id);
|
||||
@ -702,7 +711,7 @@ pub fn plan_activation_reappend(roster: &Roster, term: u64) -> MembershipRecord
|
||||
#[cfg(test)]
|
||||
#[allow(clippy::unwrap_used)]
|
||||
mod tests {
|
||||
use super::super::topology::{RegionSpec, TopologySpec};
|
||||
use super::super::topology::{RegionSpec, ShardReplicaSpec, ShardSpec, TopologySpec};
|
||||
use super::*;
|
||||
|
||||
fn topo(names: &[&str]) -> TopologySpec {
|
||||
@ -717,6 +726,7 @@ mod tests {
|
||||
http_addr: Some(format!("http://{n}.svc:9501")),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
})
|
||||
.collect(),
|
||||
write_workers: None,
|
||||
@ -724,11 +734,15 @@ mod tests {
|
||||
replication: super::super::topology::ReplicationSpec::default(),
|
||||
wal: super::super::topology::WalSpec::default(),
|
||||
election: super::super::topology::ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
|
||||
fn era0_view(self_idx: u16, names: &[&str]) -> MembershipView {
|
||||
let t = topo(names);
|
||||
// Legacy single group: every region is a replica (the synthesized
|
||||
// group is byte-for-byte the pre-m11p6 all-regions roster).
|
||||
let groups = t.resolve_shard_groups().expect("legacy synthesis");
|
||||
let mut name_to_id = HashMap::new();
|
||||
let mut id_to_name = HashMap::new();
|
||||
let mut peer_http = HashMap::new();
|
||||
@ -744,6 +758,7 @@ mod tests {
|
||||
}
|
||||
MembershipView::era0(
|
||||
RegionId(self_idx),
|
||||
&groups[0],
|
||||
&t,
|
||||
&name_to_id,
|
||||
&id_to_name,
|
||||
@ -752,6 +767,72 @@ mod tests {
|
||||
)
|
||||
}
|
||||
|
||||
/// An RF < N group's era-0 roster lists EXACTLY the group's replicas, not
|
||||
/// every region — so `voter_ids`/`voter_count` and `all_regions_for_status`
|
||||
/// agree with the group-scoped peer/quorum/election sets (the FIX-A guard
|
||||
/// against an N-voter majority on an RF-replica group).
|
||||
#[test]
|
||||
fn era0_roster_scoped_to_group_replicas_for_rf_lt_n() {
|
||||
// Three regions, one shard of RF=2 (us-east + ap-south; eu-west is NOT a
|
||||
// replica of this group). name_to_id stays cluster-wide (all three).
|
||||
let mut t = topo(&["us", "eu", "ap"]);
|
||||
t.shards = Some(vec![ShardSpec {
|
||||
id: 0,
|
||||
leader: Some("us".into()),
|
||||
replicas: vec![
|
||||
ShardReplicaSpec {
|
||||
node: "us".into(),
|
||||
grpc_addr: None,
|
||||
grpc_bind: None,
|
||||
},
|
||||
ShardReplicaSpec {
|
||||
node: "ap".into(),
|
||||
grpc_addr: None,
|
||||
grpc_bind: None,
|
||||
},
|
||||
],
|
||||
}]);
|
||||
let groups = t.resolve_shard_groups().expect("resolve RF=2 group");
|
||||
let mut name_to_id = HashMap::new();
|
||||
let mut id_to_name = HashMap::new();
|
||||
for (i, n) in ["us", "eu", "ap"].iter().enumerate() {
|
||||
let rid = RegionId(u16::try_from(i).unwrap());
|
||||
name_to_id.insert((*n).to_string(), rid);
|
||||
id_to_name.insert(rid, (*n).to_string());
|
||||
}
|
||||
// self = us (region 0); the only peer in THIS group is ap (region 2).
|
||||
let mut peer_http = HashMap::new();
|
||||
let mut peer_grpc = HashMap::new();
|
||||
peer_http.insert(RegionId(2), "http://ap.svc:9501".to_string());
|
||||
peer_grpc.insert(ShardId(2), "ap.svc:9500".to_string());
|
||||
let v = MembershipView::era0(
|
||||
RegionId(0),
|
||||
&groups[0],
|
||||
&t,
|
||||
&name_to_id,
|
||||
&id_to_name,
|
||||
&peer_http,
|
||||
&peer_grpc,
|
||||
);
|
||||
// The roster is the 2 replicas, NOT all 3 regions — eu-west is absent.
|
||||
assert_eq!(
|
||||
v.roster().voter_ids(),
|
||||
vec![RegionId(0), RegionId(2)],
|
||||
"era-0 roster must be the group's RF replicas, not every region"
|
||||
);
|
||||
assert_eq!(v.roster().members.len(), 2);
|
||||
assert!(
|
||||
v.roster().members.iter().all(|m| m.name != "eu"),
|
||||
"the non-replica region must not appear in the group roster"
|
||||
);
|
||||
// all_regions_for_status reflects the group, not the cluster.
|
||||
let all = v.all_regions_for_status();
|
||||
assert_eq!(all.len(), 2);
|
||||
assert!(all.iter().all(|(_, name, _)| name != "eu"));
|
||||
// name_to_id stays cluster-wide: a stale forward to eu still resolves.
|
||||
assert_eq!(v.name_to_id("eu"), Some(RegionId(1)));
|
||||
}
|
||||
|
||||
fn member(id: u16, name: &str, role: MemberRole) -> MemberEntry {
|
||||
MemberEntry {
|
||||
id,
|
||||
|
||||
@ -7,12 +7,15 @@
|
||||
//! name ↔ [`RegionId`] mapping. Replication between regions still traverses a
|
||||
//! real [`GrpcTransport`] on loopback (not in-process channels), but a crash
|
||||
//! takes the whole "cluster" down — there is no process isolation.
|
||||
//! * **Multi-process** ([`RegionClusterState`], the m8p10 mode, selected by
|
||||
//! `--region`): this process owns exactly ONE region — one [`TidalDb`], one
|
||||
//! [`GrpcTransport`] whose server binds this region's `grpc_addr` and whose
|
||||
//! peers are every sibling region's real `grpc_addr` — and peers with sibling
|
||||
//! processes over real gRPC. Process isolation is real; quorum-ack writes and
|
||||
//! automatic failure detection are still future work.
|
||||
//! * **Multi-process** ([`ClusterNode`], the m8p10→m11p6 mode, selected by
|
||||
//! `--region`): this process is a [`ClusterNode`] hosting a
|
||||
//! `BTreeMap<ShardId, Arc<ShardReplica>>` — one [`ShardReplica`] (one
|
||||
//! [`TidalDb`] + [`GrpcTransport`] + election/commit/membership) per shard
|
||||
//! group it replicates. Entity writes hash-route to the owning group's leader
|
||||
//! (apply locally or forward); corpus reads scatter over hosted groups. With
|
||||
//! `shards:` absent it holds ONE group spanning every region — byte-for-byte
|
||||
//! the original one-region-per-process node. Process isolation is real;
|
||||
//! quorum-ack writes (m11p3) and elected failover (m11p4) are wired per group.
|
||||
//!
|
||||
//! Both modes are gated behind the same explicit operator opt-in
|
||||
//! ([`ensure_experimental_enabled`]).
|
||||
@ -26,7 +29,8 @@
|
||||
//! * [`transport`] — single-process self-loop gRPC transport wiring
|
||||
//! * [`state`] — [`ClusterState`] (single-process) + the experimental gate
|
||||
//! * [`routes`] — the single-process router + handlers
|
||||
//! * [`node`] — [`RegionClusterState`] (multi-process) + its router/handlers
|
||||
//! * [`node`] — [`ClusterNode`] (the process) hosting [`ShardReplica`] groups
|
||||
//! (multi-process) + its gateway router/handlers
|
||||
//!
|
||||
//! [`GrpcTransport`]: tidal_net::GrpcTransport
|
||||
//! [`RegionId`]: tidaldb::replication::shard::RegionId
|
||||
@ -39,7 +43,7 @@ pub(crate) mod forward;
|
||||
/// joins an existing cluster by contacting a `--seed`.
|
||||
///
|
||||
/// Public so the binary's `run_seed_join_cluster` can invoke `seed_join_boot`
|
||||
/// before `RegionClusterState::new`.
|
||||
/// before `ShardReplica::new`.
|
||||
pub mod join_boot;
|
||||
/// The membership runtime (m11p5 §3.1–§3.3): effective roster, the one fenced
|
||||
/// apply path, conf-change gates, and leader-side join/remove planning.
|
||||
@ -48,7 +52,7 @@ pub(crate) mod node;
|
||||
/// Boot-time snapshot install + swap-recovery (m11p5 §2.2–§2.7).
|
||||
///
|
||||
/// Public so the binary's `run_region_cluster` can invoke `run_boot_install`
|
||||
/// before `RegionClusterState::new`.
