Files
marqdo/src/runtime/supervisor.rs
T
cflmyandCursor 3cabe53da2 release: v2.1.4 isolated HTTP execute (ADR 0009)
Default execute/sse-live/stream to Supervisor worker pool; MR-01 gold;
/_runtime/*; bump CLI/wasm/ext to 2.1.4.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-26 18:33:33 +08:00

304 lines
9.7 KiB
Rust

//! Supervisor: lease workers, enforce deadlines, reap orphans (M-R2).
use std::collections::HashMap;
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};
use crate::runtime::lease::{Lease, WorkerState};
use crate::runtime::protocol::InvokeRequestV2;
use crate::runtime::worker::{WorkerHandle, WorkerPool};
#[derive(Debug, Clone)]
pub struct SupervisorConfig {
pub execute_slots: usize,
pub worker_ttl: Duration,
pub cancel_grace: Duration,
pub force_kill_after: Duration,
}
impl Default for SupervisorConfig {
fn default() -> Self {
Self {
execute_slots: 2,
worker_ttl: Duration::from_secs(600),
cancel_grace: Duration::from_secs(2),
force_kill_after: Duration::from_secs(3),
}
}
}
struct ActiveWorker {
handle: WorkerHandle,
cancel_since: Option<Instant>,
}
pub struct Supervisor {
cfg: SupervisorConfig,
pool: Mutex<WorkerPool>,
active: Mutex<HashMap<String, ActiveWorker>>,
cancel_dir: PathBuf,
event_dir: PathBuf,
}
impl Supervisor {
pub fn new(
cfg: SupervisorConfig,
staging_root: impl AsRef<Path>,
marqdo_bin: impl AsRef<Path>,
cancel_dir: impl AsRef<Path>,
event_dir: impl AsRef<Path>,
) -> Self {
let cancel_dir = cancel_dir.as_ref().to_path_buf();
let event_dir = event_dir.as_ref().to_path_buf();
let _ = std::fs::create_dir_all(&cancel_dir);
let _ = std::fs::create_dir_all(&event_dir);
Self {
pool: Mutex::new(WorkerPool::new(
staging_root,
marqdo_bin,
cfg.execute_slots,
)),
cfg,
active: Mutex::new(HashMap::new()),
cancel_dir,
event_dir,
}
}
pub fn warm(&self) -> Result<(), String> {
self.pool.lock().unwrap().warm()
}
pub fn dispatch(&self, req: InvokeRequestV2) -> Result<String, String> {
req.validate()?;
{
let active = self.active.lock().unwrap();
if active.len() >= self.cfg.execute_slots {
return Err("supervisor.no_slot".into());
}
}
let mut pool = self.pool.lock().unwrap();
let worker = pool.acquire()?;
drop(pool);
let epoch = Instant::now().elapsed().as_nanos() as u64;
let lease = Lease::new(
format!("lease_{}", req.task_id),
epoch,
&req.task_id,
&req.run_id,
&req.attempt_id,
self.cfg.worker_ttl,
);
let cancel = self.cancel_dir.join(format!("{}.tok", req.task_id));
if !cancel.exists() {
std::fs::write(&cancel, b"").map_err(|e| e.to_string())?;
}
let events = self.event_dir.join(format!("{}.ndjson", req.task_id));
std::fs::write(&events, b"").map_err(|e| e.to_string())?;
// Durable emit path (not under staging — WorkerHandle::Drop removes staging).
let emit = self.event_dir.join(format!("{}.result.json", req.task_id));
let fs_root = req
.context
.fs_root
.as_ref()
.map(PathBuf::from)
.filter(|p| p.is_dir())
.ok_or_else(|| "supervisor.fs_root_required".to_string())?;
let args_path = req.args_ref.as_ref().and_then(|r| {
r.strip_prefix("file:")
.map(PathBuf::from)
.or_else(|| {
let p = PathBuf::from(r);
p.is_file().then_some(p)
})
});
let handle = {
if std::env::var_os("MARQDO_RUNTIME_DRY_RUN").is_some() {
let mut w = worker;
w.lease = Some(lease.clone());
w.emit_result = Some(emit.clone());
w.state = WorkerState::Running;
w
} else {
self.pool.lock().unwrap().lease_and_start(
worker,
lease.clone(),
&req,
&cancel,
&events,
&fs_root,
args_path.as_deref(),
&emit,
)?
