veloren_common/

slowjob.rs

use hashbrown::HashMap;
use rayon::ThreadPool;
use std::{
    collections::VecDeque,
    sync::{Arc, Mutex},
    time::Instant,
};
use tracing::{error, warn};

/// Provides a Wrapper around rayon threadpool to execute slow-jobs.
/// slow means, the job doesn't need to not complete within the same tick.
/// DO NOT USE I/O blocking jobs, but only CPU heavy jobs.
/// Jobs run here, will reduce the ammount of threads rayon can use during the
/// main tick.
///
/// ## Configuration
/// This Pool allows you to configure certain names of jobs and assign them a
/// maximum number of threads # Example
/// Your system has 16 cores, you assign 12 cores for slow-jobs.
/// Then you can configure all jobs with the name `CHUNK_GENERATOR` to spawn on
/// max 50% (6 = cores)
///
/// ## Spawn Order
/// - At least 1 job of a configuration is allowed to run if global limit isn't
///   hit.
/// - remaining capacities are spread in relation to their limit. e.g. a
///   configuration with double the limit will be sheduled to spawn double the
///   tasks, starting by a round robin.
///
/// ## States
/// - queued
/// - spawned
/// - started
/// - finished
/// ```
/// # use veloren_common::slowjob::SlowJobPool;
/// # use std::sync::Arc;
///
/// let threadpool = rayon::ThreadPoolBuilder::new()
///     .num_threads(16)
///     .build()
///     .unwrap();
/// let pool = SlowJobPool::new(3, 10, Arc::new(threadpool));
/// pool.configure("CHUNK_GENERATOR", |n| n / 2);
/// pool.spawn("CHUNK_GENERATOR", move || println!("this is a job"));
/// ```
#[derive(Clone)]
pub struct SlowJobPool {
    internal: Arc<Mutex<InternalSlowJobPool>>,
}

#[derive(Debug)]
pub struct SlowJob {
    name: String,
    id: u64,
}

type JobType = Box<dyn FnOnce() + Send + Sync + 'static>;

struct InternalSlowJobPool {
    next_id: u64,
    queue: HashMap<String, VecDeque<Queue>>,
    configs: HashMap<String, Config>,
    last_spawned_configs: Vec<String>,
    global_spawned_and_running: u64,
    global_limit: u64,
    jobs_metrics_cnt: usize,
    jobs_metrics: HashMap<String, Vec<JobMetrics>>,
    threadpool: Arc<ThreadPool>,
    internal: Option<Arc<Mutex<Self>>>,
}

#[derive(Debug)]
struct Config {
    local_limit: u64,
    local_spawned_and_running: u64,
}

struct Queue {
    id: u64,
    name: String,
    task: JobType,
}

pub struct JobMetrics {
    pub queue_created: Instant,
    pub execution_start: Instant,
    pub execution_end: Instant,
}

impl Queue {
    fn new<F>(name: &str, id: u64, internal: &Arc<Mutex<InternalSlowJobPool>>, f: F) -> Self
    where
        F: FnOnce() + Send + Sync + 'static,
    {
        let internal = Arc::clone(internal);
        let name_cloned = name.to_owned();
        let queue_created = Instant::now();
        Self {
            id,
            name: name.to_owned(),
            task: Box::new(move || {
                common_base::prof_span_alloc!(_guard, &name_cloned);
                let execution_start = Instant::now();
                f();
                let execution_end = Instant::now();
                let metrics = JobMetrics {
                    queue_created,
                    execution_start,
                    execution_end,
                };
                // directly maintain the next task afterwards
                {
                    let mut lock = internal.lock().expect("slowjob lock poisoned");
                    lock.finish(&name_cloned, metrics);
                    lock.spawn_queued();
                }
            }),
        }
    }
}