|
||||
/// before `ShardReplica::new`.
|
||||
pub mod reseed;
|
||||
pub(crate) mod routes;
|
||||
pub(crate) mod snapshot;
|
||||
@ -58,12 +62,12 @@ mod transport;
|
||||
|
||||
// ── Public API (preserved across the split) ─────────────────────────────────
|
||||
|
||||
pub use node::{RegionClusterState, build_region_router};
|
||||
pub use node::{ClusterNode, ShardReplica, build_region_router};
|
||||
pub use routes::build_cluster_router;
|
||||
pub use state::{ClusterMode, ClusterState, EXPERIMENTAL_CLUSTER_ENV, ensure_experimental_enabled};
|
||||
pub use topology::{
|
||||
ElectionSpec, GrpcTlsSpec, RegionSpec, ReplicationSpec, TimeoutsSpec, TopologySpec, WalSpec,
|
||||
load_topology, validate_multiproc,
|
||||
ElectionSpec, GrpcTlsSpec, RegionSpec, ReplicationSpec, ShardReplicaSpec, ShardSpec,
|
||||
TimeoutsSpec, TopologySpec, WalSpec, load_topology, validate_multiproc,
|
||||
};
|
||||
|
||||
// The OpenAPI documents (`crate::openapi`) reach the handlers/DTOs through
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@ -2,7 +2,7 @@
|
||||
//!
|
||||
//! When a node's data dir carries a durable `reseed_required` marker, it does
|
||||
//! NOT open the existing engine and serve degraded forever — instead, BEFORE
|
||||
//! constructing [`RegionClusterState`], it:
|
||||
//! constructing [`ShardReplica`], it:
|
||||
//!
|
||||
//! 1. **Recovers any interrupted prior swap** (§2.3) — this runs FIRST, before
|
||||
//! `ElectionStore` classification, because classifying first re-opens restart
|
||||
@ -25,7 +25,7 @@
|
||||
//! voting enabled (§2.2; the reseed needs a leader and the leader may need this
|
||||
//! node's vote, so blocking boot forever would deadlock the cluster). The
|
||||
//! post-open seed (§2.6) is driven by the install sentinel inside
|
||||
//! [`RegionClusterState::new`], not here.
|
||||
//! [`ShardReplica::new`], not here.
|
||||
//!
|
||||
//! # The swap protocol (§2.3, review-corrected ordering)
|
||||
//!
|
||||
@ -67,7 +67,7 @@ const COMPLETE_SENTINEL: &str = "COMPLETE";
|
||||
/// The install sentinel copied INTO staging BEFORE the swap (§2.6).
|
||||
///
|
||||
/// Its presence in the canonical data dir AFTER open tells
|
||||
/// `RegionClusterState::new` to run the post-open seed exactly once
|
||||
/// `ShardReplica::new` to run the post-open seed exactly once
|
||||
/// (crash-idempotent: a crash before the seed's sentinel-delete re-runs the same
|
||||
/// seed against the unchanged WAL).
|
||||
pub const INSTALL_PENDING_SENTINEL: &str = "reseed-install-pending";
|
||||
@ -370,7 +370,7 @@ struct Candidate {
|
||||
/// The outcome of a boot-time install attempt.
|
||||
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
|
||||
pub enum InstallOutcome {
|
||||
/// A snapshot was installed (the swap completed). `RegionClusterState::new`
|
||||
/// A snapshot was installed (the swap completed). `ShardReplica::new`
|
||||
/// will find the install sentinel and run the §2.6 post-open seed.
|
||||
Installed,
|
||||
/// The leader reported `needed=false` (the live WAL still serves our
|
||||
@ -385,7 +385,7 @@ pub enum InstallOutcome {
|
||||
|
||||
/// Run the boot-time snapshot install for a marker boot (§2.2–§2.7).
|
||||
///
|
||||
/// Returns the outcome; `RegionClusterState::new` is constructed afterward in
|
||||
/// Returns the outcome; `ShardReplica::new` is constructed afterward in
|
||||
/// every case (a fresh dir after `Installed`, the existing dir otherwise).
|
||||
///
|
||||
/// # Errors
|
||||
@ -532,7 +532,7 @@ pub fn run_boot_install_with(
|
||||
/// Returns:
|
||||
/// - `Ok(Some(snapshot_seq))` — a snapshot was streamed, verified, and swapped
|
||||
/// in; the new data dir carries the install sentinel (recording
|
||||
/// `leader_region`) so `RegionClusterState::new` runs the §2.6 post-open seed.
|
||||
/// `leader_region`) so `ShardReplica::new` runs the §2.6 post-open seed.
|
||||
/// - `Ok(None)` — the leader reported `needed=false` (the live WAL still serves
|
||||
/// `from_seqno`): nothing was installed, proceed empty (the small-cluster case
|
||||
/// where the stream serves from 1).
|
||||
@ -780,7 +780,7 @@ fn attempt_install(
|
||||
snapshot_seq,
|
||||
term = leader.term,
|
||||
"reseed boot: snapshot installed and swapped in; the post-open seed runs once \
|
||||
RegionClusterState opens the new data dir (§2.6)"
|
||||
ShardReplica opens the new data dir (§2.6)"
|
||||
);
|
||||
Ok(AttemptResult::Installed)
|
||||
}
|
||||
@ -999,7 +999,7 @@ fn copy_identity_into_staging(data_dir: &Path, staging: &Path) -> Result<()> {
|
||||
/// The install sentinel's body (m11p5 §2.6).
|
||||
///
|
||||
/// It carries the artifact's recovered seq plus the DISCOVERED leader's region.
|
||||
/// Both drive the post-open path: `RegionClusterState::new` seeds the frontier
|
||||
/// Both drive the post-open path: `ShardReplica::new` seeds the frontier
|
||||
/// and issues the post-install catch-up pull against the discovered leader's
|
||||
/// shard (NOT the boot topology leader, which for a reseeded ex-leader is the
|
||||
/// node itself).
|
||||
@ -1058,7 +1058,7 @@ fn write_complete_sentinel(staging: &Path) -> Result<()> {
|
||||
|
||||
/// Read the install sentinel's recorded body from `data_dir`, if present.
|
||||
///
|
||||
/// Called by `RegionClusterState::new` after `open_region_db` to drive the §2.6
|
||||
/// Called by `ShardReplica::new` after `open_region_db` to drive the §2.6
|
||||
/// post-open seed AND the post-install catch-up pull. `None` ⇒ this was not an
|
||||
/// install boot. A 10-byte sentinel carries the discovered leader's region; an
|
||||
/// 8-byte one (an interrupted older install) is read with `leader_region: None`
|
||||
|
||||
@ -24,7 +24,7 @@ use crate::{
|
||||
pub struct TopologySpec {
|
||||
/// All regions in declaration order. The 0-based index of a region in this
|
||||
/// list IS its `RegionId`, so every process that parses the same file agrees
|
||||
/// on the region → id mapping (see `ClusterState::new` / `RegionClusterState::new`).
|
||||
/// on the region → id mapping (see `ClusterState::new` / `ShardReplica::new`).
|
||||
pub regions: Vec<RegionSpec>,
|
||||
/// Name of the region that initially leads (must name a declared region).
|
||||
pub leader: String,
|
||||
@ -56,6 +56,77 @@ pub struct TopologySpec {
|
||||
/// leader lease, auto-election on).
|
||||
#[serde(default)]
|
||||
pub election: ElectionSpec,
|
||||
/// Optional shard-group declaration (m11p6). When present, the keyspace is
|
||||
/// hash-split into `shards.len()` groups, each a replication group at RF =
|
||||
/// `replicas.len()` with its own WAL/relay/commit-index and its own elected
|
||||
/// leader; a node hosts one [`ShardReplica`](super::node) per group it
|
||||
/// appears in. When ABSENT, the legacy "1 shard × RF=all-regions" group is
|
||||
/// synthesized (id 0, every region a replica with its address verbatim,
|
||||
/// `leader` = [`Self::leader`]) — byte-for-byte today's single-log cluster.
|
||||
/// See [`Self::resolve_shard_groups`].
|
||||
#[serde(default)]
|
||||
pub shards: Option<Vec<ShardSpec>>,
|
||||
}
|
||||
|
||||
/// One shard group's declaration (the optional `shards:` YAML block, m11p6).
|
||||
#[derive(Debug, Deserialize)]
|
||||
pub struct ShardSpec {
|
||||
/// Dense, unique group id in `[0, shards.len())`. The gateway hash-routes an
|
||||
/// entity to `ShardId(id)` via [`tidaldb::replication::ShardRouter`]; the id
|
||||
/// also names the group's data subdir and gRPC port offset.
