B: B2/B3 health probes, loops and the in-memory status model

This commit is contained in:
ari committed 2026-10-19 20:55:25 -04:00
1 parent bd5b21bcc4
commit d94c2b3a1d
3 files changed
+994 -126

No files matched your search

+439 -81
View File
@@ -1,47 +1,278 @@
//! One-shot peer probes used by the diagnostic CLI.
//! Peer probes and health state.
//!
//! The registry's periodic loops reuse the transport; this module only
//! classifies a single round so `slinuxm peer check` can report it.
//! One probe is cheap and bounded: a TCP connect for "server up" and an
//! authenticated `GET /api/v1/ping` for "API running". The registry schedules
//! them and stores the resulting status; `slinuxm peer check` calls the same
//! probes once.
use std::time::Duration;
use std::hash::{Hash, Hasher};
use std::sync::Arc;
use std::time::{Duration, Instant};
use jiff::Timestamp;
use serde_json::Value;
use tokio_util::sync::CancellationToken;
use super::error::PeerError;
use super::transport::{PeerRequest, PeerTransport};
use super::types::{ApiStatus, AuthState, PeerConfig, ServerStatus};
use super::transport::{PeerBody, PeerRequest, PeerResponse, PeerTransport};
use super::types::{
ApiStatus, AuthState, PeerConfig, PeerName, PeerStatus, PeersConfig, ServerStatus,
};
const PING_PATH: &str = "/api/v1/ping";
const HEALTHZ_PATH: &str = "/healthz";
fn millis(d: Duration) -> u64 {
u64::try_from(d.as_millis()).unwrap_or(u64::MAX)
}
fn failed_server(err: &PeerError, at: Timestamp) -> ServerStatus {
ServerStatus {
up: false,
since: None,
last_checked: Some(at),
last_error: Some(err.class_str()),
latency_ms: None,
/// Timing for one peer's loops, copied out of [`PeersConfig`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct LoopConfig {
pub tcp_interval: Duration,
pub tcp_timeout: Duration,
pub api_interval: Duration,
pub api_timeout: Duration,
pub fail_threshold: u32,
}
impl LoopConfig {
pub fn from_config(cfg: &PeersConfig) -> Self {
Self {
tcp_interval: Duration::from_secs(cfg.tcp_interval_secs.max(1)),
tcp_timeout: Duration::from_secs(cfg.tcp_timeout_secs.max(1)),
api_interval: Duration::from_secs(cfg.api_interval_secs.max(1)),
api_timeout: Duration::from_secs(cfg.api_timeout_secs.max(1)),
fail_threshold: cfg.fail_threshold.max(1),
}
}
}
fn failed_api(err: &PeerError, at: Timestamp) -> ApiStatus {
ApiStatus {
up: false,
since: None,
last_checked: Some(at),
last_error: Some(err.class_str()),
latency_ms: None,
auth: AuthState::Unknown,
peer_version: None,
/// Outcome of the authenticated API probe.
#[derive(Debug, Clone)]
pub struct ApiObservation {
/// The API answered as running, even if it rejected our credential.
pub ok: bool,
pub auth: AuthState,
pub version: Option<String>,
pub latency: Option<Duration>,
/// Class token, never peer text or a URL.
pub error: Option<String>,
}
impl ApiObservation {
fn down(err: &PeerError) -> Self {
Self {
ok: false,
auth: AuthState::Unknown,
version: None,
latency: None,
error: Some(err.class_str()),
}
}
fn running(auth: AuthState, latency: Duration, version: Option<String>) -> Self {
Self {
ok: true,
auth,
version,
latency: Some(latency),
error: (auth == AuthState::Rejected).then(|| "unauthorized".to_owned()),
}
}
}
/// One TCP connect and one authenticated `GET /api/v1/ping`.
/// Server-side state accumulated between checks. The `since` timestamp lives
/// here rather than in the serialized status so it survives a reload of an
/// unchanged peer.
#[derive(Debug, Clone)]
struct ServerTracker {
up: bool,
since: Option<Timestamp>,
failures: u32,
}
impl ServerTracker {
fn seeded(status: &ServerStatus) -> Self {
Self {
up: status.up,
since: status.since,
failures: 0,
}
}
fn record(
&mut self,
result: Result<Duration, PeerError>,
now: Timestamp,
threshold: u32,
) -> ServerStatus {
let first = self.since.is_none();
let (last_error, latency) = match &result {
Ok(latency) => {
self.failures = 0;
if !self.up || first {
self.since = Some(now);
}
self.up = true;
(None, Some(millis(*latency)))
}
Err(err) => {
self.failures = self.failures.saturating_add(1);
if first && self.up {
self.since = Some(now);
}
if self.failures >= threshold && self.up {
self.up = false;
self.since = Some(now);
}
if first && self.since.is_none() {
self.since = Some(now);
}
(Some(err.class_str()), None)
}
};
ServerStatus {
up: self.up,
since: self.since,
last_checked: Some(now),
last_error,
latency_ms: latency,
}
}
}
#[derive(Debug, Clone)]
struct ApiTracker {
up: bool,
since: Option<Timestamp>,
failures: u32,
auth: AuthState,
}
impl ApiTracker {
fn seeded(status: &ApiStatus) -> Self {
Self {
up: status.up,
since: status.since,
failures: 0,
auth: status.auth,
}
}
fn record(&mut self, observation: ApiObservation, now: Timestamp, threshold: u32) -> ApiStatus {
let first = self.since.is_none();
if observation.ok {
self.failures = 0;
if !self.up || first {
self.since = Some(now);
}
self.up = true;
self.auth = observation.auth;
} else {
self.failures = self.failures.saturating_add(1);
if first {
self.since = Some(now);
}
if self.failures >= threshold && self.up {
self.up = false;
self.since = Some(now);
}
if !self.up {
self.auth = AuthState::Unknown;
}
}
ApiStatus {
up: self.up,
since: self.since,
last_checked: Some(now),
last_error: observation.error,
latency_ms: observation.latency.map(millis),
auth: self.auth,
peer_version: observation.version,
}
}
}
/// TCP connect, returning the connect latency.