impl InternalSlowJobPool {
    pub fn new(
        global_limit: u64,
        jobs_metrics_cnt: usize,
        _threadpool: Arc<ThreadPool>,
    ) -> Arc<Mutex<Self>> {
        // rayon is having a bug where a ECS task could work-steal a slowjob if we use
        // the same threadpool, which would cause lagspikes we dont want!
        let threadpool = Arc::new(
            rayon::ThreadPoolBuilder::new()
                .num_threads(global_limit as usize)
                .thread_name(move |i| format!("slowjob-{}", i))
                .build()
                .unwrap(),
        );
        let link = Arc::new(Mutex::new(Self {
            next_id: 0,
            queue: HashMap::new(),
            configs: HashMap::new(),
            last_spawned_configs: Vec::new(),
            global_spawned_and_running: 0,
            global_limit: global_limit.max(1),
            jobs_metrics_cnt,
            jobs_metrics: HashMap::new(),
            threadpool,
            internal: None,
        }));

        let link_clone = Arc::clone(&link);
        link.lock()
            .expect("poisoned on InternalSlowJobPool::new")
            .internal = Some(link_clone);
        link
    }

    /// returns order of configuration which are queued next
    fn calc_queued_order(
        &self,
        mut queued: HashMap<&String, u64>,
        mut limit: usize,
    ) -> Vec<String> {
        let mut roundrobin = self.last_spawned_configs.clone();
        let mut result = vec![];
        let spawned = self
            .configs
            .iter()
            .map(|(n, c)| (n, c.local_spawned_and_running))
            .collect::<HashMap<_, u64>>();
        let mut queried_capped = self
            .configs
            .iter()
            .map(|(n, c)| {
                (
                    n,
                    queued
                        .get(&n)
                        .cloned()
                        .unwrap_or(0)
                        .min(c.local_limit - c.local_spawned_and_running),
                )
            })
            .collect::<HashMap<_, _>>();
        // grab all configs that are queued and not running. in roundrobin order
        for n in roundrobin.clone().into_iter() {
            if let Some(c) = queued.get_mut(&n)
                && *c > 0
                && spawned.get(&n).cloned().unwrap_or(0) == 0
            {
                result.push(n.clone());
                *c -= 1;
                limit -= 1;
                queried_capped.get_mut(&n).map(|v| *v -= 1);
                roundrobin
                    .iter()
                    .position(|e| e == &n)
                    .map(|i| roundrobin.remove(i));
                roundrobin.push(n);
                if limit == 0 {
                    return result;
                }
            }
        }
        //schedule rest based on their possible limites, don't use round robin here
        let total_limit = queried_capped.values().sum::<u64>() as f32;
        if total_limit < f32::EPSILON {
            return result;
        }
        let mut spawn_rates = queried_capped
            .iter()
            .map(|(&n, l)| (n, ((*l as f32 * limit as f32) / total_limit).min(*l as f32)))
            .collect::<Vec<_>>();
        while limit > 0 {
            spawn_rates.sort_by(|(_, a), (_, b)| {
                if b < a {
                    core::cmp::Ordering::Less
                } else if (b - a).abs() < f32::EPSILON {
                    core::cmp::Ordering::Equal
                } else {
                    core::cmp::Ordering::Greater
                }
            });
            match spawn_rates.first_mut() {
                Some((n, r)) => {
                    if *r > f32::EPSILON {
                        result.push(n.clone());
                        limit -= 1;
                        *r -= 1.0;
                    } else {
                        break;
                    }
                },
                None => break,
            }
        }
        result
    }

    fn can_spawn(&self, name: &str) -> bool {
        let queued = self
            .queue
            .iter()
            .map(|(n, m)| (n, m.len() as u64))
            .collect::<HashMap<_, u64>>();
        let mut to_be_queued = queued.clone();
        let name = name.to_owned();
        *to_be_queued.entry(&name).or_default() += 1;
        let limit = (self.global_limit - self.global_spawned_and_running) as usize;
        // calculate to_be_queued first
        let to_be_queued_order = self.calc_queued_order(to_be_queued, limit);
        let queued_order = self.calc_queued_order(queued, limit);
        // if its queued one time more then its okay to spawn
        let to_be_queued_cnt = to_be_queued_order
            .into_iter()
            .filter(|n| n == &name)
            .count();
        let queued_cnt = queued_order.into_iter().filter(|n| n == &name).count();
        to_be_queued_cnt > queued_cnt
    }