|
||||
pub id: u16,
|
||||
/// Term-0 / preferred leader (must name a declared region that is also one of
|
||||
/// this group's `replicas`). Defaults to the first replica when omitted.
|
||||
#[serde(default)]
|
||||
pub leader: Option<String>,
|
||||
/// The RF replica nodes hosting this group.
|
||||
pub replicas: Vec<ShardReplicaSpec>,
|
||||
}
|
||||
|
||||
/// One replica placement within a [`ShardSpec`] (m11p6).
|
||||
#[derive(Debug, Deserialize)]
|
||||
pub struct ShardReplicaSpec {
|
||||
/// The hosting node — must name a declared region.
|
||||
pub node: String,
|
||||
/// This replica's advertised gRPC address (`host:port`, may be DNS). When
|
||||
/// omitted, derived as the node's `grpc_addr` with `id` added to its port
|
||||
/// (so id 0 = the node's address verbatim — the legacy port).
|
||||
#[serde(default)]
|
||||
pub grpc_addr: Option<String>,
|
||||
/// This replica's local gRPC bind (`SocketAddr`). When omitted, derived from
|
||||
/// the node's `grpc_bind` (or `grpc_addr`) with `id` added to its port.
|
||||
#[serde(default)]
|
||||
pub grpc_bind: Option<String>,
|
||||
}
|
||||
|
||||
/// A fully-resolved shard group: ids resolved to [`RegionId`]s, addresses
|
||||
/// resolved (explicit or port-derived). Produced by
|
||||
/// [`TopologySpec::resolve_shard_groups`] — the single seam that replaces the
|
||||
/// 1:1 [`shard_of_region`] assumption.
|
||||
#[derive(Debug, Clone)]
|
||||
pub struct ResolvedShardGroup {
|
||||
/// The data-shard group id (gateway hash output; map key; subdir/port name).
|
||||
pub shard: ShardId,
|
||||
/// The group's term-0 / preferred leader node.
|
||||
pub leader: RegionId,
|
||||
/// `leader`'s region name (display).
|
||||
pub leader_name: String,
|
||||
/// The group's RF replica placements.
|
||||
pub replicas: Vec<ResolvedReplica>,
|
||||
}
|
||||
|
||||
/// One resolved replica placement within a [`ResolvedShardGroup`] (m11p6).
|
||||
#[derive(Debug, Clone)]
|
||||
pub struct ResolvedReplica {
|
||||
/// The hosting node's region id.
|
||||
pub region: RegionId,
|
||||
/// The hosting node's region name.
|
||||
pub name: String,
|
||||
/// The replica's advertised gRPC address (explicit or port-derived).
|
||||
pub grpc_addr: String,
|
||||
/// The replica's local gRPC bind, when explicit/derived (else `None` ⇒ the
|
||||
/// node derives it from `grpc_addr` at boot per `resolve_grpc_bind_addr`).
|
||||
pub grpc_bind: Option<String>,
|
||||
}
|
||||
|
||||
/// Election / failure-detector tuning (the optional `election:` YAML block,
|
||||
@ -257,6 +328,13 @@ pub struct RegionSpec {
|
||||
/// metrics listener before m11p1 and the roadmap calls that out as a gap.
|
||||
#[serde(default)]
|
||||
pub metrics_addr: Option<String>,
|
||||
/// Optional placement label (m11p6), e.g. an availability zone (`"az-a"`),
|
||||
/// demoting "region" from a shard identity to a placement hint. **Parsed and
|
||||
/// accepted but NOT yet honored:** the read-affinity router that would prefer
|
||||
/// a replica in the reader's zone is a later layer (L4/p8), so setting this
|
||||
/// today has no runtime effect. Carried now so the schema is forward-stable.
|
||||
#[serde(default)]
|
||||
pub zone: Option<String>,
|
||||
}
|
||||
|
||||
/// TLS material for one region's gRPC transport (the optional `grpc_tls:`
|
||||
@ -440,13 +518,14 @@ fn validate_spec_values(spec: &TopologySpec) -> Result<()> {
|
||||
)));
|
||||
}
|
||||
}
|
||||
validate_shards(spec)?;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
impl TopologySpec {
|
||||
/// Resolve the cluster-write-pool config from `write_workers`.
|
||||
///
|
||||
/// Shared by both `ClusterState::new` and `RegionClusterState::new` so the
|
||||
/// Shared by both `ClusterState::new` and `ShardReplica::new` so the
|
||||
/// two cluster modes size their write pool the same way from the same field.
|
||||
#[must_use]
|
||||
pub(crate) fn write_pool_config(&self) -> ClusterWritePoolConfig {
|
||||
@ -493,6 +572,236 @@ impl TopologySpec {
|
||||
.find(|r| r.name == region_name)
|
||||
.and_then(|r| r.metrics_addr.clone())
|
||||
}
|
||||
|
||||
/// The number of shard groups (m11p6): `shards.len()`, or 1 when `shards:`
|
||||
/// is absent (the legacy single replicated log). Drives the gateway's
|
||||
/// [`tidaldb::replication::ShardRouter`] (`Hash(n)` for `n > 1`, `Single`
|
||||
/// for `n == 1`).
|
||||
#[must_use]
|
||||
pub fn shard_count(&self) -> usize {
|
||||
self.shards.as_ref().map_or(1, Vec::len)
|
||||
}
|
||||
|
||||
/// Resolve the shard-group assignment (m11p6) — the single seam that
|
||||
/// generalizes the 1:1 [`shard_of_region`] mapping. With `shards:` present,
|
||||
/// resolves each replica's node name to a [`RegionId`] and its gRPC address
|
||||
/// (explicit, or the node's address with the shard id added to its port).
|
||||
/// Absent ⇒ one group spanning every region (id 0, the topology `leader`,
|
||||
/// addresses verbatim) — byte-for-byte today's single-log cluster (id 0 ⇒
|
||||
/// port offset 0 ⇒ the declared address).
|
||||
///
|
||||
/// # Errors
|
||||
///
|
||||
/// Returns [`ServerError::SchemaConfig`] if a replica/leader names an
|
||||
/// undeclared region, a group has no replicas, ids are not dense/unique, a
|
||||
/// derived port overflows, or two shards on one node collide on a gRPC
|
||||
/// address.
|
||||
#[allow(clippy::too_many_lines)]
|
||||
pub fn resolve_shard_groups(&self) -> Result<Vec<ResolvedShardGroup>> {
|
||||
// The positional region→id map, overflow-guarded ONCE here (a RegionId is
|
||||
// a u16). Every id below reuses this map rather than re-deriving the cast
|
||||
// with a different overflow policy — one source of truth for the seam
|
||||
// every durable id depends on.
|
||||
let mut name_to_id: std::collections::HashMap<&str, RegionId> =
|
||||
std::collections::HashMap::with_capacity(self.regions.len());
|
||||
for (i, r) in self.regions.iter().enumerate() {
|
||||
let id = RegionId(u16::try_from(i).map_err(|_| {
|
||||
ServerError::SchemaConfig("topology declares more than 65535 regions".into())
|
||||
})?);
|
||||
name_to_id.insert(r.name.as_str(), id);
|
||||
}
|
||||
let region_of = |name: &str| -> Result<RegionId> {
|
||||
name_to_id.get(name).copied().ok_or_else(|| {
|
||||
ServerError::SchemaConfig(format!(
|
||||
"shards: replica/leader names undeclared region '{name}'"
|
||||
))
|
||||
})
|
||||
};
|
||||
|
||||
let Some(shards) = self.shards.as_ref() else {
|
||||
// Legacy synthesis: one group, RF = all regions, addresses verbatim.