pub async fn probe_server(
transport: &dyn PeerTransport,
peer: &PeerConfig,
timeout: Duration,
) -> Result<Duration, PeerError> {
transport.tcp_check(peer, timeout).await
}
/// Authenticated ping. A 401/403 means the API is up and our credential is
/// not; older peers without the route fall back to public `/healthz`.
pub async fn probe_api(
transport: &dyn PeerTransport,
peer: &PeerConfig,
timeout: Duration,
) -> ApiObservation {
let started = Instant::now();
let req = PeerRequest {
path: PING_PATH.into(),
query: Vec::new(),
headers: Vec::new(),
accept: None,
};
let resp = match transport.get(peer, req, timeout).await {
Ok(resp) => resp,
Err(PeerError::Rejected) => {
return ApiObservation::running(AuthState::Rejected, started.elapsed(), None)
}
Err(err) => return ApiObservation::down(&err),
};
let latency = started.elapsed();
match resp.status {
200..=299 => ApiObservation::running(AuthState::Ok, latency, version_of(&resp)),
401 | 403 => ApiObservation::running(AuthState::Rejected, latency, None),
404 => probe_healthz(transport, peer, timeout, latency).await,
status => ApiObservation::down(&PeerError::Http(status)),
}
}
async fn probe_healthz(
transport: &dyn PeerTransport,
peer: &PeerConfig,
timeout: Duration,
so_far: Duration,
) -> ApiObservation {
let req = PeerRequest {
path: HEALTHZ_PATH.into(),
query: Vec::new(),
headers: Vec::new(),
accept: None,
};
match transport.get(peer, req, timeout).await {
Ok(resp) if (200..300).contains(&resp.status) => {
ApiObservation::running(AuthState::Unknown, so_far, version_of(&resp))
}
Ok(resp) if resp.status == 401 || resp.status == 403 => {
ApiObservation::running(AuthState::Rejected, so_far, None)
}
Ok(resp) => ApiObservation::down(&PeerError::Http(resp.status)),
Err(PeerError::Rejected) => ApiObservation::running(AuthState::Rejected, so_far, None),
Err(err) => ApiObservation::down(&err),
}
}
fn version_of(resp: &PeerResponse) -> Option<String> {
let PeerBody::Bytes(body) = &resp.body else {
return None;
};
let value: Value = serde_json::from_slice(body).ok()?;
value
.get("version")
.and_then(Value::as_str)
.map(str::to_owned)
}
/// One TCP connect and one API probe, for the diagnostic CLI.
///
/// A peer that answers 401 is up with `auth = rejected`: the API is running
/// and our credential is the fault. Version and error classification beyond
/// that belong to the registry loops.
/// Both probes run concurrently so the caller waits for the slower one, not
/// the sum. The threshold is one: a single failure is reported as down.
pub async fn check_once(
transport: &dyn PeerTransport,
peer: &PeerConfig,
@@ -49,69 +280,93 @@ pub async fn check_once(
api_timeout: Duration,
) -> (ServerStatus, ApiStatus) {
let now = Timestamp::now();
let server = match transport.tcp_check(peer, tcp_timeout).await {
Ok(latency) => ServerStatus {
up: true,
since: Some(now),
last_checked: Some(now),
last_error: None,
latency_ms: Some(millis(latency)),
},
Err(err) => failed_server(&err, now),
};
let req = PeerRequest {
path: "/api/v1/ping".into(),
query: Vec::new(),
headers: Vec::new(),
accept: None,
};
// The GET is not timed, so `latency_ms` stays empty. The probe that
// replaces this stub measures it; `peer check` prints that value.
let api = match transport.get(peer, req, api_timeout).await {
Ok(resp) if (200..300).contains(&resp.status) => ApiStatus {
up: true,
since: Some(now),
last_checked: Some(now),
last_error: None,
latency_ms: None,
auth: AuthState::Ok,
peer_version: None,
},
Ok(resp) if resp.status == 401 || resp.status == 403 => ApiStatus {
up: true,
since: Some(now),
last_checked: Some(now),
last_error: Some("unauthorized".into()),
latency_ms: None,
auth: AuthState::Rejected,
peer_version: None,
},
Ok(resp) => failed_api(&PeerError::Http(resp.status), now),
Err(PeerError::Rejected) => ApiStatus {
up: true,
since: Some(now),
last_checked: Some(now),
last_error: Some("unauthorized".into()),
latency_ms: None,
auth: AuthState::Rejected,
peer_version: None,
},
Err(err) => failed_api(&err, now),
};
let (server, api) = tokio::join!(
probe_server(transport, peer, tcp_timeout),
probe_api(transport, peer, api_timeout)
);
let server = ServerTracker::seeded(&ServerStatus::default()).record(server, now, 1);
let api = ApiTracker::seeded(&ApiStatus::default()).record(api, now, 1);
(server, api)
}
/// Written by the loop after every probe.