    pub fn spawn<F>(&mut self, name: &str, f: F) -> SlowJob
    where
        F: FnOnce() + Send + Sync + 'static,
    {
        let id = self.next_id;
        self.next_id += 1;
        let queue = Queue::new(name, id, self.internal.as_ref().expect("internal empty"), f);
        self.queue
            .entry(name.to_string())
            .or_default()
            .push_back(queue);
        debug_assert!(
            self.configs.contains_key(name),
            "Can't spawn unconfigured task!"
        );
        //spawn already queued
        self.spawn_queued();
        SlowJob {
            name: name.to_string(),
            id,
        }
    }

    fn finish(&mut self, name: &str, metrics: JobMetrics) {
        let metric = self.jobs_metrics.entry(name.to_string()).or_default();

        if metric.len() < self.jobs_metrics_cnt {
            metric.push(metrics);
        }
        self.global_spawned_and_running -= 1;
        if let Some(c) = self.configs.get_mut(name) {
            c.local_spawned_and_running -= 1;
        } else {
            warn!(?name, "sync_maintain on a no longer existing config");
        }
    }

    fn spawn_queued(&mut self) {
        let queued = self
            .queue
            .iter()
            .map(|(n, m)| (n, m.len() as u64))
            .collect::<HashMap<_, u64>>();
        let limit = self.global_limit as usize;
        let queued_order = self.calc_queued_order(queued, limit);
        for name in queued_order.into_iter() {
            match self.queue.get_mut(&name) {
                Some(deque) => match deque.pop_front() {
                    Some(queue) => {
                        //fire
                        self.global_spawned_and_running += 1;
                        self.configs
                            .get_mut(&queue.name)
                            .expect("cannot fire a unconfigured job")
                            .local_spawned_and_running += 1;
                        self.last_spawned_configs
                            .iter()
                            .position(|e| e == &queue.name)
                            .map(|i| self.last_spawned_configs.remove(i));
                        self.last_spawned_configs.push(queue.name.to_owned());
                        self.threadpool.spawn(queue.task);
                    },
                    None => error!(
                        "internal calculation is wrong, we extected a schedulable job to be \
                         present in the queue"
                    ),
                },
                None => error!(
                    "internal calculation is wrong, we marked a queue as schedulable which \
                     doesn't exist"
                ),
            }
        }
    }

    pub fn take_metrics(&mut self) -> HashMap<String, Vec<JobMetrics>> {
        core::mem::take(&mut self.jobs_metrics)
    }
}

impl SlowJobPool {
    pub fn new(global_limit: u64, jobs_metrics_cnt: usize, threadpool: Arc<ThreadPool>) -> Self {
        Self {
            internal: InternalSlowJobPool::new(global_limit, jobs_metrics_cnt, threadpool),
        }
    }

    /// configure a NAME to spawn up to f(n) threads, depending on how many
    /// threads we globally have available
    pub fn configure<F>(&self, name: &str, f: F)
    where
        F: Fn(u64) -> u64,
    {
        let mut lock = self.internal.lock().expect("lock poisoned while configure");
        let cnf = Config {
            local_limit: f(lock.global_limit).max(1),
            local_spawned_and_running: 0,
        };
        lock.configs.insert(name.to_owned(), cnf);
        lock.last_spawned_configs.push(name.to_owned());
    }

    /// spawn a new slow job on a certain NAME IF it can run immediately
    #[expect(clippy::result_unit_err)]
    pub fn try_run<F>(&self, name: &str, f: F) -> Result<SlowJob, ()>
    where
        F: FnOnce() + Send + Sync + 'static,
    {
        let mut lock = self.internal.lock().expect("lock poisoned while try_run");
        //spawn already queued
        lock.spawn_queued();
        if lock.can_spawn(name) {
            Ok(lock.spawn(name, f))
        } else {
            Err(())
        }
    }

    pub fn spawn<F>(&self, name: &str, f: F) -> SlowJob
    where
        F: FnOnce() + Send + Sync + 'static,
    {
        self.internal
            .lock()
            .expect("lock poisoned while spawn")
            .spawn(name, f)
    }