|
||||
let leader = region_of(&self.leader)?;
|
||||
let replicas = self
|
||||
.regions
|
||||
.iter()
|
||||
.map(|r| ResolvedReplica {
|
||||
region: name_to_id[r.name.as_str()],
|
||||
name: r.name.clone(),
|
||||
grpc_addr: r.grpc_addr.clone().unwrap_or_default(),
|
||||
grpc_bind: r.grpc_bind.clone(),
|
||||
})
|
||||
.collect();
|
||||
return Ok(vec![ResolvedShardGroup {
|
||||
shard: ShardId(0),
|
||||
leader,
|
||||
leader_name: self.leader.clone(),
|
||||
replicas,
|
||||
}]);
|
||||
};
|
||||
|
||||
let mut groups = Vec::with_capacity(shards.len());
|
||||
for spec in shards {
|
||||
if spec.replicas.is_empty() {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard {} declares no replicas",
|
||||
spec.id
|
||||
)));
|
||||
}
|
||||
let mut replicas = Vec::with_capacity(spec.replicas.len());
|
||||
let mut seen_nodes = std::collections::HashSet::new();
|
||||
for r in &spec.replicas {
|
||||
let region = region_of(&r.node)?;
|
||||
if !seen_nodes.insert(region) {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard {} lists node '{}' twice",
|
||||
spec.id, r.node
|
||||
)));
|
||||
}
|
||||
let node = &self.regions[usize::from(region.0)];
|
||||
let grpc_addr = match &r.grpc_addr {
|
||||
Some(a) => a.clone(),
|
||||
None => offset_host_port(
|
||||
node.grpc_addr.as_deref().ok_or_else(|| {
|
||||
ServerError::SchemaConfig(format!(
|
||||
"shard {} replica '{}' has no grpc_addr and node declares none \
|
||||
to derive from",
|
||||
spec.id, r.node
|
||||
))
|
||||
})?,
|
||||
spec.id,
|
||||
)?,
|
||||
};
|
||||
let grpc_bind = match &r.grpc_bind {
|
||||
Some(b) => Some(b.clone()),
|
||||
None => node
|
||||
.grpc_bind
|
||||
.as_deref()
|
||||
.map(|b| offset_host_port(b, spec.id))
|
||||
.transpose()?,
|
||||
};
|
||||
replicas.push(ResolvedReplica {
|
||||
region,
|
||||
name: r.node.clone(),
|
||||
grpc_addr,
|
||||
grpc_bind,
|
||||
});
|
||||
}
|
||||
let leader_name = spec
|
||||
.leader
|
||||
.clone()
|
||||
.unwrap_or_else(|| replicas[0].name.clone());
|
||||
let leader = region_of(&leader_name)?;
|
||||
if !replicas.iter().any(|r| r.region == leader) {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard {} leader '{leader_name}' is not one of its replicas",
|
||||
spec.id
|
||||
)));
|
||||
}
|
||||
groups.push(ResolvedShardGroup {
|
||||
shard: ShardId(spec.id),
|
||||
leader,
|
||||
leader_name,
|
||||
replicas,
|
||||
});
|
||||
}
|
||||
validate_no_port_collisions(&groups)?;
|
||||
Ok(groups)
|
||||
}
|
||||
}
|
||||
|
||||
/// Add `offset` to the port of a `host:port` address; `offset == 0` returns the
|
||||
/// input verbatim (the legacy-shard guarantee — no reformatting).
|
||||
fn offset_host_port(addr: &str, offset: u16) -> Result<String> {
|
||||
if offset == 0 {
|
||||
return Ok(addr.to_string());
|
||||
}
|
||||
let (host, port) = addr.rsplit_once(':').ok_or_else(|| {
|
||||
ServerError::SchemaConfig(format!(
|
||||
"cannot derive a shard port from '{addr}' (no ':port' to offset)"
|
||||
))
|
||||
})?;
|
||||
let base: u16 = port.parse().map_err(|_| {
|
||||
ServerError::SchemaConfig(format!("'{addr}' port is not a number to offset"))
|
||||
})?;
|
||||
let derived = base.checked_add(offset).ok_or_else(|| {
|
||||
ServerError::SchemaConfig(format!(
|
||||
"deriving shard port from '{addr}' + {offset} overflows u16 — set an explicit \
|
||||
grpc_addr for this replica"
|
||||
))
|
||||
})?;
|
||||
Ok(format!("{host}:{derived}"))
|
||||
}
|
||||
|
||||
/// Reject two shard replicas on the SAME node that resolve to the same gRPC
|
||||
/// address (a derived/explicit port clash would make two of the node's shard
|
||||
/// transports fight for one socket).
|
||||
fn validate_no_port_collisions(groups: &[ResolvedShardGroup]) -> Result<()> {
|
||||
let mut seen: std::collections::HashMap<(RegionId, &str), ShardId> =
|
||||
std::collections::HashMap::new();
|
||||
for g in groups {
|
||||
for r in &g.replicas {
|
||||
if let Some(prev) = seen.insert((r.region, r.grpc_addr.as_str()), g.shard) {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"node '{}' resolves shards {} and {} to the same grpc_addr '{}' — set \
|
||||
distinct ports (derived ports add the shard id to the node port)",
|
||||
r.name, prev.0, g.shard.0, r.grpc_addr
|
||||
)));
|
||||
}
|
||||
}
|
||||
}
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Structural validation of the optional `shards:` block (mode-independent;
|
||||
/// address resolution is deferred to [`TopologySpec::resolve_shard_groups`]).
|
||||
/// Checks dense unique ids in `[0, n)`, non-empty replicas, and that named
|
||||
/// nodes/leaders exist — so a typo fails at load, not at routing time.
|
||||
fn validate_shards(spec: &TopologySpec) -> Result<()> {
|
||||
let Some(shards) = spec.shards.as_ref() else {
|
||||
return Ok(());
|
||||
};
|
||||
if shards.is_empty() {
|
||||
return Err(ServerError::SchemaConfig(
|
||||
"shards: declared but empty (omit the key for the legacy single group)".into(),
|
||||
));
|
||||
}
|
||||
let names: std::collections::HashSet<&str> =
|
||||
spec.regions.iter().map(|r| r.name.as_str()).collect();
|
||||
let mut ids: Vec<u16> = shards.iter().map(|s| s.id).collect();
|
||||
ids.sort_unstable();
|
||||
for (want, got) in ids.iter().enumerate() {
|
||||
if usize::from(*got) != want {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard ids must be dense and unique in [0, {}); got {ids:?}",
|
||||
shards.len()
|
||||
)));
|
||||
}
|
||||
}
|
||||
for s in shards {
|
||||
if s.replicas.is_empty() {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard {} declares no replicas (RF must be >= 1)",
|
||||
s.id
|
||||
)));
|
||||
}
|
||||
for r in &s.replicas {
|
||||
if !names.contains(r.node.as_str()) {
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard {} replica node '{}' is not a declared region",
|
||||
s.id, r.node
|
||||
)));
|
||||
}
|
||||
}
|
||||
if let Some(l) = s.leader.as_deref()
|
||||
&& !s.replicas.iter().any(|r| r.node == l)
|
||||
{
|
||||
return Err(ServerError::SchemaConfig(format!(
|
||||
"shard {} leader '{l}' must be one of its replicas",
|
||||
s.id
|
||||
)));
|
||||
}
|
||||
}
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Validate a topology for **multi-process** (`--region`) mode.
|
||||
@ -628,6 +937,7 @@ mod tests {
|
||||
http_addr: http.map(str::to_string),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
}
|
||||
}
|
||||
|
||||
@ -643,6 +953,7 @@ mod tests {
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
|
||||
@ -837,4 +1148,120 @@ leader: us-east
|
||||
assert_eq!(tls.ca_cert, PathBuf::from("/etc/tidal/ca.pem"));
|
||||
assert!(tls.client_cert.is_none(), "mTLS pair defaults to off");
|
||||
}
|
||||
|
||||
// ── m11p6 shard-group schema ─────────────────────────────────────────────
|
||||
|
||||
#[test]
|
||||
fn absent_shards_synthesizes_one_group_verbatim() {
|
||||
// The legacy guarantee: no `shards:` ⇒ one group, RF = all regions,
|
||||
// leader = topology.leader, addresses byte-for-byte (offset 0).
|
||||
let t = full_topology();
|
||||
assert_eq!(t.shard_count(), 1);
|
||||
let groups = t.resolve_shard_groups().expect("legacy synthesis");
|
||||
assert_eq!(groups.len(), 1);
|
||||
let g = &groups[0];
|
||||
assert_eq!(g.shard, ShardId(0));
|
||||
assert_eq!(g.leader, RegionId(0));
|
||||
assert_eq!(g.leader_name, "us-east");
|
||||
assert_eq!(g.replicas.len(), 2);
|
||||
assert_eq!(g.replicas[0].region, RegionId(0));
|
||||
assert_eq!(g.replicas[0].grpc_addr, "127.0.0.1:9601"); // verbatim
|
||||
assert_eq!(g.replicas[1].grpc_addr, "127.0.0.1:9602"); // verbatim
|
||||
}
|
||||
|
||||
fn three_node_yaml(shards: &str) -> TopologySpec {
|
||||
let yaml = format!(
|
||||
"regions:\n - {{ name: us-east, grpc_addr: \"127.0.0.1:9601\", \
|
||||
http_addr: \"127.0.0.1:9501\", zone: az-a }}\n - {{ name: eu-west, \
|
||||
grpc_addr: \"127.0.0.1:9611\", http_addr: \"127.0.0.1:9511\", zone: az-b }}\n \
|
||||
- {{ name: ap-south, grpc_addr: \"127.0.0.1:9621\", http_addr: \"127.0.0.1:9521\" }}\n\
|
||||
leader: us-east\n{shards}"
|
||||
);
|
||||
load_topology_from_str(&yaml).expect("topology parses + validates")
|
||||
}
|
||||
|
||||
fn load_topology_from_str(raw: &str) -> Result<TopologySpec> {
|
||||
let spec: TopologySpec = serde_yml::from_str(raw)
|
||||
.map_err(|e| ServerError::SchemaConfig(format!("parse: {e}")))?;
|
||||
validate_spec_values(&spec)?;
|
||||
Ok(spec)
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn explicit_shards_resolve_with_derived_and_explicit_ports() {
|
||||
let t = three_node_yaml(
|
||||
"shards:\n - id: 0\n leader: us-east\n replicas: [{node: us-east}, \
|
||||
{node: eu-west}, {node: ap-south}]\n - id: 1\n leader: eu-west\n \
|
||||
replicas: [{node: us-east, grpc_addr: \"127.0.0.1:7000\"}, {node: eu-west}, \
|
||||
{node: ap-south}]\n",
|
||||
);
|
||||
assert_eq!(t.shard_count(), 2);
|
||||
// `zone:` parses onto the RegionSpec (accepted, not yet honored).
