datum/dynamic/

hub.rs

use std::{
    collections::{BTreeMap, VecDeque},
    fmt,
    sync::{
        Arc, Condvar, Mutex, MutexGuard,
        atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering},
    },
    time::Duration,
};

use arc_swap::ArcSwap;
use smallvec::SmallVec;

use crate::stream::{BoxStream, NotUsed, Sink, Source, SourceRuntimeHints, StreamCompletion};
use crate::{StreamError, StreamResult};

type Partitioner<T> = Arc<dyn Fn(&PartitionConsumerInfo, &T) -> isize + Send + Sync>;

const MERGE_HUB_BATCH_LIMIT: usize = 256;
const BROADCAST_HUB_BATCH_LIMIT: usize = 256;
const BROADCAST_HUB_SINGLE_CONSUMER_BATCH_LIMIT: usize = 64;
const PARTITION_HUB_BATCH_LIMIT: usize = 1024;
const PARTITION_HUB_WIDE_BATCH_LIMIT: usize = 2048;
const PARTITION_HUB_SINGLE_CONSUMER_BATCH_LIMIT: usize = 256;
const FAN_OUT_CONSUMER_BATCH_LIMIT: usize = 256;

fn fan_out_wait_timeout() -> Duration {
    Duration::from_millis(1)
}

#[derive(Clone)]
pub struct MergeHubDrainingControl {
    state: Arc<MergeHubState>,
    on_drain: Arc<dyn Fn() + Send + Sync>,
}

impl fmt::Debug for MergeHubDrainingControl {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("MergeHubDrainingControl").finish()
    }
}

impl MergeHubDrainingControl {
    pub fn drain_and_complete(&self) {
        let mut state = self.state.lock();
        state.draining = true;
        self.state.condvar.notify_all();
        drop(state);
        (self.on_drain)();
    }
}

pub struct MergeHub;

impl MergeHub {
    /// Materializes a reusable sink that lets many producers attach to one running downstream.
    #[must_use]
    pub fn source<T: Send + 'static>(
        per_producer_buffer_size: usize,
    ) -> Source<T, Sink<T, NotUsed>> {
        Self::source_with_draining(per_producer_buffer_size)
            .map_materialized_value(|(sink, _)| sink)
    }

    /// Materializes a reusable sink plus a control that stops accepting new producers and
    /// completes once currently attached producers finish.
    #[must_use]
    pub fn source_with_draining<T: Send + 'static>(
        per_producer_buffer_size: usize,
    ) -> Source<T, (Sink<T, NotUsed>, MergeHubDrainingControl)> {
        assert!(
            per_producer_buffer_size > 0,
            "MergeHub per_producer_buffer_size must be greater than zero"
        );
        Source::from_terminal_batch_materialized_factory(move |_| {
            let state = Arc::new(MergeHubShared::<T>::new(per_producer_buffer_size));
            let source = Box::new(MergeHubSourceStream {
                state: Arc::clone(&state),
                local: VecDeque::new(),
                prefer_direct: false,
            }) as BoxStream<T>;
            let sink = merge_hub_sink(Arc::clone(&state));
            let control = MergeHubDrainingControl {
                state: Arc::clone(&state.state),
                on_drain: Arc::new({
                    let state = Arc::clone(&state);
                    move || state.finish_if_draining()
                }),
            };
            Ok((source, (sink, control)))
        })
    }
}

pub struct BroadcastHub;

impl BroadcastHub {
    /// Materializes a reusable source for dynamically attaching consumers.
    ///
    /// As in Akka, one producer adapts to the slowest active consumer unless callers add their own
    /// buffering or dropping stages around the materialized consumer sources.
    ///
    /// Unlike Akka, Datum blocks upstream immediately when there are zero consumers instead of
    /// pre-buffering up to `buffer_size` elements.
    #[must_use]
    pub fn sink<T: Clone + Send + 'static>(
        buffer_size: usize,
    ) -> Sink<T, BroadcastHubConsumerSource<T>> {
        Self::sink_starting_after(0, buffer_size)
    }

    /// Delays upstream consumption until at least `start_after_nr_of_consumers` consumers attach.
    #[must_use]
    pub fn sink_starting_after<T: Clone + Send + 'static>(
        start_after_nr_of_consumers: usize,
        buffer_size: usize,
    ) -> Sink<T, BroadcastHubConsumerSource<T>> {
        assert!(
            buffer_size > 0,
            "BroadcastHub buffer_size must be greater than zero"
        );
        Sink::from_hinted_runner(move |input, materializer, hints| {
            let state = Arc::new(FanOutHubShared::new(
                FanOutMode::Broadcast,
                start_after_nr_of_consumers,
                buffer_size,
                None::<Partitioner<T>>,
            ));
            let source = BroadcastHubConsumerSource {
                state: Arc::clone(&state),
                completion: Arc::new(Mutex::new(None)),
            };
            let completion = materializer.spawn_stream(move |cancelled| {
                FanOutProducer::new(input, state).run(cancelled, hints)
            });
            source.attach_completion(completion);
            Ok(source)
        })
    }
}

pub struct PartitionHub;