pub trait StatusSink: Send + Sync {
fn apply_server(&self, name: &PeerName, status: ServerStatus);
fn apply_api(&self, name: &PeerName, status: ApiStatus);
}
/// How long to wait before a peer's first check, so a restart of the
/// consolidator does not hit every peer at the same instant.
pub fn start_jitter(name: &PeerName, interval: Duration) -> Duration {
let mut hasher = std::collections::hash_map::DefaultHasher::new();
name.hash(&mut hasher);
let window_ms = (u64::try_from(interval.as_millis()).unwrap_or(u64::MAX) / 4).max(1);
Duration::from_millis(hasher.finish() % window_ms)
}
/// One loop per peer: both checks share a task so a peer's own slow check
/// cannot fan out into more tasks. A hung peer only delays its own next
/// round; every await is bounded by its timeout and cancellation.
pub async fn run_peer(
transport: Arc<dyn PeerTransport>,
peer: PeerConfig,
cfg: LoopConfig,
initial: Option<PeerStatus>,
sink: Arc<dyn StatusSink>,
token: CancellationToken,
) {
let mut server = ServerTracker::seeded(
initial
.as_ref()
.map(|s| &s.server)
.unwrap_or(&ServerStatus::default()),
);
let mut api = ApiTracker::seeded(
initial
.as_ref()
.map(|s| &s.api)
.unwrap_or(&ApiStatus::default()),
);
let jitter = start_jitter(&peer.name, cfg.tcp_interval.min(cfg.api_interval));
let first = tokio::time::Instant::now() + jitter;
let mut tcp_deadline = first;
let mut api_deadline = first;
tokio::select! {
_ = token.cancelled() => return,
_ = tokio::time::sleep_until(first) => {}
}
loop {
tokio::select! {
_ = token.cancelled() => break,
_ = tokio::time::sleep_until(tcp_deadline) => {
let result = probe_server(transport.as_ref(), &peer, cfg.tcp_timeout).await;
sink.apply_server(&peer.name, server.record(result, Timestamp::now(), cfg.fail_threshold));
tcp_deadline = tokio::time::Instant::now() + cfg.tcp_interval;
}
_ = tokio::time::sleep_until(api_deadline) => {
let observation = probe_api(transport.as_ref(), &peer, cfg.api_timeout).await;
sink.apply_api(&peer.name, api.record(observation, Timestamp::now(), cfg.fail_threshold));
api_deadline = tokio::time::Instant::now() + cfg.api_interval;
}
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::peers::error::PeerError;
use crate::peers::testing::{fake_peers_config, FakeTransport};
fn cfg() -> PeersConfig {
fake_peers_config(&["beta"])
}
#[tokio::test]
async fn healthy_peer() {
let fake = FakeTransport::new();
let cfg = fake_peers_config(&["beta"]);
let cfg = cfg();
let (server, api) = check_once(
fake.as_ref(),
&cfg.peer[0],
@@ -123,6 +378,7 @@ mod tests {
assert_eq!(server.latency_ms, Some(1));
assert!(api.up);
assert_eq!(api.auth, AuthState::Ok);
assert_eq!(api.peer_version.as_deref(), Some("0.0.0-fake"));
assert_eq!(fake.calls().len(), 1);
}
@@ -131,7 +387,7 @@ mod tests {
let fake = FakeTransport::new();
fake.peer("beta")
.respond_err("/api/v1/ping", PeerError::Rejected);
let cfg = fake_peers_config(&["beta"]);
let cfg = cfg();
let (_, api) = check_once(
fake.as_ref(),
&cfg.peer[0],
@@ -144,13 +400,50 @@ mod tests {
assert_eq!(api.last_error.as_deref(), Some("unauthorized"));
}
#[tokio::test]
async fn a_401_status_is_rejected_not_down() {
let fake = FakeTransport::new();
fake.peer("beta")
.respond("/api/v1/ping", 401, "application/json", "{}");
let cfg = cfg();
let (_, api) = check_once(
fake.as_ref(),
&cfg.peer[0],
Duration::from_secs(1),
Duration::from_secs(1),
)
.await;
assert!(api.up);
assert_eq!(api.auth, AuthState::Rejected);
}
#[tokio::test]
async fn older_peer_falls_back_to_healthz_with_unknown_auth() {
let fake = FakeTransport::new();
fake.peer("beta")
.respond("/api/v1/ping", 404, "application/json", "{}")
.respond("/healthz", 200, "application/json", r#"{"status":"ok"}"#);
let cfg = cfg();
let (_, api) = check_once(
fake.as_ref(),
&cfg.peer[0],
Duration::from_secs(1),
Duration::from_secs(1),
)
.await;
assert!(api.up);
assert_eq!(api.auth, AuthState::Unknown);
let paths: Vec<_> = fake.calls().into_iter().map(|(_, r)| r.path).collect();
assert_eq!(paths, ["/api/v1/ping", "/healthz"]);
}
#[tokio::test]
async fn down_peer_reports_class() {
let fake = FakeTransport::new();
fake.peer("beta").tcp(Err(PeerError::Timeout));
fake.peer("beta")
.respond_err("/api/v1/ping", PeerError::Timeout);
let cfg = fake_peers_config(&["beta"]);
let cfg = cfg();
let (server, api) = check_once(
fake.as_ref(),
&cfg.peer[0],
@@ -170,7 +463,7 @@ mod tests {