    pub fn cancel(&self, job: SlowJob) -> Result<(), SlowJob> {
        let mut lock = self.internal.lock().expect("lock poisoned while cancel");
        if let Some(m) = lock.queue.get_mut(&job.name) {
            let p = match m.iter().position(|p| p.id == job.id) {
                Some(p) => p,
                None => return Err(job),
            };
            if m.remove(p).is_some() {
                return Ok(());
            }
        }
        Err(job)
    }

    pub fn take_metrics(&self) -> HashMap<String, Vec<JobMetrics>> {
        self.internal
            .lock()
            .expect("lock poisoned while take_metrics")
            .take_metrics()
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::{
        sync::{
            Barrier,
            atomic::{AtomicBool, AtomicU64, Ordering},
        },
        time::Duration,
    };

    fn mock_pool(
        pool_threads: usize,
        global_threads: u64,
        metrics: usize,
        foo: u64,
        bar: u64,
        baz: u64,
    ) -> SlowJobPool {
        let threadpool = rayon::ThreadPoolBuilder::new()
            .num_threads(pool_threads)
            .build()
            .unwrap();
        let pool = SlowJobPool::new(global_threads, metrics, Arc::new(threadpool));
        if foo != 0 {
            pool.configure("FOO", |x| x / foo);
        }
        if bar != 0 {
            pool.configure("BAR", |x| x / bar);
        }
        if baz != 0 {
            pool.configure("BAZ", |x| x / baz);
        }
        pool
    }

    #[test]
    fn simple_queue() {
        let pool = mock_pool(4, 4, 0, 1, 0, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 1u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 4);
        assert_eq!(result.len(), 1);
        assert_eq!(result[0], "FOO");
    }

    #[test]
    fn multiple_queue() {
        let pool = mock_pool(4, 4, 0, 1, 0, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 2u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 4);
        assert_eq!(result.len(), 2);
        assert_eq!(result[0], "FOO");
        assert_eq!(result[1], "FOO");
    }

    #[test]
    fn limit_queue() {
        let pool = mock_pool(5, 5, 0, 1, 0, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 80u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 4);
        assert_eq!(result.len(), 4);
        assert_eq!(result[0], "FOO");
        assert_eq!(result[1], "FOO");
        assert_eq!(result[2], "FOO");
        assert_eq!(result[3], "FOO");
    }

    #[test]
    fn simple_queue_2() {
        let pool = mock_pool(4, 4, 0, 1, 1, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 1u64), ("BAR", 1u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 4);
        assert_eq!(result.len(), 2);
        assert_eq!(result.iter().filter(|&x| x == "FOO").count(), 1);
        assert_eq!(result.iter().filter(|&x| x == "BAR").count(), 1);
    }

    #[test]
    fn multiple_queue_3() {
        let pool = mock_pool(4, 4, 0, 1, 1, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 2u64), ("BAR", 2u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 4);
        assert_eq!(result.len(), 4);
        assert_eq!(result.iter().filter(|&x| x == "FOO").count(), 2);
        assert_eq!(result.iter().filter(|&x| x == "BAR").count(), 2);
    }

    #[test]
    fn multiple_queue_4() {
        let pool = mock_pool(4, 4, 0, 2, 1, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 3u64), ("BAR", 3u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 4);
        assert_eq!(result.len(), 4);
        assert_eq!(result.iter().filter(|&x| x == "FOO").count(), 2);
        assert_eq!(result.iter().filter(|&x| x == "BAR").count(), 2);
    }

    #[test]
    fn multiple_queue_5() {
        let pool = mock_pool(4, 4, 0, 2, 1, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 5u64), ("BAR", 5u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 5);
        assert_eq!(result.len(), 5);
        assert_eq!(result.iter().filter(|&x| x == "FOO").count(), 2);
        assert_eq!(result.iter().filter(|&x| x == "BAR").count(), 3);
    }

    #[test]
    fn multiple_queue_6() {
        let pool = mock_pool(40, 40, 0, 2, 1, 0);
        let internal = pool.internal.lock().unwrap();
        let queue_data = [("FOO", 5u64), ("BAR", 5u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        let result = internal.calc_queued_order(queued, 11);
        assert_eq!(result.len(), 10);
        assert_eq!(result.iter().filter(|&x| x == "FOO").count(), 5);
        assert_eq!(result.iter().filter(|&x| x == "BAR").count(), 5);
    }