impl PartitionHub {
    /// Materializes a reusable source whose elements are routed to one selected consumer.
    ///
    /// The producer still adapts to the slowest active consumer queue unless callers add their own
    /// buffering or dropping stages around the materialized consumer sources.
    ///
    /// The partitioner sees a consistent snapshot of currently attached consumers and queued
    /// elements, but it is not run while Datum's internal hub lock is held.
    #[must_use]
    pub fn sink<T: Clone + Send + 'static, F>(
        partitioner: F,
        start_after_nr_of_consumers: usize,
        buffer_size: usize,
    ) -> Sink<T, PartitionHubConsumerSource<T>>
    where
        F: Fn(&PartitionConsumerInfo, &T) -> isize + Send + Sync + 'static,
    {
        assert!(
            buffer_size > 0,
            "PartitionHub buffer_size must be greater than zero"
        );
        let partitioner = Arc::new(partitioner);
        Sink::from_hinted_runner(move |input, materializer, hints| {
            let partitioner = Arc::clone(&partitioner);
            let state = Arc::new(FanOutHubShared::new(
                FanOutMode::Partition,
                start_after_nr_of_consumers,
                buffer_size,
                Some(partitioner),
            ));
            let source = PartitionHubConsumerSource {
                state: Arc::clone(&state),
                completion: Arc::new(Mutex::new(None)),
            };
            let completion = materializer.spawn_stream(move |cancelled| {
                FanOutProducer::new(input, state).run(cancelled, hints)
            });
            source.attach_completion(completion);
            Ok(source)
        })
    }
}

#[derive(Clone)]
pub struct BroadcastHubConsumerSource<T> {
    state: Arc<FanOutHubShared<T>>,
    completion: Arc<Mutex<Option<StreamCompletion<NotUsed>>>>,
}

impl<T: Clone + Send + 'static> BroadcastHubConsumerSource<T> {
    fn attach_completion(&self, completion: StreamCompletion<NotUsed>) {
        *self
            .completion
            .lock()
            .expect("broadcast hub completion poisoned") = Some(completion);
    }

    #[must_use]
    pub fn source(&self) -> Source<T, NotUsed> {
        let state = Arc::clone(&self.state);
        Source::from_materialized_factory(move |_| {
            let lane = state.register_consumer();
            let stream = Box::new(FanOutConsumerStream {
                state: Arc::clone(&state),
                lane,
                local: None,
                detached: false,
            }) as BoxStream<T>;
            Ok((stream, NotUsed))
        })
    }
}

impl<T: Clone + Send + 'static> fmt::Debug for BroadcastHubConsumerSource<T> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("BroadcastHubConsumerSource").finish()
    }
}

#[derive(Clone)]
pub struct PartitionHubConsumerSource<T> {
    state: Arc<FanOutHubShared<T>>,
    completion: Arc<Mutex<Option<StreamCompletion<NotUsed>>>>,
}

impl<T: Clone + Send + 'static> PartitionHubConsumerSource<T> {
    fn attach_completion(&self, completion: StreamCompletion<NotUsed>) {
        *self
            .completion
            .lock()
            .expect("partition hub completion poisoned") = Some(completion);
    }

    #[must_use]
    pub fn source(&self) -> Source<T, NotUsed> {
        let state = Arc::clone(&self.state);
        Source::from_materialized_factory(move |_| {
            let lane = state.register_consumer();
            let stream = Box::new(FanOutConsumerStream {
                state: Arc::clone(&state),
                lane,
                local: None,
                detached: false,
            }) as BoxStream<T>;
            Ok((stream, NotUsed))
        })
    }
}

impl<T: Clone + Send + 'static> fmt::Debug for PartitionHubConsumerSource<T> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("PartitionHubConsumerSource").finish()
    }
}

#[derive(Clone, Debug)]
pub struct PartitionConsumerInfo {
    consumer_ids: SmallVec<[u64; 16]>,
    queue_sizes: SmallVec<[(u64, usize); 16]>,
}

impl PartitionConsumerInfo {
    #[must_use]
    pub fn size(&self) -> usize {
        self.consumer_ids.len()
    }

    #[must_use]
    pub fn consumer_ids(&self) -> &[u64] {
        &self.consumer_ids
    }

    #[must_use]
    pub fn consumer_id_by_idx(&self, idx: usize) -> u64 {
        self.consumer_ids[idx]
    }

    #[must_use]
    pub fn queue_size(&self, consumer_id: u64) -> usize {
        self.queue_sizes
            .iter()
            .find_map(|(id, size)| (*id == consumer_id).then_some(*size))
            .unwrap_or(0)
    }
}

fn merge_hub_sink<T: Send + 'static>(state: Arc<MergeHubShared<T>>) -> Sink<T, NotUsed> {
    Sink::from_raw_hinted_runner(move |input, materializer, hints| {
        if hints.inline_micro_max_success_items.is_some() {
            state.register_direct_producer(input)?;
            return Ok(NotUsed);
        }

        let input = materializer.checked_stream(input, None);
        let producer_id = state.register_producer()?;
        let hub = Arc::clone(&state);
        let completion = materializer.spawn_stream(move |cancelled| {
            let mut input = input;
            loop {
                if cancelled.load(std::sync::atomic::Ordering::SeqCst) {
                    hub.fail(StreamError::Cancelled);
                    hub.deregister_producer(producer_id);
                    return Err(StreamError::Cancelled);
                }

                match input.next() {
                    Some(Ok(item)) => hub.push_item(producer_id, item)?,
                    Some(Err(error)) => {
                        hub.fail(error.clone());
                        hub.deregister_producer(producer_id);
                        return Err(error);
                    }
                    None => {
                        hub.deregister_producer(producer_id);
                        return Ok(NotUsed);
                    }
                }
            }
        });
        state.store_producer_completion(completion);
        Ok(NotUsed)
    })
}

struct MergeHubShared<T> {
    state: Arc<MergeHubState>,
    shared: Mutex<MergeHubInner<T>>,
    condvar: Condvar,
    failed: AtomicBool,
}