let fake = FakeTransport::new();
fake.peer("beta")
.respond_err("/api/v1/ping", PeerError::BadRequest(leaked.into()));
let cfg = fake_peers_config(&["beta"]);
let cfg = cfg();
let (_, api) = check_once(
fake.as_ref(),
&cfg.peer[0],
@@ -183,4 +476,69 @@ mod tests {
assert!(!stored.contains("token"));
assert!(!stored.contains("http://"));
}
#[test]
fn threshold_needs_consecutive_failures_and_recovers_immediately() {
let mut server = ServerTracker::seeded(&ServerStatus::default());
let t0 = Timestamp::now();
let ok = Duration::from_millis(2);
let first = server.record(Ok(ok), t0, 2);
assert!(first.up);
assert_eq!(first.since, Some(t0));
let since_up = first.since;
let t1 = t0 + Duration::from_secs(1);
let still_up = server.record(Err(PeerError::Timeout), t1, 2);
assert!(still_up.up, "one failure below the threshold");
assert_eq!(still_up.since, since_up);
assert_eq!(still_up.last_error.as_deref(), Some("timeout"));
let t2 = t0 + Duration::from_secs(2);
let down = server.record(Err(PeerError::Timeout), t2, 2);
assert!(!down.up);
assert_eq!(down.since, Some(t2));
let t3 = t0 + Duration::from_secs(3);
let back = server.record(Ok(ok), t3, 2);
assert!(back.up, "recovery does not wait for a threshold");
assert_eq!(back.since, Some(t3));
}
#[test]
fn first_failed_check_still_records_a_since() {
let mut server = ServerTracker::seeded(&ServerStatus::default());
let t0 = Timestamp::now();
let st = server.record(Err(PeerError::Unreachable), t0, 2);
assert!(!st.up);
assert_eq!(st.since, Some(t0));
}
#[test]
fn api_tracker_keeps_auth_when_up_and_clears_it_when_down() {
let mut api = ApiTracker::seeded(&ApiStatus::default());
let t0 = Timestamp::now();
let up = api.record(
ApiObservation::running(AuthState::Rejected, Duration::from_millis(3), None),
t0,
2,
);
assert!(up.up);
assert_eq!(up.auth, AuthState::Rejected);
assert_eq!(up.last_error.as_deref(), Some("unauthorized"));
let t1 = t0 + Duration::from_secs(1);
let down = api.record(ApiObservation::down(&PeerError::Unreachable), t1, 1);
assert!(!down.up);
assert_eq!(down.auth, AuthState::Unknown);
assert_eq!(down.peer_version, None);
}
#[test]
fn jitter_is_bounded_and_stable() {
let interval = Duration::from_secs(10);
let name = PeerName::parse("beta").unwrap();
let j = start_jitter(&name, interval);
assert!(j < interval / 3, "{j:?}");
assert_eq!(j, start_jitter(&name, interval));
}
}
+397 -40
View File
@@ -2,11 +2,16 @@
//! side never awaits on a peer and is cheap enough for a metrics scrape.
use std::collections::BTreeMap;
use std::sync::{Arc, RwLock};
use std::sync::{Arc, Mutex, RwLock};
use std::time::Duration;
use tokio::task::JoinHandle;
use tokio_util::sync::CancellationToken;
use crate::config::{CliOverrides, Config};
use crate::error::AppError;
use crate::peers::error::PeerConfigError;
use crate::peers::health::{self, LoopConfig, StatusSink};
use crate::peers::transport::{PeerResponse, PeerTransport};
use crate::peers::types::{ApiStatus, PeerConfig, PeerName, PeerStatus, PeersConfig, ServerStatus};
@@ -29,9 +34,39 @@ struct RegistryState {
statuses: BTreeMap<PeerName, PeerStatus>,
}
/// Status storage shared with the loops; this is the only mutable state a
/// loop needs, so the registry can be dropped without leaking a task.
struct Shared {
state: RwLock<RegistryState>,
}
impl StatusSink for Shared {
fn apply_server(&self, name: &PeerName, status: ServerStatus) {
if let Some(entry) = lock_write(&self.state).statuses.get_mut(name) {
entry.server = status;
}
}
fn apply_api(&self, name: &PeerName, status: ApiStatus) {
if let Some(entry) = lock_write(&self.state).statuses.get_mut(name) {
entry.api = status;
}
}
}
struct LoopEntry {
peer: PeerConfig,
params: LoopConfig,
token: CancellationToken,
join: JoinHandle<()>,
}
pub struct PeerRegistry {
transport: Arc<dyn PeerTransport>,
state: RwLock<RegistryState>,
shared: Arc<Shared>,
loops: Mutex<BTreeMap<PeerName, LoopEntry>>,
loops_enabled: bool,
shutdown: CancellationToken,
}
fn initial_status(peer: &PeerConfig) -> PeerStatus {
@@ -52,8 +87,22 @@ fn lock_write<T>(lock: &RwLock<T>) -> std::sync::RwLockWriteGuard<'_, T> {
lock.write().unwrap_or_else(|e| e.into_inner())
}
fn lock_mutex<T>(lock: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
lock.lock().unwrap_or_else(|e| e.into_inner())
}
impl PeerRegistry {
/// Registry without loops, for tests that drive the transport directly.