|
||||
assert_eq!(t.regions[0].zone.as_deref(), Some("az-a"));
|
||||
let groups = t.resolve_shard_groups().expect("resolve");
|
||||
assert_eq!(groups.len(), 2);
|
||||
// Shard 0: derived = node port + 0 (verbatim).
|
||||
assert_eq!(groups[0].shard, ShardId(0));
|
||||
assert_eq!(groups[0].leader, RegionId(0));
|
||||
assert_eq!(groups[0].replicas[0].grpc_addr, "127.0.0.1:9601");
|
||||
assert_eq!(groups[0].replicas[1].grpc_addr, "127.0.0.1:9611");
|
||||
// Shard 1: us-east explicit 7000; eu-west/ap-south derived = port + 1.
|
||||
assert_eq!(groups[1].shard, ShardId(1));
|
||||
assert_eq!(groups[1].leader, RegionId(1));
|
||||
assert_eq!(groups[1].replicas[0].grpc_addr, "127.0.0.1:7000");
|
||||
assert_eq!(groups[1].replicas[1].grpc_addr, "127.0.0.1:9612");
|
||||
assert_eq!(groups[1].replicas[2].grpc_addr, "127.0.0.1:9622");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn shards_reject_non_dense_ids() {
|
||||
let yaml = "shards:\n - {id: 0, replicas: [{node: us-east}]}\n \
|
||||
- {id: 2, replicas: [{node: eu-west}]}\n";
|
||||
let err = load_topology_from_str(&format!(
|
||||
"regions:\n - {{name: us-east, grpc_addr: \"1:1\", http_addr: \"1:1\"}}\n \
|
||||
- {{name: eu-west, grpc_addr: \"1:2\", http_addr: \"1:2\"}}\nleader: us-east\n{yaml}"
|
||||
))
|
||||
.expect_err("non-dense ids rejected");
|
||||
assert!(err.to_string().contains("dense"), "got {err}");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn shards_reject_unknown_node_and_bad_leader() {
|
||||
let bad_node = load_topology_from_str(
|
||||
"regions:\n - {name: us-east, grpc_addr: \"1:1\", http_addr: \"1:1\"}\nleader: us-east\n\
|
||||
shards:\n - {id: 0, replicas: [{node: ghost}]}\n",
|
||||
)
|
||||
.expect_err("unknown node");
|
||||
assert!(bad_node.to_string().contains("ghost"), "got {bad_node}");
|
||||
|
||||
let bad_leader = load_topology_from_str(
|
||||
"regions:\n - {name: us-east, grpc_addr: \"1:1\", http_addr: \"1:1\"}\n \
|
||||
- {name: eu-west, grpc_addr: \"1:2\", http_addr: \"1:2\"}\nleader: us-east\n\
|
||||
shards:\n - {id: 0, leader: eu-west, replicas: [{node: us-east}]}\n",
|
||||
)
|
||||
.expect_err("leader not a replica");
|
||||
assert!(
|
||||
bad_leader
|
||||
.to_string()
|
||||
.contains("must be one of its replicas"),
|
||||
"got {bad_leader}"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn shards_reject_same_node_port_collision() {
|
||||
// us-east hosts shards 0 and 1 but both resolve to the same explicit addr.
|
||||
let t = load_topology_from_str(
|
||||
"regions:\n - {name: us-east, grpc_addr: \"127.0.0.1:9601\", http_addr: \"1:1\"}\n \
|
||||
- {name: eu-west, grpc_addr: \"127.0.0.1:9611\", http_addr: \"1:2\"}\nleader: us-east\n\
|
||||
shards:\n - {id: 0, replicas: [{node: us-east, grpc_addr: \"127.0.0.1:5000\"}, \
|
||||
{node: eu-west}]}\n - {id: 1, leader: eu-west, replicas: [{node: us-east, \
|
||||
grpc_addr: \"127.0.0.1:5000\"}, {node: eu-west}]}\n",
|
||||
)
|
||||
.expect("structural validation passes");
|
||||
let err = t
|
||||
.resolve_shard_groups()
|
||||
.expect_err("port collision rejected");
|
||||
assert!(err.to_string().contains("same grpc_addr"), "got {err}");
|
||||
}
|
||||
}
|
||||
|
||||
@ -3,7 +3,7 @@ use std::{net::SocketAddr, path::PathBuf, sync::Arc};
|
||||
use clap::{Args, Parser, Subcommand};
|
||||
use tidal_server::{
|
||||
cluster::{
|
||||
ClusterMode, ClusterState, RegionClusterState, build_cluster_router, build_region_router,
|
||||
ClusterMode, ClusterNode, ClusterState, build_cluster_router, build_region_router,
|
||||
ensure_experimental_enabled, join_boot, load_topology,
|
||||
},
|
||||
config::{CONFIG_DIR_SCHEMA_FILE, CONFIG_DIR_TOPOLOGY_FILE, load_schema, resolve_config_path},
|
||||
@ -270,16 +270,16 @@ async fn run_region_cluster(args: ClusterArgs, region: String) -> Result<()> {
|
||||
|
||||
let data_dir = args.data_dir.clone();
|
||||
let api_key = read_api_key();
|
||||
// `RegionClusterState::new` builds the GrpcTransport via `GrpcTransport::new`,
|
||||
// `ShardReplica::new` builds the GrpcTransport via `GrpcTransport::new`,
|
||||
// which blocks on its own tokio runtime — must run off this reactor. The
|
||||
// m11p5 boot-time snapshot install (§2.2–§2.7) ALSO blocks (a leader-
|
||||
// discovery loop + a synchronous snapshot fetch), so both run on the same
|
||||
// dedicated thread, install FIRST: swap-recovery and the snapshot swap must
|
||||
// complete BEFORE `RegionClusterState::new` opens the engine (a marker boot
|
||||
// complete BEFORE `ShardReplica::new` opens the engine (a marker boot
|
||||
// opens the NEW data dir; a fallback opens the existing one degraded).
|
||||
let state = std::thread::Builder::new()
|
||||
.name("region-build".into())
|
||||
.spawn(move || -> Result<RegionClusterState> {
|
||||
.spawn(move || -> Result<ClusterNode> {
|
||||
// §2.2: the install pre-open step runs ONLY when a reseed marker is
|
||||
// present (plain topology boots never contact a peer before serving).
|
||||
// `run_boot_install` returns immediately (NotNeeded) when no marker
|
||||
@ -289,7 +289,10 @@ async fn run_region_cluster(args: ClusterArgs, region: String) -> Result<()> {
|
||||
tidal_server::cluster::reseed::run_boot_install(&topology, ®ion, dir)?;
|
||||
tracing::info!(?outcome, region = %region, "boot-time reseed install evaluated");
|
||||
}
|
||||
RegionClusterState::new(
|
||||
// m11p6: `ClusterNode::new` resolves the shard-group assignment and
|
||||
// opens one `ShardReplica` per group this node hosts (S=1 is the
|
||||
// legacy group spanning every region, the node data dir verbatim).
|
||||
ClusterNode::new(
|
||||
&topology,
|
||||
®ion,
|
||||
schema,
|
||||
@ -307,7 +310,7 @@ async fn run_region_cluster(args: ClusterArgs, region: String) -> Result<()> {
|
||||
|
||||
/// Seed-join boot (m11p5 §3.4–§3.6): this node is NOT declared in the local
|
||||
/// topology file. It learns its roster, assigned id, and term from a `--seed`,
|
||||
/// installs a snapshot when needed, then constructs `RegionClusterState` from
|
||||
/// installs a snapshot when needed, then constructs `ShardReplica` from
|
||||
/// the LEARNED roster (not the topology's `regions:` list).