#[derive(Debug)]
struct MergeHubState {
    inner: Mutex<MergeHubFlags>,
    condvar: Condvar,
}

#[derive(Debug, Default)]
struct MergeHubFlags {
    draining: bool,
}

impl MergeHubState {
    fn lock(&self) -> MutexGuard<'_, MergeHubFlags> {
        self.inner.lock().expect("merge hub flags poisoned")
    }
}

struct MergeHubInner<T> {
    queue: VecDeque<(u64, T)>,
    direct_producers: VecDeque<MergeHubDirectProducer<T>>,
    queued_per_producer: BTreeMap<u64, usize>,
    producer_completions: Vec<StreamCompletion<NotUsed>>,
    active_producers: usize,
    next_producer_id: u64,
    source_closed: bool,
    completed: bool,
    failed: Option<StreamError>,
    per_producer_buffer_size: usize,
}

struct MergeHubDirectProducer<T> {
    id: u64,
    input: BoxStream<T>,
}

enum MergeHubProducerTerminal {
    Active,
    Completed,
    Failed(StreamError),
}

impl<T> MergeHubShared<T> {
    fn new(per_producer_buffer_size: usize) -> Self {
        Self {
            state: Arc::new(MergeHubState {
                inner: Mutex::new(MergeHubFlags::default()),
                condvar: Condvar::new(),
            }),
            shared: Mutex::new(MergeHubInner {
                queue: VecDeque::new(),
                direct_producers: VecDeque::new(),
                queued_per_producer: BTreeMap::new(),
                producer_completions: Vec::new(),
                active_producers: 0,
                next_producer_id: 0,
                source_closed: false,
                completed: false,
                failed: None,
                per_producer_buffer_size,
            }),
            condvar: Condvar::new(),
            failed: AtomicBool::new(false),
        }
    }

    fn register_direct_producer(&self, input: BoxStream<T>) -> StreamResult<()> {
        let mut inner = self.shared.lock().expect("merge hub poisoned");
        prune_finished_producer_completions(&mut inner.producer_completions);
        let flags = self.state.lock();
        if flags.draining || inner.source_closed || inner.completed {
            return Err(StreamError::Failed(
                "merge hub is draining or closed to new producers".to_owned(),
            ));
        }
        if let Some(error) = inner.failed.clone() {
            return Err(error);
        }
        let id = inner.next_producer_id;
        inner.next_producer_id += 1;
        inner.active_producers += 1;
        inner
            .direct_producers
            .push_back(MergeHubDirectProducer { id, input });
        drop(flags);
        drop(inner);
        self.condvar.notify_all();
        Ok(())
    }

    fn register_producer(&self) -> StreamResult<u64> {
        let mut inner = self.shared.lock().expect("merge hub poisoned");
        prune_finished_producer_completions(&mut inner.producer_completions);
        let flags = self.state.lock();
        if flags.draining || inner.source_closed || inner.completed {
            return Err(StreamError::Failed(
                "merge hub is draining or closed to new producers".to_owned(),
            ));
        }
        if let Some(error) = inner.failed.clone() {
            return Err(error);
        }
        let id = inner.next_producer_id;
        inner.next_producer_id += 1;
        inner.active_producers += 1;
        inner.queued_per_producer.insert(id, 0);
        Ok(id)
    }

    fn store_producer_completion(&self, completion: StreamCompletion<NotUsed>) {
        let mut inner = self.shared.lock().expect("merge hub poisoned");
        prune_finished_producer_completions(&mut inner.producer_completions);
        inner.producer_completions.push(completion);
    }

    fn push_item(&self, producer_id: u64, item: T) -> StreamResult<()> {
        let mut inner = self.shared.lock().expect("merge hub poisoned");
        loop {
            if let Some(error) = inner.failed.clone() {
                inner.queued_per_producer.remove(&producer_id);
                return Err(error);
            }
            if inner.source_closed {
                inner.queued_per_producer.remove(&producer_id);
                return Err(StreamError::Cancelled);
            }
            let queued = inner
                .queued_per_producer
                .get(&producer_id)
                .copied()
                .unwrap_or(0);
            if queued < inner.per_producer_buffer_size {
                inner.queue.push_back((producer_id, item));
                inner.queued_per_producer.insert(producer_id, queued + 1);
                self.condvar.notify_all();
                return Ok(());
            }
            inner = self
                .condvar
                .wait(inner)
                .expect("merge hub poisoned while waiting");
        }
    }

    fn deregister_producer(&self, producer_id: u64) {
        let mut inner = self.shared.lock().expect("merge hub poisoned");
        prune_finished_producer_completions(&mut inner.producer_completions);
        inner.queued_per_producer.remove(&producer_id);
        inner
            .direct_producers
            .retain(|producer| producer.id != producer_id);
        inner.active_producers = inner.active_producers.saturating_sub(1);
        if inner.active_producers == 0 {
            let flags = self.state.lock();
            if flags.draining {
                inner.completed = true;
            }
        }
        self.condvar.notify_all();
    }

    fn pop_direct_producer(
        &self,
        inner: &mut MergeHubInner<T>,
    ) -> Option<MergeHubDirectProducer<T>> {
        inner.direct_producers.pop_front()
    }