pub(crate) fn new(config: PeersConfig, transport: Arc<dyn PeerTransport>) -> Self {
Self::build(config, transport, CancellationToken::new(), false)
}
fn build(
config: PeersConfig,
transport: Arc<dyn PeerTransport>,
shutdown: CancellationToken,
loops_enabled: bool,
) -> Self {
let statuses = config
.peer
.iter()
@@ -61,10 +110,15 @@ impl PeerRegistry {
.collect();
Self {
transport,
state: RwLock::new(RegistryState {
config: Arc::new(config),
statuses,
shared: Arc::new(Shared {
state: RwLock::new(RegistryState {
config: Arc::new(config),
statuses,
}),
}),
loops: Mutex::new(BTreeMap::new()),
loops_enabled,
shutdown,
}
}
@@ -73,40 +127,114 @@ impl PeerRegistry {
config: PeersConfig,
transport: Arc<dyn PeerTransport>,
) -> Self {
Self {
transport,
state: RwLock::new(RegistryState {
config: Arc::new(config),
statuses: statuses.into_iter().map(|s| (s.name.clone(), s)).collect(),
}),
let reg = Self::build(config, transport, CancellationToken::new(), false);
lock_write(&reg.shared.state).statuses =
statuses.into_iter().map(|s| (s.name.clone(), s)).collect();
reg
}
/// Real registry for the server: builds the outbound transport and starts
/// one loop per enabled peer. `shutdown` is the process token; cancelling
/// it stops the loops.
pub fn start(
config: PeersConfig,
shutdown: CancellationToken,
) -> Result<Arc<Self>, PeerConfigError> {
let transport = Arc::new(ReqwestTransport::new(&config)?);
Ok(Self::start_with_transport(config, transport, shutdown))
}
/// Loop-enabled registry over a supplied transport, for tests.
pub fn start_with_transport(
config: PeersConfig,
transport: Arc<dyn PeerTransport>,
shutdown: CancellationToken,
) -> Arc<Self> {
let registry = Arc::new(Self::build(config, transport, shutdown.child_token(), true));
registry.spawn_all();
registry
}
fn spawn_all(&self) {
let peers: Vec<PeerConfig> = if self.config().enabled {
self.config()
.peer
.iter()
.filter(|peer| peer.enabled)
.cloned()
.collect()
} else {
Vec::new()
};
for peer in peers {
self.spawn_peer(&peer);
}
}
fn spawn_peer(&self, peer: &PeerConfig) {
let params = LoopConfig::from_config(&self.config());
let initial = self.status_internal(&peer.name);
let token = self.shutdown.child_token();
let sink: Arc<dyn StatusSink> = self.shared.clone();
let transport = self.transport.clone();
let peer_copy = peer.clone();
let name = peer.name.clone();
let task_token = token.clone();
let join = tokio::spawn(async move {
health::run_peer(transport, peer_copy, params, initial, sink, task_token).await;
});
lock_mutex(&self.loops).insert(
name,
LoopEntry {
peer: peer.clone(),
params,
token,
join,
},
);
}
fn stop_loop(&self, entry: LoopEntry) {
// Cancel, then let the task finish its bounded probe on its own.
entry.token.cancel();
tokio::spawn(async move {
let _ = entry.join.await;
});
}
/// Sorted by name; disabled peers included.
pub fn snapshot(&self) -> Vec<PeerStatus> {
lock_read(&self.state).statuses.values().cloned().collect()
lock_read(&self.shared.state)
.statuses
.values()
.cloned()
.collect()
}
pub fn status(&self, name: &str) -> Result<PeerStatus, AppError> {
let Ok(key) = PeerName::parse(name) else {
return Err(AppError::NotFound(format!("peer {name}: unknown peer")));
};
lock_read(&self.state)
lock_read(&self.shared.state)
.statuses
.get(&key)
.cloned()
.ok_or_else(|| AppError::NotFound(format!("peer {name}: unknown peer")))
}
fn status_internal(&self, name: &PeerName) -> Option<PeerStatus> {
lock_read(&self.shared.state).statuses.get(name).cloned()
}
pub fn contains(&self, name: &str) -> bool {
let Ok(key) = PeerName::parse(name) else {
return false;
};
lock_read(&self.state).statuses.contains_key(&key)
lock_read(&self.shared.state).statuses.contains_key(&key)
}
pub fn config(&self) -> Arc<PeersConfig> {
lock_read(&self.state).config.clone()
lock_read(&self.shared.state).config.clone()
}
/// Single proxy entry point for REST and MCP.
@@ -119,7 +247,7 @@ impl PeerRegistry {
) -> Result<PeerResponse, AppError> {
let key = PeerName::parse(name).ok();
let (peer, timeout) = {
let st = lock_read(&self.state);
let st = lock_read(&self.shared.state);
let peer = key.and_then(|key| st.config.peer.iter().find(|p| p.name == key).cloned());
(peer, Duration::from_secs(st.config.request_timeout_secs))
};
@@ -128,30 +256,114 @@ impl PeerRegistry {
}
/// Replace the peer set. Statuses of peers that survive a reload are kept
/// so `since` survives a config change. Validation arrives with the
/// config and secrets tasks; this body only swaps state.
/// so `since` survives a config change. Loops for new or changed peers are
/// started, loops for removed or disabled peers are cancelled.