|
||||
///
|
||||
/// §3.5 knob-source rule: the local topology/config file is STILL required for
|
||||
@ -350,7 +353,7 @@ async fn run_seed_join_cluster(args: ClusterArgs, region: String) -> Result<()>
|
||||
// We do NOT call `validate_multiproc` on this file's `regions:` (it would
|
||||
// demand the joiner be declared there); the runtime roster is the join
|
||||
// response, and the synthesized topology IS validated by `validate_multiproc`
|
||||
// inside `RegionClusterState::new`.
|
||||
// inside `ShardReplica::new`.
|
||||
let knobs = load_topology(topology_path.as_deref())?;
|
||||
|
||||
let advertise_grpc = args.advertise_grpc.clone().ok_or_else(|| {
|
||||
@ -396,12 +399,12 @@ async fn run_seed_join_cluster(args: ClusterArgs, region: String) -> Result<()>
|
||||
let metrics = args.metrics.clone();
|
||||
let seeds = args.seed.clone();
|
||||
// The §2.7 join loop + snapshot install BLOCK (leader discovery, a
|
||||
// synchronous JoinCluster, a snapshot fetch), and `RegionClusterState::new`
|
||||
// synchronous JoinCluster, a snapshot fetch), and `ShardReplica::new`
|
||||
// builds the GrpcTransport on its own runtime — both must run off this
|
||||
// reactor, on the same dedicated thread, install FIRST.
|
||||
let state = std::thread::Builder::new()
|
||||
.name("region-seed-join-build".into())
|
||||
.spawn(move || -> Result<RegionClusterState> {
|
||||
.spawn(move || -> Result<ClusterNode> {
|
||||
let join_boot::SeedJoinBoot {
|
||||
topology: synth,
|
||||
region: synth_region,
|
||||
@ -415,7 +418,9 @@ async fn run_seed_join_cluster(args: ClusterArgs, region: String) -> Result<()>
|
||||
data_dir: &data_dir,
|
||||
api_key: api_key_for_join,
|
||||
})?;
|
||||
RegionClusterState::new(
|
||||
// m11p6: a seed-join synthesizes the legacy single group (the joiner
|
||||
// learns its shard assignment from the membership log, not the file).
|
||||
ClusterNode::new(
|
||||
&synth,
|
||||
&synth_region,
|
||||
schema,
|
||||
@ -475,10 +480,11 @@ impl ServeState for ClusterState {
|
||||
}
|
||||
}
|
||||
|
||||
impl ServeState for RegionClusterState {
|
||||
impl ServeState for ClusterNode {
|
||||
const WHAT: &'static str = "region";
|
||||
fn started(self: &Arc<Self>) {
|
||||
self.start_election_driver();
|
||||
// m11p6: start every hosted shard group's election driver.
|
||||
self.start_all_elections();
|
||||
}
|
||||
fn set_shutting_down(&self) {
|
||||
Self::set_shutting_down(self); // the inherent method, as above
|
||||
|
||||
@ -30,6 +30,7 @@ fn three_region_topology() -> TopologySpec {
|
||||
http_addr: None,
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
})
|
||||
.into_iter()
|
||||
.collect(),
|
||||
@ -39,6 +40,7 @@ fn three_region_topology() -> TopologySpec {
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
//! m8p10 in-process multi-process-cluster tests.
|
||||
//!
|
||||
//! Each test builds TWO `RegionClusterState`s in ONE test process — distinct
|
||||
//! Each test builds TWO `ShardReplica`s in ONE test process — distinct
|
||||
//! topologies pointing at each other's REAL loopback gRPC addresses — and drives
|
||||
//! them over real HTTP (axum on loopback) + real `GrpcTransport` replication.
|
||||
//! Unlike `cluster_e2e.rs` (tier-3, spawns OS processes), these run in the
|
||||
@ -24,10 +24,11 @@ use std::{
|
||||
};
|
||||
|
||||
use tidal_server::cluster::{
|
||||
ElectionSpec, RegionClusterState, RegionSpec, ReplicationSpec, TimeoutsSpec, TopologySpec,
|
||||
WalSpec, build_region_router,
|
||||
ClusterNode, ElectionSpec, RegionSpec, ReplicationSpec, ShardReplicaSpec, ShardSpec,
|
||||
TimeoutsSpec, TopologySpec, WalSpec, build_region_router,
|
||||
};
|
||||
use tidaldb::schema::{DecaySpec, EntityKind, Schema, SchemaBuilder, Window};
|
||||
use tidaldb::replication::shard::ShardRouter;
|
||||
use tidaldb::schema::{DecaySpec, EntityId, EntityKind, Schema, SchemaBuilder, Window};
|
||||
use tidaldb::wal::format::{MemberEntry, MemberRole, MembershipRecord};
|
||||
|
||||
/// A single-`view`-signal schema with a `hide` hard-negative signal (so the
|
||||
@ -95,6 +96,7 @@ impl Pair {
|
||||
http_addr: Some(self.leader_http.to_string()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
},
|
||||
RegionSpec {
|
||||
name: self.follower_name.clone(),
|
||||
@ -103,6 +105,7 @@ impl Pair {
|
||||
http_addr: Some(self.follower_http.to_string()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
},
|
||||
],
|
||||
leader: self.leader_name.clone(),
|
||||
@ -111,6 +114,7 @@ impl Pair {
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -131,15 +135,11 @@ fn region_dir() -> tempfile::TempDir {
|
||||
/// Build one region node off the reactor (GrpcTransport::new blocks on its own
|
||||
/// runtime, so it must run on a plain thread). `dir` is the node's data dir;
|
||||
/// see [`region_dir`] for the declaration-order contract.
|
||||
fn build_region(
|
||||
topology: TopologySpec,
|
||||
region: &str,
|
||||
dir: &tempfile::TempDir,
|
||||
) -> RegionClusterState {
|
||||
fn build_region(topology: TopologySpec, region: &str, dir: &tempfile::TempDir) -> ClusterNode {
|
||||
let region = region.to_string();
|
||||
let data_dir = dir.path().to_path_buf();
|
||||
std::thread::spawn(move || {
|
||||
RegionClusterState::new(
|
||||
ClusterNode::new(
|
||||
&topology,
|
||||
®ion,
|
||||
region_schema(),
|
||||
@ -310,6 +310,330 @@ fn region_node_replicates_over_grpc() {
|
||||
rt.shutdown_timeout(Duration::from_secs(2));
|
||||
}
|
||||
|
||||
/// m11p6 sharding × replication (in-process): 2 nodes × 2 shards × RF=2, with
|
||||
/// shard 0 led by node A and shard 1 led by node B (balanced leaders). Every
|
||||
/// node hosts a replica of BOTH groups. A write routes by entity hash to its
|
||||
/// shard's leader (A applies its shard-0 writes locally and FORWARDS its
|
||||
/// shard-1 writes to B, and vice versa — the unified path, no /sharded needed);
|
||||
/// each shard replicates to its follower on the other node; and a /feed read on
|
||||
/// either node scatters over both local groups and merges, returning items from
|
||||
/// BOTH shards. This is the exit-gate shape at 2×2 — proof the
|
||||
/// replicated-XOR-sharded split is gone.
|
||||
#[test]
|
||||
fn region_sharded_writes_route_per_shard_and_reads_scatter() {
|
||||
// 2 nodes, 4 gRPC ports (one per (node, shard)), 2 HTTP ports.
|
||||
let names = ["node-a".to_string(), "node-b".to_string()];
|
||||
let http = [free_addr(), free_addr()];
|
||||
// grpc[node][shard]
|
||||
let grpc = [[free_addr(), free_addr()], [free_addr(), free_addr()]];
|
||||
|
||||
let topology = || -> TopologySpec {
|
||||
let regions = (0..2)
|
||||
.map(|i| RegionSpec {
|
||||
name: names[i].clone(),
|
||||
// The node's base grpc_addr is shard 0's; shard 1 sets its own.
|
||||
grpc_addr: Some(grpc[i][0].to_string()),
|
||||
grpc_bind: None,
|
||||
http_addr: Some(http[i].to_string()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
})
|
||||
.collect();
|
||||
let shard = |id: u16, leader: usize| ShardSpec {
|
||||
id,
|
||||
leader: Some(names[leader].clone()),
|
||||
replicas: (0..2)
|
||||
.map(|n| ShardReplicaSpec {
|
||||
node: names[n].clone(),
|
||||
grpc_addr: Some(grpc[n][id as usize].to_string()),
|
||||
grpc_bind: None,
|
||||
})
|
||||
.collect(),
|
||||
};
|
||||
TopologySpec {
|
||||
regions,
|
||||
leader: names[0].clone(), // legacy field unused once shards: is set
|
||||
write_workers: None,
|
||||
timeouts: TimeoutsSpec::default(),
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: Some(vec![shard(0, 0), shard(1, 1)]),
|
||||
}
|
||||
};
|
||||
|
||||
let dirs: Vec<tempfile::TempDir> = (0..2).map(|_| region_dir()).collect();
|
||||
let node_a = Arc::new(build_region(topology(), &names[0], &dirs[0]));
|
||||
let node_b = Arc::new(build_region(topology(), &names[1], &dirs[1]));
|
||||
|
||||
let rt = tokio::runtime::Builder::new_multi_thread()
|
||||
.worker_threads(4)
|
||||
.enable_all()
|
||||
.build()
|
||||
.unwrap();
|
||||
serve(&rt, build_region_router(Arc::clone(&node_a), None), http[0]);
|
||||
serve(&rt, build_region_router(Arc::clone(&node_b), None), http[1]);
|
||||
|
||||
let client = reqwest::blocking::Client::new();
|
||||
let base_a = format!("http://{}", http[0]);
|
||||
let base_b = format!("http://{}", http[1]);
|
||||
|
||||
// Partition entity ids by their shard so we can assert BOTH shards receive
|
||||
// writes (the same FNV-1a router the gateway uses).