    fn restore_direct_producer(
        &self,
        producer: MergeHubDirectProducer<T>,
        terminal: MergeHubProducerTerminal,
    ) -> StreamResult<()> {
        match terminal {
            MergeHubProducerTerminal::Active => {
                let mut inner = self.shared.lock().expect("merge hub poisoned");
                if let Some(error) = inner.failed.clone() {
                    inner.active_producers = inner.active_producers.saturating_sub(1);
                    return Err(error);
                }
                if inner.source_closed {
                    inner.active_producers = inner.active_producers.saturating_sub(1);
                    return Err(StreamError::Cancelled);
                }
                inner.direct_producers.push_back(producer);
                Ok(())
            }
            MergeHubProducerTerminal::Completed => {
                self.deregister_producer(producer.id);
                Ok(())
            }
            MergeHubProducerTerminal::Failed(error) => {
                self.fail(error.clone());
                self.deregister_producer(producer.id);
                Err(error)
            }
        }
    }

    fn fail(&self, error: StreamError) {
        let mut inner = self.shared.lock().expect("merge hub poisoned");
        if inner.failed.is_none() {
            inner.failed = Some(error);
            self.failed.store(true, Ordering::SeqCst);
        }
        self.condvar.notify_all();
    }

    fn failed_error(&self) -> Option<StreamError> {
        if !self.failed.load(Ordering::SeqCst) {
            return None;
        }
        self.shared
            .lock()
            .expect("merge hub poisoned")
            .failed
            .clone()
    }

    fn finish_if_draining(&self) {
        let flags = self.state.lock();
        if !flags.draining {
            return;
        }
        drop(flags);

        let mut inner = self.shared.lock().expect("merge hub poisoned");
        prune_finished_producer_completions(&mut inner.producer_completions);
        if inner.active_producers == 0 {
            inner.completed = true;
            self.condvar.notify_all();
        }
    }
}

fn prune_finished_producer_completions(completions: &mut Vec<StreamCompletion<NotUsed>>) {
    let mut index = 0;
    while index < completions.len() {
        if completions[index].try_wait().is_some() {
            drop(completions.swap_remove(index));
        } else {
            index += 1;
        }
    }
}

struct MergeHubSourceStream<T> {
    state: Arc<MergeHubShared<T>>,
    local: VecDeque<T>,
    prefer_direct: bool,
}

impl<T> Iterator for MergeHubSourceStream<T> {
    type Item = StreamResult<T>;

    fn next(&mut self) -> Option<Self::Item> {
        if let Some(error) = self.state.failed_error() {
            self.local.clear();
            let mut inner = self.state.shared.lock().expect("merge hub poisoned");
            inner.source_closed = true;
            return Some(Err(error));
        }
        if let Some(item) = self.local.pop_front() {
            return Some(Ok(item));
        }

        let mut inner = self.state.shared.lock().expect("merge hub poisoned");
        loop {
            if let Some(error) = inner.failed.clone() {
                inner.source_closed = true;
                return Some(Err(error));
            }
            let should_drain_direct = !inner.direct_producers.is_empty()
                && (self.prefer_direct || inner.queue.is_empty());
            if should_drain_direct {
                let Some(mut producer) = self.state.pop_direct_producer(&mut inner) else {
                    continue;
                };
                let drain_limit = inner
                    .per_producer_buffer_size
                    .clamp(1, MERGE_HUB_BATCH_LIMIT);
                drop(inner);

                let mut batch = std::mem::take(&mut self.local);
                batch.clear();
                batch.reserve(drain_limit.saturating_sub(batch.capacity()));
                let mut terminal = MergeHubProducerTerminal::Active;
                for _ in 0..drain_limit {
                    match producer.input.next() {
                        Some(Ok(item)) => batch.push_back(item),
                        Some(Err(error)) => {
                            terminal = MergeHubProducerTerminal::Failed(error);
                            break;
                        }
                        None => {
                            terminal = MergeHubProducerTerminal::Completed;
                            break;
                        }
                    }
                }

                let restore_result = self.state.restore_direct_producer(producer, terminal);
                match restore_result {
                    Ok(()) => {
                        self.prefer_direct = false;
                        if let Some(first) = batch.pop_front() {
                            self.local = batch;
                            return Some(Ok(first));
                        }
                        inner = self.state.shared.lock().expect("merge hub poisoned");
                        continue;
                    }
                    Err(error) => {
                        self.local.clear();
                        return Some(Err(error));
                    }
                }
            }
            if !inner.queue.is_empty() {
                let drain_n = inner.queue.len().min(MERGE_HUB_BATCH_LIMIT);
                let mut batch = std::mem::take(&mut self.local);
                batch.clear();
                batch.reserve(drain_n.saturating_sub(batch.capacity()));
                for _ in 0..drain_n {
                    if let Some((producer_id, item)) = inner.queue.pop_front() {
                        if let Some(queued) = inner.queued_per_producer.get_mut(&producer_id) {
                            *queued = queued.saturating_sub(1);
                        }
                        batch.push_back(item);
                    }
                }
                self.state.condvar.notify_all();
                drop(inner);
                let first = batch
                    .pop_front()
                    .expect("merge hub drained non-empty batch");
                self.local = batch;
                self.prefer_direct = true;
                return Some(Ok(first));
            }
            if inner.completed {
                inner.source_closed = true;
                return None;
            }
            inner = self
                .state
                .condvar
                .wait(inner)
                .expect("merge hub poisoned while waiting");
        }
    }
}

impl<T> Drop for MergeHubSourceStream<T> {
    fn drop(&mut self) {
        let mut inner = self.state.shared.lock().expect("merge hub poisoned");
        inner.source_closed = true;
        self.state.condvar.notify_all();
    }
}