///
/// A bad `[peers]` block never reaches this method: the handler logs the
/// config error and keeps the old registry, as TLS reload does.
pub async fn reload(&self, new: PeersConfig) -> Result<(), PeerConfigError> {
let mut st = lock_write(&self.state);
let old = std::mem::take(&mut st.statuses);
st.statuses = new
.peer
.iter()
.map(|peer| {
let mut status = old
.get(&peer.name)
.cloned()
.unwrap_or_else(|| initial_status(peer));
status.enabled = peer.enabled;
status.url = peer.display_url();
(peer.name.clone(), status)
})
.collect();
st.config = Arc::new(new);
let params = LoopConfig::from_config(&new);
{
let mut st = lock_write(&self.shared.state);
let old = std::mem::take(&mut st.statuses);
st.statuses = new
.peer
.iter()
.map(|peer| {
let mut status = old
.get(&peer.name)
.cloned()
.unwrap_or_else(|| initial_status(peer));
status.enabled = peer.enabled;
status.url = peer.display_url();
(peer.name.clone(), status)
})
.collect();
st.config = Arc::new(new.clone());
}
if self.loops_enabled {
self.resync_loops(&new, params);
}
Ok(())
}
/// Health loops stop on this; nothing is running yet.
pub async fn shutdown(&self) {}
fn resync_loops(&self, new: &PeersConfig, params: LoopConfig) {
let mut old = std::mem::take(&mut *lock_mutex(&self.loops));
let mut kept = BTreeMap::new();
for (name, entry) in std::mem::take(&mut old) {
let wanted = new
.enabled
.then(|| {
new.peer
.iter()
.find(|peer| peer.name == name && peer.enabled)
})
.flatten();
match wanted {
Some(peer) if entry.peer == *peer && entry.params == params => {
kept.insert(name, entry);
}
_ => self.stop_loop(entry),
}
}
*lock_mutex(&self.loops) = kept;
if new.enabled {
for peer in new.peer.iter().filter(|peer| peer.enabled) {
if !lock_mutex(&self.loops).contains_key(&peer.name) {
self.spawn_peer(peer);
}
}
}
}
/// Stop every loop and wait (bounded) for the tasks to exit.
pub async fn shutdown(&self) {
self.shutdown.cancel();
let joins: Vec<JoinHandle<()>> = std::mem::take(&mut *lock_mutex(&self.loops))
.into_values()
.map(|entry| entry.join)
.collect();
let join_all = async {
for join in joins {
let _ = join.await;
}
};
if tokio::time::timeout(Duration::from_secs(5), join_all)
.await
.is_err()
{
tracing::warn!("peer health loops did not stop in time");
}
}
}
/// Everything the SIGHUP handler needs to re-read the `[peers]` block.
#[derive(Clone)]
pub struct PeersReloader {
registry: Arc<PeerRegistry>,
path: Option<std::path::PathBuf>,
overrides: CliOverrides,
}
impl PeersReloader {
pub fn new(
registry: Arc<PeerRegistry>,
path: Option<std::path::PathBuf>,
overrides: CliOverrides,
) -> Self {
Self {
registry,
path,
overrides,
}
}
/// Re-read the whole config and apply only `[peers]`. A failure returns
/// the config error; the old registry stays in place.
pub async fn reload(&self) -> Result<(), PeerConfigError> {
let cfg = Config::load(self.path.as_deref(), &self.overrides)?;
self.registry.reload(cfg.peers).await
}
}
#[cfg(test)]
@@ -159,6 +371,7 @@ mod tests {
use super::*;
use crate::peers::testing::{fake_peers_config, registry_with, static_registry, FakeTransport};
use crate::peers::transport::PeerBody;
use crate::peers::types::AuthState;
fn status(name: &str, up: bool) -> PeerStatus {
PeerStatus {
@@ -184,7 +397,7 @@ mod tests {
assert_eq!(names, ["beta", "gamma"]);
let b = reg.status("beta").unwrap();
assert!(b.server.up);
assert_eq!(b.api.auth, crate::peers::types::AuthState::Unknown);
assert_eq!(b.api.auth, AuthState::Unknown);
}
#[test]
@@ -247,7 +460,7 @@ mod tests {
async fn reload_keeps_surviving_status_and_prunes_removed() {
let reg = registry_with(FakeTransport::new(), fake_peers_config(&["beta", "gamma"]));
{
let mut st = lock_write(&reg.state);
let mut st = lock_write(&reg.shared.state);
let beta = st
.statuses
.get_mut(&PeerName::parse("beta").unwrap())
@@ -256,8 +469,7 @@ mod tests {
}
let mut new = fake_peers_config(&["beta", "delta"]);
new.peer[0].enabled = false;
let reloaded: Result<(), PeerConfigError> = reg.reload(new).await;
reloaded.unwrap();
reg.reload(new).await.unwrap();
let snap = reg.snapshot();
assert_eq!(snap.len(), 2);
@@ -267,4 +479,149 @@ mod tests {
assert_eq!(snap[1].name.as_str(), "delta");
assert!(!snap[1].server.up);
}
async fn wait_for<F: Fn(&PeerRegistry) -> bool>(reg: &PeerRegistry, f: F) {
let deadline = tokio::time::Instant::now() + Duration::from_secs(5);
loop {
if f(reg) {
return;
}
assert!(
tokio::time::Instant::now() < deadline,
"condition not met: {:?}",
reg.snapshot()
);
tokio::time::sleep(Duration::from_millis(10)).await;
}
}
#[tokio::test]
async fn loops_mark_up_and_honour_the_failure_threshold() {
let fake = FakeTransport::new();
let mut cfg = fake_peers_config(&["beta"]);
cfg.fail_threshold = 2;
let reg = PeerRegistry::start_with_transport(cfg, fake.clone(), CancellationToken::new());
wait_for(&reg, |r| {
let s = r.status("beta").unwrap();
s.server.up && s.api.up && s.api.auth == AuthState::Ok
})
.await;
let up = reg.status("beta").unwrap();
assert!(up.server.since.is_some());
assert_eq!(up.api.peer_version.as_deref(), Some("0.0.0-fake"));
fake.peer("beta").tcp(Err(crate::peers::PeerError::Timeout));
fake.peer("beta")
.respond_err("/api/v1/ping", crate::peers::PeerError::Timeout);
// One failure each: still up with the error visible.