|
||||
let router = ShardRouter::hash(2).unwrap();
|
||||
let mut shard0_ids = Vec::new();
|
||||
let mut shard1_ids = Vec::new();
|
||||
for i in 1..=24u64 {
|
||||
match router.route(EntityId::new(i)).0 {
|
||||
0 => shard0_ids.push(i),
|
||||
_ => shard1_ids.push(i),
|
||||
}
|
||||
}
|
||||
assert!(
|
||||
!shard0_ids.is_empty() && !shard1_ids.is_empty(),
|
||||
"the hash must spread 24 ids across both shards (got {} / {})",
|
||||
shard0_ids.len(),
|
||||
shard1_ids.len()
|
||||
);
|
||||
|
||||
// Write every item + a view signal through node A ONLY. A leads shard 0
|
||||
// (applies locally) and follows shard 1 (forwards to B) — one client, one
|
||||
// endpoint, both shards.
|
||||
for i in 1..=24u64 {
|
||||
let resp = client
|
||||
.post(format!("{base_a}/items"))
|
||||
.json(&serde_json::json!({
|
||||
"entity_id": i, "metadata": { "title": format!("item {i}") }
|
||||
}))
|
||||
.send()
|
||||
.unwrap();
|
||||
assert!(
|
||||
resp.status().is_success(),
|
||||
"POST /items id={i}: {}",
|
||||
resp.status()
|
||||
);
|
||||
let resp = client
|
||||
.post(format!("{base_a}/signals"))
|
||||
.json(&serde_json::json!({ "entity_id": i, "signal": "view", "weight": 1.0 }))
|
||||
.send()
|
||||
.unwrap();
|
||||
assert!(
|
||||
resp.status().is_success(),
|
||||
"POST /signals id={i}: {}",
|
||||
resp.status()
|
||||
);
|
||||
}
|
||||
|
||||
// Both nodes' feeds must, after convergence, return EVERY item — proof that
|
||||
// (1) A forwarded shard-1 writes to B, (2) each shard replicated to its
|
||||
// follower, and (3) the read scatters over both local groups and merges.
|
||||
let feed_ids = |base: &str| -> std::collections::HashSet<u64> {
|
||||
let feed: serde_json::Value = client
|
||||
.get(format!("{base}/feed?profile=trending&limit=50"))
|
||||
.send()
|
||||
.unwrap()
|
||||
.json()
|
||||
.unwrap();
|
||||
feed["items"]
|
||||
.as_array()
|
||||
.unwrap()
|
||||
.iter()
|
||||
.filter_map(|it| it["entity_id"].as_u64())
|
||||
.collect()
|
||||
};
|
||||
|
||||
let deadline = Instant::now() + Duration::from_secs(10);
|
||||
loop {
|
||||
let a = feed_ids(&base_a);
|
||||
let b = feed_ids(&base_b);
|
||||
let all: std::collections::HashSet<u64> = (1..=24u64).collect();
|
||||
if a == all && b == all {
|
||||
// Sanity: both shards are actually represented (not one big shard).
|
||||
assert!(
|
||||
shard0_ids.iter().all(|i| a.contains(i))
|
||||
&& shard1_ids.iter().all(|i| a.contains(i)),
|
||||
"node A feed must merge BOTH shards"
|
||||
);
|
||||
break;
|
||||
}
|
||||
assert!(
|
||||
Instant::now() <= deadline,
|
||||
"feeds did not converge to all 24 items across both shards within 10s \
|
||||
(A={}, B={})",
|
||||
a.len(),
|
||||
b.len()
|
||||
);
|
||||
std::thread::sleep(Duration::from_millis(50));
|
||||
}
|
||||
|
||||
rt.shutdown_timeout(Duration::from_secs(2));
|
||||
}
|
||||
|
||||
/// m11p6 PARTIAL placement (the `EntityRoute::Remote` path + group-local reads):
|
||||
/// 3 nodes, 2 shards, RF=2, placed so shard 0 = {node-a (leader), node-b} and
|
||||
/// shard 1 = {node-b, node-c (leader)}. So node-a hosts ONLY shard 0, node-c
|
||||
/// ONLY shard 1, node-b BOTH. Writing every entity through node-a forces its
|
||||
/// shard-1 writes down the not-hosted (`Remote`) path — forwarded leader-first to
|
||||
/// a shard-1 replica node, which applies. node-b (full placement) then merges the
|
||||
/// whole corpus; node-a / node-c (partial) feed only their own group — the
|
||||
/// documented L2 read scope, asserted so it is intentional, not a silent gap.
|
||||
#[test]
|
||||
fn region_sharded_subset_placement_forwards_and_reads_are_group_scoped() {
|
||||
let names = [
|
||||
"node-a".to_string(),
|
||||
"node-b".to_string(),
|
||||
"node-c".to_string(),
|
||||
];
|
||||
let http = [free_addr(), free_addr(), free_addr()];
|
||||
let (a_s0, b_s0, b_s1, c_s1) = (free_addr(), free_addr(), free_addr(), free_addr());
|
||||
|
||||
let topology = || -> TopologySpec {
|
||||
// Each node's RegionSpec base grpc_addr (a hosted addr); replicas are
|
||||
// explicit per (node, shard), so the base is only the declared advertise.
|
||||
let bases = [a_s0, b_s0, c_s1];
|
||||
let regions = (0..3)
|
||||
.map(|i| RegionSpec {
|
||||
name: names[i].clone(),
|
||||
grpc_addr: Some(bases[i].to_string()),
|
||||
grpc_bind: None,
|
||||
http_addr: Some(http[i].to_string()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
})
|
||||
.collect();
|
||||
let replica = |node: usize, addr: SocketAddr| ShardReplicaSpec {
|
||||
node: names[node].clone(),
|
||||
grpc_addr: Some(addr.to_string()),
|
||||
grpc_bind: None,
|
||||
};
|
||||
TopologySpec {
|
||||
regions,
|
||||
leader: names[0].clone(),
|
||||
write_workers: None,
|
||||
timeouts: TimeoutsSpec::default(),
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: Some(vec![
|
||||
ShardSpec {
|
||||
id: 0,
|
||||
leader: Some(names[0].clone()),
|
||||
replicas: vec![replica(0, a_s0), replica(1, b_s0)],
|
||||
},
|
||||
ShardSpec {
|
||||
id: 1,
|
||||
leader: Some(names[1].clone()),
|
||||
replicas: vec![replica(1, b_s1), replica(2, c_s1)],
|
||||
},
|
||||
]),
|
||||
}
|
||||
};
|
||||
|
||||
let dirs: Vec<tempfile::TempDir> = (0..3).map(|_| region_dir()).collect();
|
||||
let node_a = Arc::new(build_region(topology(), &names[0], &dirs[0]));
|
||||
let node_b = Arc::new(build_region(topology(), &names[1], &dirs[1]));
|
||||
let node_c = Arc::new(build_region(topology(), &names[2], &dirs[2]));
|
||||
|
||||
let rt = tokio::runtime::Builder::new_multi_thread()
|
||||
.worker_threads(4)
|
||||
.enable_all()
|
||||
.build()
|
||||
.unwrap();
|
||||
serve(&rt, build_region_router(Arc::clone(&node_a), None), http[0]);
|
||||
serve(&rt, build_region_router(Arc::clone(&node_b), None), http[1]);
|
||||
serve(&rt, build_region_router(Arc::clone(&node_c), None), http[2]);
|
||||
|
||||
let client = reqwest::blocking::Client::new();
|
||||
let base_a = format!("http://{}", http[0]);
|
||||
let base_b = format!("http://{}", http[1]);
|
||||
let base_c = format!("http://{}", http[2]);
|
||||
|
||||
let router = ShardRouter::hash(2).unwrap();
|
||||
let mut shard0_ids = std::collections::HashSet::new();
|
||||
let mut shard1_ids = std::collections::HashSet::new();
|
||||
for i in 1..=24u64 {
|
||||
if router.route(EntityId::new(i)).0 == 0 {
|
||||
shard0_ids.insert(i);
|
||||
} else {
|
||||
shard1_ids.insert(i);
|
||||
}
|
||||
}
|
||||
assert!(
|
||||
!shard0_ids.is_empty() && !shard1_ids.is_empty(),
|
||||
"the hash must spread 24 ids across both shards"
|
||||
);
|
||||
|
||||
// Write every item + a view signal through node-A ONLY. Shard-0 ids apply
|
||||
// locally (A leads shard 0); shard-1 ids take EntityRoute::Remote (A hosts no
|
||||
// shard-1 replica) and forward to a shard-1 replica node — proof the
|
||||
// not-hosted gateway path works for both items and signals.