#[derive(Clone, Copy)]
enum FanOutMode {
    Broadcast,
    Partition,
}

struct FanOutHubShared<T> {
    registry: Mutex<FanOutRegistry<T>>,
    snapshot: ArcSwap<FanOutSnapshot<T>>,
    producer_wait: Mutex<()>,
    producer_condvar: Condvar,
    producer_epoch: AtomicU64,
    topology_epoch: AtomicU64,
    mode: FanOutMode,
    start_after_nr_of_consumers: usize,
    buffer_size: usize,
    partitioner: Option<Partitioner<T>>,
}

struct FanOutRegistry<T> {
    consumers: BTreeMap<u64, Arc<FanOutConsumerLane<T>>>,
    next_consumer_id: u64,
    terminal: Option<FanOutTerminal>,
}

struct FanOutSnapshot<T> {
    consumers: Vec<Arc<FanOutConsumerLane<T>>>,
    terminal: Option<FanOutTerminal>,
}

#[derive(Clone, Debug)]
enum FanOutTerminal {
    Completed,
    Failed(StreamError),
}

impl FanOutTerminal {
    fn producer_error(&self) -> StreamError {
        match self {
            Self::Completed => StreamError::Cancelled,
            Self::Failed(error) => error.clone(),
        }
    }
}

struct FanOutConsumerLane<T> {
    id: u64,
    state: Mutex<FanOutLaneState<T>>,
    condvar: Condvar,
    queued: AtomicUsize,
    active: AtomicBool,
    failed: AtomicBool,
}

struct FanOutLaneState<T> {
    chunks: VecDeque<Vec<T>>,
    queued: usize,
    terminal: Option<FanOutTerminal>,
}

impl<T> FanOutConsumerLane<T> {
    fn new(id: u64, buffer_size: usize) -> Self {
        Self {
            id,
            state: Mutex::new(FanOutLaneState {
                chunks: VecDeque::with_capacity((buffer_size / FAN_OUT_CONSUMER_BATCH_LIMIT) + 1),
                queued: 0,
                terminal: None,
            }),
            condvar: Condvar::new(),
            queued: AtomicUsize::new(0),
            active: AtomicBool::new(true),
            failed: AtomicBool::new(false),
        }
    }

    fn terminal(id: u64, terminal: FanOutTerminal) -> Self {
        let failed = matches!(terminal, FanOutTerminal::Failed(_));
        Self {
            id,
            state: Mutex::new(FanOutLaneState {
                chunks: VecDeque::new(),
                queued: 0,
                terminal: Some(terminal),
            }),
            condvar: Condvar::new(),
            queued: AtomicUsize::new(0),
            active: AtomicBool::new(false),
            failed: AtomicBool::new(failed),
        }
    }

    fn id(&self) -> u64 {
        self.id
    }

    fn queued_len(&self) -> usize {
        self.queued.load(Ordering::Acquire)
    }

    fn is_active(&self) -> bool {
        self.active.load(Ordering::Acquire)
    }

    fn deactivate(&self) {
        self.active.store(false, Ordering::Release);
        self.condvar.notify_all();
    }

    fn set_terminal(&self, terminal: FanOutTerminal) {
        let mut state = self.state.lock().expect("fan-out lane poisoned");
        if matches!(terminal, FanOutTerminal::Failed(_)) {
            state.chunks.clear();
            state.queued = 0;
            self.queued.store(0, Ordering::Release);
            self.failed.store(true, Ordering::Release);
        }
        state.terminal = Some(terminal);
        drop(state);
        self.condvar.notify_all();
    }
}

impl PartitionConsumerInfo {
    fn from_cached_lanes<T>(consumer_ids: &[u64], lanes: &[Arc<FanOutConsumerLane<T>>]) -> Self {
        Self {
            consumer_ids: consumer_ids.iter().copied().collect(),
            queue_sizes: lanes
                .iter()
                .map(|lane| (lane.id(), lane.queued_len()))
                .collect(),
        }
    }
}

struct PartitionTopologyCache<T> {
    epoch: u64,
    consumer_ids: SmallVec<[u64; 16]>,
    lanes: SmallVec<[Arc<FanOutConsumerLane<T>>; 16]>,
}

type PartitionRoutedBatches<T> = SmallVec<[(Arc<FanOutConsumerLane<T>>, VecDeque<T>); 16]>;

impl<T> PartitionTopologyCache<T> {
    fn new() -> Self {
        Self {
            epoch: u64::MAX,
            consumer_ids: SmallVec::new(),
            lanes: SmallVec::new(),
        }
    }

    fn clear(&mut self) {
        self.epoch = u64::MAX;
        self.consumer_ids.clear();
        self.lanes.clear();
    }
}

impl<T> FanOutHubShared<T> {
    fn new(
        mode: FanOutMode,
        start_after_nr_of_consumers: usize,
        buffer_size: usize,
        partitioner: Option<Partitioner<T>>,
    ) -> Self {
        Self {
            registry: Mutex::new(FanOutRegistry {
                consumers: BTreeMap::new(),
                next_consumer_id: 0,
                terminal: None,
            }),
            snapshot: ArcSwap::from_pointee(FanOutSnapshot {
                consumers: Vec::new(),
                terminal: None,
            }),
            producer_wait: Mutex::new(()),
            producer_condvar: Condvar::new(),
            producer_epoch: AtomicU64::new(0),
            topology_epoch: AtomicU64::new(0),
            mode,
            start_after_nr_of_consumers,
            buffer_size,
            partitioner,
        }
    }