    #[test]
    fn roundrobin() {
        let pool = mock_pool(4, 4, 0, 2, 2, 0);
        let queue_data = [("FOO", 5u64), ("BAR", 5u64)]
            .iter()
            .map(|(n, c)| ((*n).to_owned(), *c))
            .collect::<Vec<_>>();
        let queued = queue_data
            .iter()
            .map(|(s, c)| (s, *c))
            .collect::<HashMap<_, _>>();
        // Spawn a FOO task.
        pool.internal
            .lock()
            .unwrap()
            .spawn("FOO", || println!("foo"));
        // a barrier in f doesnt work as we need to wait for the cleanup
        while pool.internal.lock().unwrap().global_spawned_and_running != 0 {
            std::thread::yield_now();
        }
        let result = pool
            .internal
            .lock()
            .unwrap()
            .calc_queued_order(queued.clone(), 1);
        assert_eq!(result.len(), 1);
        assert_eq!(result[0], "BAR");
        // keep order if no new is spawned
        let result = pool
            .internal
            .lock()
            .unwrap()
            .calc_queued_order(queued.clone(), 1);
        assert_eq!(result.len(), 1);
        assert_eq!(result[0], "BAR");
        // spawn a BAR task
        pool.internal
            .lock()
            .unwrap()
            .spawn("BAR", || println!("bar"));
        while pool.internal.lock().unwrap().global_spawned_and_running != 0 {
            std::thread::yield_now();
        }
        let result = pool.internal.lock().unwrap().calc_queued_order(queued, 1);
        assert_eq!(result.len(), 1);
        assert_eq!(result[0], "FOO");
    }

    #[test]
    #[should_panic]
    fn unconfigured() {
        let pool = mock_pool(4, 4, 0, 2, 1, 0);
        let mut internal = pool.internal.lock().unwrap();
        internal.spawn("UNCONFIGURED", || println!());
    }

    #[test]
    fn correct_spawn_doesnt_panic() {
        let pool = mock_pool(4, 4, 0, 2, 1, 0);
        let mut internal = pool.internal.lock().unwrap();
        internal.spawn("FOO", || println!("foo"));
        internal.spawn("BAR", || println!("bar"));
    }

    #[test]
    fn can_spawn() {
        let pool = mock_pool(4, 4, 0, 2, 1, 0);
        let internal = pool.internal.lock().unwrap();
        assert!(internal.can_spawn("FOO"));
        assert!(internal.can_spawn("BAR"));
    }

    #[test]
    fn try_run_works() {
        let pool = mock_pool(4, 4, 0, 2, 1, 0);
        pool.try_run("FOO", || println!("foo")).unwrap();
        pool.try_run("BAR", || println!("bar")).unwrap();
    }

    #[test]
    fn try_run_exhausted() {
        let pool = mock_pool(8, 8, 0, 4, 2, 0);
        let func = || loop {
            std::thread::sleep(Duration::from_secs(1))
        };
        pool.try_run("FOO", func).unwrap();
        pool.try_run("BAR", func).unwrap();
        pool.try_run("FOO", func).unwrap();
        pool.try_run("BAR", func).unwrap();
        pool.try_run("FOO", func).unwrap_err();
        pool.try_run("BAR", func).unwrap();
        pool.try_run("FOO", func).unwrap_err();
        pool.try_run("BAR", func).unwrap();
        pool.try_run("FOO", func).unwrap_err();
        pool.try_run("BAR", func).unwrap_err();
        pool.try_run("FOO", func).unwrap_err();
    }

    #[test]
    fn actually_runs_1() {
        let pool = mock_pool(4, 4, 0, 0, 0, 1);
        let barrier = Arc::new(Barrier::new(2));
        let barrier_clone = Arc::clone(&barrier);
        pool.try_run("BAZ", move || {
            barrier_clone.wait();
        })
        .unwrap();
        barrier.wait();
    }