|
||||
for i in 1..=24u64 {
|
||||
let item = client
|
||||
.post(format!("{base_a}/items"))
|
||||
.json(&serde_json::json!({ "entity_id": i, "metadata": { "title": format!("item {i}") } }))
|
||||
.send()
|
||||
.unwrap();
|
||||
assert!(
|
||||
item.status().is_success(),
|
||||
"POST /items id={i} via node-a: {}",
|
||||
item.status()
|
||||
);
|
||||
let sig = client
|
||||
.post(format!("{base_a}/signals"))
|
||||
.json(&serde_json::json!({ "entity_id": i, "signal": "view", "weight": 1.0 }))
|
||||
.send()
|
||||
.unwrap();
|
||||
assert!(
|
||||
sig.status().is_success(),
|
||||
"POST /signals id={i} via node-a: {}",
|
||||
sig.status()
|
||||
);
|
||||
}
|
||||
|
||||
let feed_ids = |base: &str| -> std::collections::HashSet<u64> {
|
||||
let feed: serde_json::Value = client
|
||||
.get(format!("{base}/feed?profile=trending&limit=50"))
|
||||
.send()
|
||||
.unwrap()
|
||||
.json()
|
||||
.unwrap();
|
||||
feed["items"]
|
||||
.as_array()
|
||||
.unwrap()
|
||||
.iter()
|
||||
.filter_map(|it| it["entity_id"].as_u64())
|
||||
.collect()
|
||||
};
|
||||
let all: std::collections::HashSet<u64> = (1..=24u64).collect();
|
||||
|
||||
let deadline = Instant::now() + Duration::from_secs(10);
|
||||
loop {
|
||||
let a = feed_ids(&base_a);
|
||||
let b = feed_ids(&base_b);
|
||||
let c = feed_ids(&base_c);
|
||||
// node-b hosts BOTH groups ⇒ the whole corpus once replication converges;
|
||||
// node-a / node-c host ONE group ⇒ exactly that group's ids (group-local
|
||||
// read by design).
|
||||
if b == all && a == shard0_ids && c == shard1_ids {
|
||||
break;
|
||||
}
|
||||
assert!(
|
||||
Instant::now() <= deadline,
|
||||
"subset-placement feeds did not converge (a={}/{}, b={}/24, c={}/{})",
|
||||
a.len(),
|
||||
shard0_ids.len(),
|
||||
b.len(),
|
||||
c.len(),
|
||||
shard1_ids.len()
|
||||
);
|
||||
std::thread::sleep(Duration::from_millis(50));
|
||||
}
|
||||
|
||||
rt.shutdown_timeout(Duration::from_secs(2));
|
||||
}
|
||||
|
||||
/// A write to the FOLLOWER (a non-leader node) is FORWARDED to the leader; with
|
||||
/// the leader process not running, the forward fails and degrades to a 503 whose
|
||||
/// body names the (unreachable) leader — the task-03 leader-unreachable contract.
|
||||
@ -673,6 +997,7 @@ impl Trio {
|
||||
http_addr: Some(self.http[i].to_string()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
})
|
||||
.collect(),
|
||||
leader: self.names[0].clone(),
|
||||
@ -681,6 +1006,7 @@ impl Trio {
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -724,7 +1050,7 @@ fn region_node_heal_backfills_missed_items() {
|
||||
// DOWN during the leader's item/signal broadcast. The leader's best-effort
|
||||
// HTTP broadcast to it will fail (connection refused) and land in `failed`,
|
||||
// exactly as it would for a crashed/restarting region. The follower's gRPC
|
||||
// receiver IS running (started in RegionClusterState::new), so once the
|
||||
// receiver IS running (started in ShardReplica::new), so once the
|
||||
// leader heals it, the relay re-ships signals — but the item broadcast that
|
||||
// failed during downtime is what heal must backfill.
|
||||
|
||||
@ -850,7 +1176,7 @@ fn region_node_lag_honest_across_promote() {
|
||||
// Three nodes: us-east (leader, shard 0), eu-west (shard 1), ap-south
|
||||
// (shard 2). Dirs FIRST (see `region_dir` for the drop-order contract).
|
||||
let dirs: Vec<tempfile::TempDir> = (0..3).map(|_| region_dir()).collect();
|
||||
let nodes: Vec<Arc<RegionClusterState>> = (0..3)
|
||||
let nodes: Vec<Arc<ClusterNode>> = (0..3)
|
||||
.map(|i| Arc::new(build_region(trio.topology(), &trio.names[i], &dirs[i])))
|
||||
.collect();
|
||||
// Build the runtime AFTER the nodes (and their data-dir guards): locals
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
//! m8p10 task 03: cross-process route tests.
|
||||
//!
|
||||
//! Builds 3-node in-process clusters — distinct `RegionClusterState`s pointing at
|
||||
//! Builds 3-node in-process clusters — distinct `ShardReplica`s pointing at
|
||||
//! each other's REAL loopback gRPC + HTTP addresses — and drives them over real
|
||||
//! HTTP (axum on loopback) + real `GrpcTransport` replication + real reqwest
|
||||
//! forwarding. No OS processes (tier-3 OS-process coverage is tasks 04/05); these
|
||||
@ -26,8 +26,8 @@ use std::{
|
||||
};
|
||||
|
||||
use tidal_server::cluster::{
|
||||
ElectionSpec, RegionClusterState, RegionSpec, ReplicationSpec, TimeoutsSpec, TopologySpec,
|
||||
WalSpec, build_region_router,
|
||||
ClusterNode, ElectionSpec, RegionSpec, ReplicationSpec, TimeoutsSpec, TopologySpec, WalSpec,
|
||||
build_region_router,
|
||||
};
|
||||
use tidaldb::schema::{DecaySpec, EntityKind, Schema, SchemaBuilder, Window};
|
||||
|
||||
@ -87,6 +87,7 @@ impl Cluster3 {
|
||||
http_addr: Some(self.http[i].to_string()),
|
||||
grpc_tls: None,
|
||||
metrics_addr: None,
|
||||
zone: None,
|
||||
})
|
||||
.collect(),
|
||||
leader: self.names[0].clone(),
|
||||
@ -95,6 +96,7 @@ impl Cluster3 {
|
||||
replication: ReplicationSpec::default(),
|
||||
wal: WalSpec::default(),
|
||||
election: ElectionSpec::default(),
|
||||
shards: None,
|
||||
}
|
||||
}
|
||||
|
||||
@ -113,15 +115,11 @@ fn region_dir() -> tempfile::TempDir {
|
||||
|
||||
/// Build one region node off the reactor (GrpcTransport::new blocks on its own
|
||||
/// runtime, so it must run on a plain thread).
|
||||
fn build_region(
|
||||
topology: TopologySpec,
|
||||
region: &str,
|
||||
dir: &tempfile::TempDir,
|
||||
) -> RegionClusterState {
|
||||
fn build_region(topology: TopologySpec, region: &str, dir: &tempfile::TempDir) -> ClusterNode {
|
||||
let region = region.to_string();
|
||||
let data_dir = dir.path().to_path_buf();
|
||||
std::thread::spawn(move || {
|
||||
RegionClusterState::new(
|
||||
ClusterNode::new(
|
||||
&topology,
|
||||
®ion,
|
||||
region_schema(),
|
||||
|
||||
@ -7,7 +7,7 @@
|
||||
//! write to `nodes[0]` replicated inside node 0's own fabric and never crossed
|
||||
//! into `nodes[1]`'s process. That harness's own module docs flag this as the
|
||||
//! gap. `MultiProcCluster` closes it: each process runs `tidal-server cluster
|
||||
//! --region <name>` and owns exactly ONE region (a `RegionClusterState`), peering
|
||||
//! --region <name>` and owns exactly ONE region (a `ShardReplica`), peering
|
||||
//! with siblings over real gRPC and forwarding over real HTTP. Convergence is
|
||||
//! therefore verified against EVERY follower process's own `/cluster/status/local`,
|
||||
//! not a single node's internal view.
|
||||
|
||||
Loading…
Reference in New Issue
Block a user