    fn register_consumer(&self) -> Arc<FanOutConsumerLane<T>> {
        let mut registry = self.registry.lock().expect("fan-out hub poisoned");
        let id = registry.next_consumer_id;
        registry.next_consumer_id += 1;

        if let Some(terminal) = registry.terminal.clone() {
            return Arc::new(FanOutConsumerLane::terminal(id, terminal));
        }

        let lane = Arc::new(FanOutConsumerLane::new(id, self.buffer_size));
        registry.consumers.insert(id, Arc::clone(&lane));
        self.publish_snapshot_locked(&registry);
        drop(registry);
        self.notify_topology_transition();
        lane
    }

    fn remove_consumer(&self, consumer_id: u64) {
        let lane = {
            let mut registry = self.registry.lock().expect("fan-out hub poisoned");
            let lane = registry.consumers.remove(&consumer_id);
            if let Some(lane) = &lane {
                lane.deactivate();
                self.publish_snapshot_locked(&registry);
            }
            lane
        };

        if let Some(lane) = lane {
            lane.condvar.notify_all();
            self.notify_topology_transition();
        }
    }

    fn wait_for_broadcast_capacity(&self, max_items: usize) -> StreamResult<usize> {
        loop {
            let observed = self.producer_epoch.load(Ordering::Acquire);
            let snapshot = self.snapshot.load();
            if let Some(terminal) = &snapshot.terminal {
                return Err(terminal.producer_error());
            }

            let lanes = self.active_lanes(&snapshot);
            if lanes.len() < self.start_after_nr_of_consumers || lanes.is_empty() {
                self.wait_for_producer_transition(observed);
                continue;
            }

            let free = lanes
                .iter()
                .map(|lane| self.buffer_size.saturating_sub(lane.queued_len()))
                .min()
                .unwrap_or(0);
            if free > 0 {
                let max_items = if lanes.len() == 1 {
                    max_items.min(BROADCAST_HUB_SINGLE_CONSUMER_BATCH_LIMIT)
                } else {
                    max_items
                };
                return Ok(free.min(max_items).max(1));
            }
            self.wait_for_producer_transition(observed);
        }
    }

    fn push_broadcast_item(&self, item: T) -> StreamResult<()>
    where
        T: Clone,
    {
        let mut batch = VecDeque::new();
        batch.push_back(item);
        self.push_broadcast_batch(&mut batch)
    }

    fn push_broadcast_batch(&self, batch: &mut VecDeque<T>) -> StreamResult<()>
    where
        T: Clone,
    {
        if batch.is_empty() {
            return Ok(());
        }

        while !batch.is_empty() {
            let observed = self.producer_epoch.load(Ordering::Acquire);
            let snapshot = self.snapshot.load();
            if let Some(terminal) = &snapshot.terminal {
                batch.clear();
                return Err(terminal.producer_error());
            }

            let lanes = self.active_lanes(&snapshot);
            if lanes.len() < self.start_after_nr_of_consumers || lanes.is_empty() {
                self.wait_for_producer_transition(observed);
                continue;
            }

            let free = lanes
                .iter()
                .map(|lane| self.buffer_size.saturating_sub(lane.queued_len()))
                .min()
                .unwrap_or(0);
            if free == 0 {
                self.wait_for_producer_transition(observed);
                continue;
            }

            let take_n = free.min(batch.len());
            if !self.try_push_broadcast_batch(&lanes, batch, take_n)? {
                self.wait_for_producer_transition(observed);
            }
        }
        Ok(())
    }

    fn try_push_broadcast_batch(
        &self,
        lanes: &[Arc<FanOutConsumerLane<T>>],
        batch: &mut VecDeque<T>,
        take_n: usize,
    ) -> StreamResult<bool>
    where
        T: Clone,
    {
        if take_n == 0 || lanes.is_empty() {
            return Ok(false);
        }

        let mut guards = SmallVec::<
            [(
                Arc<FanOutConsumerLane<T>>,
                MutexGuard<'_, FanOutLaneState<T>>,
            ); 16],
        >::new();
        for lane in lanes {
            if !lane.is_active() {
                return Ok(false);
            }
            let guard = lane.state.lock().expect("fan-out lane poisoned");
            if !lane.is_active() {
                return Ok(false);
            }
            if let Some(terminal) = &guard.terminal {
                return Err(terminal.producer_error());
            }
            if self.buffer_size.saturating_sub(guard.queued) < take_n {
                return Ok(false);
            }
            guards.push((Arc::clone(lane), guard));
        }

        let lane_count = guards.len();
        if lane_count == 0 {
            return Ok(false);
        }

        let mut notify_lanes = SmallVec::<[Arc<FanOutConsumerLane<T>>; 16]>::new();
        for (index, (lane, guard)) in guards.iter_mut().enumerate() {
            let chunk = if index + 1 == lane_count {
                batch.drain(..take_n).collect()
            } else {
                batch.iter().take(take_n).cloned().collect()
            };
            guard.chunks.push_back(chunk);
            guard.queued += take_n;
            lane.queued.store(guard.queued, Ordering::Release);
            notify_lanes.push(Arc::clone(lane));
        }
        drop(guards);

        for lane in notify_lanes {
            lane.condvar.notify_one();
        }
        Ok(true)
    }