wait_for(&reg, |r| {
let s = r.status("beta").unwrap();
s.server.last_error.as_deref() == Some("timeout")
&& s.api.last_error.as_deref() == Some("timeout")
})
.await;
assert!(reg.status("beta").unwrap().server.up);
reg.shutdown().await;
let after = reg.status("beta").unwrap();
assert!(after.server.up, "one failure below threshold = still up");
}
#[tokio::test]
async fn a_hung_peer_does_not_delay_another_peers_checks() {
let fake = FakeTransport::new();
fake.peer("beta").delay(Duration::from_secs(2));
fake.peer("gamma")
.respond("/api/v1/ping", 200, "application/json", "{}");
let reg = PeerRegistry::start_with_transport(
fake_peers_config(&["beta", "gamma"]),
fake.clone(),
CancellationToken::new(),
);
wait_for(&reg, |r| {
r.status("gamma").unwrap().api.last_checked.is_some()
})
.await;
let gamma_checked = reg.status("gamma").unwrap().api.last_checked;
assert!(reg.status("beta").unwrap().api.last_checked.is_none());
assert!(gamma_checked.is_some());
reg.shutdown().await;
}
#[test]
#[allow(clippy::result_large_err)] // figment::Jail closures return figment::Error
fn reloader_rereads_the_peers_block_and_keeps_the_old_set_on_error() {
use figment::Jail;
let reg = registry_with(FakeTransport::new(), fake_peers_config(&["beta"]));
Jail::expect_with(|jail| {
jail.create_file(
"config.toml",
r#"
[peers]
enabled = true
api_interval_secs = 2
[[peers.peer]]
name = "gamma"
url = "http://gamma:1984"
token_file = "/run/credentials/skyg-linux-tracker.service/peer-gamma"
allow_plain_http = true
"#,
)?;
let reloader = PeersReloader::new(
reg.clone(),
Some(std::path::PathBuf::from("config.toml")),
CliOverrides::default(),
);
tokio::runtime::Builder::new_current_thread()
.enable_all()
.build()
.unwrap()
.block_on(reloader.reload())
.unwrap();
let snap = reg.snapshot();
assert_eq!(snap.len(), 1);
assert_eq!(snap[0].name.as_str(), "gamma");
assert_eq!(reg.config().api_interval_secs, 2);
let broken = PeersReloader::new(
reg.clone(),
Some(std::path::PathBuf::from("missing.toml")),
CliOverrides::default(),
);
assert!(tokio::runtime::Builder::new_current_thread()
.enable_all()
.build()
.unwrap()
.block_on(broken.reload())
.is_err());
assert!(reg.contains("gamma"), "a bad reload keeps the old peers");
Ok(())
});
}
#[tokio::test]
async fn reload_moves_loops_to_the_new_peer_set() {
let fake = FakeTransport::new();
let reg = PeerRegistry::start_with_transport(
fake_peers_config(&["beta"]),
fake.clone(),
CancellationToken::new(),
);
wait_for(&reg, |r| r.status("beta").unwrap().server.up).await;
reg.reload(fake_peers_config(&["beta", "gamma"]))
.await
.unwrap();
wait_for(&reg, |r| r.status("gamma").unwrap().server.up).await;
assert!(
reg.status("beta").unwrap().server.up,
"kept peer keeps state"
);
reg.reload(fake_peers_config(&["gamma"])).await.unwrap();
assert!(reg.status("beta").is_err());
assert!(reg.contains("gamma"));
reg.shutdown().await;
}
}
+158 -5
View File
@@ -1,17 +1,28 @@
//! Peer client against a local HTTP listener.
//! Peer client against a local HTTP listener, then real health loops against
//! a second in-process `slinuxm` with fake sources.
//!
//! The listener is not a full `slinuxm`; it only has to speak enough HTTP/1.1
//! to prove the real transport. Later cases in this file drive a second
//! in-process server with fake sources.
//! The first cases use a hand-written listener: it only has to speak enough
//! HTTP/1.1 to prove the transport. The later cases run the assembled router
//! so the `PeerTransport`, probing and registry state are exercised together.
use std::path::Path;
use std::sync::Arc;
use std::time::Duration;
use skyg_linux_tracker::peers::registry::ReqwestTransport;
use skyg_linux_tracker::auth::tokens;
use skyg_linux_tracker::config::Config;
use skyg_linux_tracker::peers::registry::{PeerRegistry, ReqwestTransport};
use skyg_linux_tracker::peers::transport::{PeerBody, PeerRequest, PeerTransport};
use skyg_linux_tracker::peers::types::AuthState;
use skyg_linux_tracker::peers::types::PeerStatus;
use skyg_linux_tracker::peers::types::{PeerConfig, PeerName, PeersConfig};
use skyg_linux_tracker::peers::PeerError;
use skyg_linux_tracker::server;
use skyg_linux_tracker::sources::fake::fake_state;
use sqlx::sqlite::SqlitePoolOptions;
use tokio::io::{AsyncReadExt, AsyncWriteExt};
use tokio::net::TcpListener;
use tokio_util::sync::CancellationToken;
fn transport() -> ReqwestTransport {
ReqwestTransport::new(&PeersConfig {
@@ -131,3 +142,145 @@ async fn client_get_is_capped_and_does_not_follow_redirects() {
assert!(matches!(err, PeerError::Oversize), "{err:?}");
let _ = std::fs::remove_file(&token);
}
/// A second in-process `slinuxm` (fake sources) acting as the peer.