    #[test]
    fn actually_runs_2() {
        let pool = mock_pool(4, 4, 0, 0, 0, 1);
        let barrier = Arc::new(Barrier::new(2));
        let barrier_clone = Arc::clone(&barrier);
        pool.spawn("BAZ", move || {
            barrier_clone.wait();
        });
        barrier.wait();
    }

    #[test]
    fn actually_waits() {
        let pool = mock_pool(4, 4, 0, 4, 0, 1);
        let ops_i_ran = Arc::new(AtomicBool::new(false));
        let ops_i_ran_clone = Arc::clone(&ops_i_ran);
        let barrier = Arc::new(Barrier::new(2));
        let barrier_clone = Arc::clone(&barrier);
        let barrier2 = Arc::new(Barrier::new(2));
        let barrier2_clone = Arc::clone(&barrier2);
        pool.try_run("FOO", move || {
            barrier_clone.wait();
        })
        .unwrap();
        pool.spawn("FOO", move || {
            ops_i_ran_clone.store(true, Ordering::SeqCst);
            barrier2_clone.wait();
        });
        // in this case we have to sleep
        std::thread::sleep(Duration::from_secs(1));
        assert!(!ops_i_ran.load(Ordering::SeqCst));
        // now finish the first job
        barrier.wait();
        // now wait on the second job to be actually finished
        barrier2.wait();
    }

    #[test]
    fn verify_metrics() {
        let pool = mock_pool(4, 4, 2, 1, 0, 4);
        let barrier = Arc::new(Barrier::new(5));
        for name in &["FOO", "BAZ", "FOO", "FOO"] {
            let barrier_clone = Arc::clone(&barrier);
            pool.spawn(name, move || {
                barrier_clone.wait();
            });
        }
        // now finish all jobs
        barrier.wait();
        // in this case we have to sleep to give it some time to store all the metrics
        std::thread::sleep(Duration::from_secs(2));
        let metrics = pool.take_metrics();
        let foo = metrics.get("FOO").expect("FOO doesn't exist in metrics");
        //its limited to 2, even though we had 3 jobs
        assert_eq!(foo.len(), 2);
        assert!(metrics.get("BAR").is_none());
        let baz = metrics.get("BAZ").expect("BAZ doesn't exist in metrics");
        assert_eq!(baz.len(), 1);
    }

    fn work_barrier(counter: &Arc<AtomicU64>, ms: u64) -> impl std::ops::FnOnce() -> () + use<> {
        let counter = Arc::clone(counter);
        println!("Create work_barrier");
        move || {
            println!(".{}..", ms);
            std::thread::sleep(Duration::from_millis(ms));
            println!(".{}..Done", ms);
            counter.fetch_add(1, Ordering::SeqCst);
        }
    }

    #[test]
    fn verify_that_spawn_doesnt_block_par_iter() {
        let threadpool = Arc::new(
            rayon::ThreadPoolBuilder::new()
                .num_threads(20)
                .build()
                .unwrap(),
        );
        let pool = SlowJobPool::new(2, 100, Arc::<rayon::ThreadPool>::clone(&threadpool));
        pool.configure("BAZ", |_| 2);
        let counter = Arc::new(AtomicU64::new(0));
        let start = Instant::now();

        threadpool.install(|| {
            use rayon::prelude::*;
            (0..100)
                .into_par_iter()
                .map(|i| {
                    std::thread::sleep(Duration::from_millis(10));
                    if i == 50 {
                        pool.spawn("BAZ", work_barrier(&counter, 2000));
                    }
                    if i == 99 {
                        println!("The first ITER end, at {}ms", start.elapsed().as_millis());
                    }
                })
                .collect::<Vec<_>>();
            let elapsed = start.elapsed().as_millis();
            println!("The first ITER finished, at {}ms", elapsed);
            assert!(
                elapsed < 1900,
                "It seems like the par_iter waited on the 2s sleep task to finish"
            );
        });

        while counter.load(Ordering::SeqCst) == 0 {
            println!("waiting for BAZ task to finish");
            std::thread::sleep(Duration::from_secs(1));
        }
    }
}