    fn select_partition(
        &self,
        item: &T,
        cache: &mut PartitionTopologyCache<T>,
    ) -> StreamResult<Option<Arc<FanOutConsumerLane<T>>>> {
        let Some(partitioner) = &self.partitioner else {
            return Err(StreamError::Failed(
                "partition hub partitioner missing".to_owned(),
            ));
        };
        let topology_epoch = self.topology_epoch.load(Ordering::Acquire);
        if cache.epoch != topology_epoch || cache.lanes.is_empty() {
            self.refresh_partition_topology(cache)?;
        }

        let info = PartitionConsumerInfo::from_cached_lanes(&cache.consumer_ids, &cache.lanes);
        let selected = partitioner(&info, item);
        if selected < 0 {
            return Ok(None);
        }
        let selected = selected as u64;
        let selected_idx = selected as usize;
        if cache.consumer_ids.get(selected_idx).copied() == Some(selected) {
            return Ok(Some(Arc::clone(&cache.lanes[selected_idx])));
        }
        cache
            .consumer_ids
            .iter()
            .position(|id| *id == selected)
            .map(|idx| Some(Arc::clone(&cache.lanes[idx])))
            .ok_or_else(|| {
                StreamError::Failed("partition hub selected unknown consumer".to_owned())
            })
    }

    fn refresh_partition_topology(
        &self,
        cache: &mut PartitionTopologyCache<T>,
    ) -> StreamResult<()> {
        loop {
            let observed = self.producer_epoch.load(Ordering::Acquire);
            let topology_epoch = self.topology_epoch.load(Ordering::Acquire);
            let snapshot = self.snapshot.load();
            if let Some(terminal) = &snapshot.terminal {
                cache.clear();
                return Err(terminal.producer_error());
            }

            let lanes = self.active_lanes(&snapshot);
            if lanes.len() >= self.start_after_nr_of_consumers && !lanes.is_empty() {
                cache.epoch = topology_epoch;
                cache.consumer_ids.clear();
                cache
                    .consumer_ids
                    .extend(lanes.iter().map(|lane| lane.id()));
                cache.lanes = lanes;
                return Ok(());
            }
            self.wait_for_producer_transition(observed);
        }
    }

    fn enqueue_partition(&self, selected: Arc<FanOutConsumerLane<T>>, item: T) -> StreamResult<()> {
        let mut batch = VecDeque::new();
        batch.push_back(item);
        self.enqueue_partition_batch(selected, &mut batch)
    }

    fn enqueue_partition_batch(
        &self,
        selected: Arc<FanOutConsumerLane<T>>,
        batch: &mut VecDeque<T>,
    ) -> StreamResult<()> {
        if batch.is_empty() {
            return Ok(());
        }

        let mut state = selected.state.lock().expect("fan-out lane poisoned");
        loop {
            if !selected.is_active() {
                batch.clear();
                return Err(StreamError::Failed(
                    "partition hub selected unknown consumer".to_owned(),
                ));
            }
            if let Some(terminal) = &state.terminal {
                batch.clear();
                return Err(terminal.producer_error());
            }
            let free = self.buffer_size.saturating_sub(state.queued);
            if free > 0 {
                let take_n = free.min(batch.len());
                let chunk = batch.drain(..take_n).collect();
                state.chunks.push_back(chunk);
                state.queued += take_n;
                selected.queued.store(state.queued, Ordering::Release);
                selected.condvar.notify_one();
                if batch.is_empty() {
                    return Ok(());
                }
            }
            let (guard, _) = selected
                .condvar
                .wait_timeout(state, fan_out_wait_timeout())
                .expect("fan-out lane poisoned while waiting");
            state = guard;
        }
    }

    fn complete(&self) {
        let lanes = {
            let mut registry = self.registry.lock().expect("fan-out hub poisoned");
            if registry.terminal.is_none() {
                registry.terminal = Some(FanOutTerminal::Completed);
                self.publish_snapshot_locked(&registry);
            }
            registry.consumers.values().cloned().collect::<Vec<_>>()
        };

        for lane in lanes {
            lane.set_terminal(FanOutTerminal::Completed);
        }
        self.notify_topology_transition();
    }

    fn fail(&self, error: StreamError) {
        let terminal = FanOutTerminal::Failed(error);
        let lanes = {
            let mut registry = self.registry.lock().expect("fan-out hub poisoned");
            if registry.terminal.is_none() {
                registry.terminal = Some(terminal.clone());
                self.publish_snapshot_locked(&registry);
            }
            registry.consumers.values().cloned().collect::<Vec<_>>()
        };

        for lane in lanes {
            lane.set_terminal(terminal.clone());
        }
        self.notify_topology_transition();
    }

    fn publish_snapshot_locked(&self, registry: &FanOutRegistry<T>) {
        self.snapshot.store(Arc::new(FanOutSnapshot {
            consumers: registry.consumers.values().cloned().collect(),
            terminal: registry.terminal.clone(),
        }));
    }

    fn active_lanes(
        &self,
        snapshot: &FanOutSnapshot<T>,
    ) -> SmallVec<[Arc<FanOutConsumerLane<T>>; 16]> {
        snapshot
            .consumers
            .iter()
            .filter(|lane| lane.is_active())
            .cloned()
            .collect()
    }

    fn partition_batch_limit(&self) -> usize {
        let snapshot = self.snapshot.load();
        let active = snapshot
            .consumers
            .iter()
            .filter(|lane| lane.is_active())
            .count();
        match active {
            0 | 1 => PARTITION_HUB_SINGLE_CONSUMER_BATCH_LIMIT,
            2..=7 => PARTITION_HUB_BATCH_LIMIT,
            _ => PARTITION_HUB_WIDE_BATCH_LIMIT,
        }
    }