struct RunningPeer {
port: u16,
shutdown: CancellationToken,
join: tokio::task::JoinHandle<anyhow::Result<()>>,
}
impl RunningPeer {
async fn stop(self) {
self.shutdown.cancel();
let _ = self.join.await;
}
}
async fn start_peer_at(port: u16, auth: bool) -> RunningPeer {
let listener = TcpListener::bind(("127.0.0.1", port)).await.unwrap();
let port = listener.local_addr().unwrap().port();
let mut config = Config::default();
config.server.port = port;
config.auth.enabled = auth;
let mut state = fake_state(config);
if auth {
let pool = SqlitePoolOptions::new()
.max_connections(1)
.connect("sqlite::memory:")
.await
.unwrap();
skyg_linux_tracker::db::migrate(&pool).await.unwrap();
let _ = tokens::create(&pool, "peer-consumer").await.unwrap();
state.db = Some(pool);
}
let shutdown = state.shutdown.clone();
let app = server::build_router(state);
let join = tokio::spawn(server::serve_listener(
listener,
None,
app,
shutdown.clone(),
server::Limits::default(),
));
RunningPeer {
port,
shutdown,
join,
}
}
fn registry_for(port: u16, token_path: &Path, shutdown: CancellationToken) -> Arc<PeerRegistry> {
let mut cfg = skyg_linux_tracker::peers::testing::fake_peers_config(&["beta"]);
cfg.peer[0].url = url::Url::parse(&format!("http://127.0.0.1:{port}")).unwrap();
cfg.peer[0].token_file = token_path.to_path_buf();
PeerRegistry::start(cfg, shutdown).unwrap()
}
async fn wait_for<F: Fn(&PeerStatus) -> bool>(reg: &PeerRegistry, f: F) -> PeerStatus {
let deadline = tokio::time::Instant::now() + Duration::from_secs(8);
loop {
let status = reg.status("beta").unwrap();
if f(&status) {
return status;
}
assert!(
tokio::time::Instant::now() < deadline,
"peer status stayed {:?}",
status
);
tokio::time::sleep(Duration::from_millis(25)).await;
}
}
#[tokio::test]
async fn loops_observe_up_down_and_recovery() {
let token = token_file("anything\n");
let peer = start_peer_at(0, false).await;
let port = peer.port;
let shutdown = CancellationToken::new();
let reg = registry_for(port, &token, shutdown.clone());
let up = wait_for(&reg, |s| s.server.up && s.api.up).await;
assert_eq!(up.api.auth, AuthState::Ok);
let since_up = up.server.since.expect("first check records since");
// Port closed: both checks fail within their own timeout.
peer.stop().await;
let down = wait_for(&reg, |s| !s.server.up && !s.api.up).await;
assert!(down.server.last_error.is_some());
assert_eq!(down.api.auth, AuthState::Unknown);
let since_down = down.server.since.expect("flip records since");
assert_ne!(since_down, since_up);
// Same port again: recovery is immediate, not threshold-bound.
let peer = start_peer_at(port, false).await;
let back = wait_for(&reg, |s| s.server.up && s.api.up).await;
assert_ne!(back.server.since, Some(since_down));
reg.shutdown().await;
peer.stop().await;
let _ = std::fs::remove_file(&token);
}
#[tokio::test]
async fn api_down_with_the_listener_still_accepting() {
let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
let port = listener.local_addr().unwrap().port();
let hang = tokio::spawn(async move {
let mut held = Vec::new();
while let Ok((sock, _)) = listener.accept().await {
held.push(sock);
if held.len() > 8 {
held.remove(0);
}
}
});
let token = token_file("anything\n");
let shutdown = CancellationToken::new();
let reg = registry_for(port, &token, shutdown.clone());
let status = wait_for(&reg, |s| s.api.last_checked.is_some() && !s.api.up).await;
assert!(status.server.up, "TCP connect succeeds");
assert!(!status.api.up, "no HTTP answer before the timeout");
assert_eq!(status.api.last_error.as_deref(), Some("timeout"));
reg.shutdown().await;
hang.abort();
let _ = std::fs::remove_file(&token);
}
#[tokio::test]
async fn a_rejected_token_leaves_the_api_up_and_names_the_auth_state() {
let peer = start_peer_at(0, true).await;
let token = token_file("not-the-peer-token\n");
let shutdown = CancellationToken::new();
let reg = registry_for(peer.port, &token, shutdown.clone());
let status = wait_for(&reg, |s| s.api.last_checked.is_some()).await;
assert!(status.server.up);
assert!(status.api.up, "a 401 proves the API is running");
assert_eq!(status.api.auth, AuthState::Rejected);
assert_eq!(status.api.last_error.as_deref(), Some("unauthorized"));
reg.shutdown().await;
peer.stop().await;
let _ = std::fs::remove_file(&token);
}