}
};
let lease_id = lease.id.0.clone();
self.active.lock().unwrap().insert(
lease_id.clone(),
ActiveWorker {
handle,
cancel_since: None,
},
);
Ok(lease_id)
}
pub fn event_path(&self, task_id: &str) -> PathBuf {
self.event_dir.join(format!("{task_id}.ndjson"))
}
pub fn cancel_path(&self, task_id: &str) -> PathBuf {
self.cancel_dir.join(format!("{task_id}.tok"))
}
pub fn emit_result_path(&self, lease_id: &str) -> Option<PathBuf> {
self.active
.lock()
.unwrap()
.get(lease_id)
.and_then(|w| w.handle.emit_result.clone())
}
/// Non-blocking poll: if worker exited, remove from active and return exit code.
pub fn try_reap(&self, lease_id: &str) -> Result<Option<i32>, String> {
let mut active = self.active.lock().unwrap();
let Some(w) = active.get_mut(lease_id) else {
return Ok(None);
};
match w.handle.try_wait()? {
Some(st) => {
let code = st.code().unwrap_or(-1);
active.remove(lease_id);
let _ = self.pool.lock().unwrap().warm();
Ok(Some(code))
}
None => Ok(None),
}
}
pub fn signal_cancel(&self, lease_id: &str, task_id: &str) -> Result<(), String> {
let tok = self.cancel_dir.join(format!("{task_id}.tok"));
std::fs::write(&tok, b"cancelled").map_err(|e| e.to_string())?;
let mut active = self.active.lock().unwrap();
if let Some(w) = active.get_mut(lease_id) {
w.cancel_since = Some(Instant::now());
w.handle.state = WorkerState::Draining;
}
Ok(())
}
/// Reap expired / force-kill after cancel grace. Returns orphan count.
pub fn janitor_tick(&self) -> usize {
let mut orphans = 0usize;
let mut active = self.active.lock().unwrap();
let mut done = Vec::new();
for (id, w) in active.iter_mut() {
if let Some(lease) = &w.handle.lease {
if lease.expired() {
w.handle.kill_process_group();
orphans += 1;
done.push(id.clone());
continue;
}
}
if let Some(since) = w.cancel_since {
if since.elapsed() > self.cfg.cancel_grace + self.cfg.force_kill_after {
w.handle.kill_process_group();
orphans += 1;
done.push(id.clone());
}
}
}
for id in done {
active.remove(&id);
}
orphans
}
pub fn active_count(&self) -> usize {
self.active.lock().unwrap().len()
}
pub fn release(&self, lease_id: &str) {
let mut active = self.active.lock().unwrap();
if let Some(mut w) = active.remove(lease_id) {
w.handle.state = WorkerState::Draining;
w.handle.kill_process_group();
}
// Async replenish hint.
let _ = self.pool.lock().unwrap().warm();
}
}
/// Shared supervisor handle for HTTP/control wiring.
pub type SharedSupervisor = Arc<Supervisor>;
#[cfg(test)]
mod tests {
use super::*;
use crate::runtime::protocol::{EntryRef, InvokeContext, LeaseRef};
fn req(task: &str) -> InvokeRequestV2 {
InvokeRequestV2 {
schema: "InvokeRequestV2".into(),
schema_version: 2,
task_id: task.into(),
run_id: format!("run_{task}"),
attempt_id: format!("attempt_{task}"),
lease: LeaseRef {
id: format!("lease_{task}"),
epoch: 1,
},
entry: EntryRef {
module: "agent/qdagent.mq.md".into(),
function: "run".into(),
},
args_ref: None,
context: InvokeContext {
workspace_id: "ws".into(),
snapshot_id: None,
selection_ref: None,
capabilities_ref: None,
deadline_at: "2099-01-01T00:00:00Z".into(),
fs_root: Some("/tmp".into()),
},
}
}
#[test]
fn dry_run_two_parallel_leases() {
std::env::set_var("MARQDO_RUNTIME_DRY_RUN", "1");
let root = std::env::temp_dir().join(format!("mq-sup-{}", std::process::id()));
let _ = std::fs::remove_dir_all(&root);
let bin = std::env::current_exe().unwrap();
let _ = std::fs::create_dir_all(root.join("fs"));
let mut ra = req("a");
ra.context.fs_root = Some(root.join("fs").to_string_lossy().into());
let mut rb = req("b");
rb.context.fs_root = ra.context.fs_root.clone();
let mut rc = req("c");
rc.context.fs_root = ra.context.fs_root.clone();
let sup = Supervisor::new(
SupervisorConfig {
execute_slots: 2,
..SupervisorConfig::default()
},
root.join("stage"),
&bin,
root.join("cancel"),
root.join("events"),
);
sup.warm().unwrap();
let a = sup.dispatch(ra).unwrap();
let b = sup.dispatch(rb).unwrap();
assert_ne!(a, b);
assert_eq!(sup.active_count(), 2);
assert!(sup.dispatch(rc).is_err());
sup.release(&a);
assert_eq!(sup.active_count(), 1);
let _ = std::fs::remove_dir_all(&root);
std::env::remove_var("MARQDO_RUNTIME_DRY_RUN");
}
}