    fn notify_producer_transition(&self) {
        self.producer_epoch.fetch_add(1, Ordering::Release);
        self.producer_condvar.notify_one();
    }

    fn notify_topology_transition(&self) {
        self.topology_epoch.fetch_add(1, Ordering::Release);
        self.notify_producer_transition();
    }

    fn wait_for_producer_transition(&self, observed_epoch: u64) {
        let guard = self
            .producer_wait
            .lock()
            .expect("fan-out producer wait poisoned");
        if self.producer_epoch.load(Ordering::Acquire) == observed_epoch {
            let (_guard, _) = self
                .producer_condvar
                .wait_timeout(guard, fan_out_wait_timeout())
                .expect("fan-out producer wait poisoned while waiting");
        }
    }
}

fn push_partition_routed<T>(
    routed: &mut PartitionRoutedBatches<T>,
    lane: Arc<FanOutConsumerLane<T>>,
    item: T,
) {
    for (existing, batch) in routed.iter_mut() {
        if existing.id() == lane.id() {
            batch.push_back(item);
            return;
        }
    }

    let mut batch = VecDeque::new();
    batch.push_back(item);
    routed.push((lane, batch));
}

struct FanOutProducer<T> {
    input: BoxStream<T>,
    state: Arc<FanOutHubShared<T>>,
}

impl<T> FanOutProducer<T> {
    fn new(input: BoxStream<T>, state: Arc<FanOutHubShared<T>>) -> Self {
        Self { input, state }
    }
}

impl<T: Send + 'static + Clone> FanOutProducer<T> {
    fn run(
        mut self,
        cancelled: Arc<std::sync::atomic::AtomicBool>,
        hints: SourceRuntimeHints,
    ) -> StreamResult<NotUsed> {
        struct ProducerDropGuard<T> {
            state: Arc<FanOutHubShared<T>>,
            disarmed: bool,
        }

        impl<T> ProducerDropGuard<T> {
            fn new(state: Arc<FanOutHubShared<T>>) -> Self {
                Self {
                    state,
                    disarmed: false,
                }
            }

            fn disarm(&mut self) {
                self.disarmed = true;
            }
        }

        impl<T> Drop for ProducerDropGuard<T> {
            fn drop(&mut self) {
                if !self.disarmed && std::thread::panicking() {
                    self.state.fail(StreamError::Failed(
                        "fan-out hub producer panicked".to_owned(),
                    ));
                }
            }
        }

        let mut guard = ProducerDropGuard::new(Arc::clone(&self.state));
        match self.state.mode {
            FanOutMode::Broadcast => {
                let batch_producer = hints.inline_micro_max_success_items.is_some();
                let mut batch = VecDeque::new();
                loop {
                    if cancelled.load(Ordering::SeqCst) {
                        self.state.fail(StreamError::Cancelled);
                        guard.disarm();
                        return Err(StreamError::Cancelled);
                    }

                    if batch_producer {
                        let capacity = self
                            .state
                            .wait_for_broadcast_capacity(BROADCAST_HUB_BATCH_LIMIT)?;
                        batch.clear();
                        batch.reserve(capacity.saturating_sub(batch.capacity()));
                        let mut terminal = None;
                        for _ in 0..capacity {
                            if cancelled.load(Ordering::SeqCst) {
                                terminal = Some(Err(StreamError::Cancelled));
                                break;
                            }
                            match self.input.next() {
                                Some(Ok(item)) => batch.push_back(item),
                                Some(Err(error)) => {
                                    terminal = Some(Err(error));
                                    break;
                                }
                                None => {
                                    terminal = Some(Ok(()));
                                    break;
                                }
                            }
                        }
                        if !batch.is_empty() {
                            self.state.push_broadcast_batch(&mut batch)?;
                        }
                        match terminal {
                            Some(Err(StreamError::Cancelled)) => {
                                self.state.fail(StreamError::Cancelled);
                                guard.disarm();
                                return Err(StreamError::Cancelled);
                            }
                            Some(Err(error)) => {
                                self.state.fail(error.clone());
                                guard.disarm();
                                return Err(error);
                            }
                            Some(Ok(())) => {
                                self.state.complete();
                                guard.disarm();
                                return Ok(NotUsed);
                            }
                            None => continue,
                        }
                    }

                    match self.input.next() {
                        Some(Ok(item)) => self.state.push_broadcast_item(item)?,
                        Some(Err(error)) => {
                            self.state.fail(error.clone());
                            guard.disarm();
                            return Err(error);
                        }
                        None => {
                            self.state.complete();
                            guard.disarm();
                            return Ok(NotUsed);
                        }
                    }
                }
            }
            FanOutMode::Partition => {
                let batch_producer = hints.inline_micro_max_success_items.is_some();
                let mut topology_cache = PartitionTopologyCache::new();
                let mut routed = PartitionRoutedBatches::new();
                loop {
                    if cancelled.load(Ordering::SeqCst) {
                        self.state.fail(StreamError::Cancelled);
                        guard.disarm();
                        return Err(StreamError::Cancelled);
                    }

                    if batch_producer {
                        for (_, batch) in routed.iter_mut() {
                            batch.clear();
                        }
                        let partition_batch_limit = self.state.partition_batch_limit();
                        let mut terminal = None;
                        for _ in 0..partition_batch_limit {
                            if cancelled.load(Ordering::SeqCst) {
                                terminal = Some(Err(StreamError::Cancelled));
                                break;
                            }
                            match self.input.next() {
                                Some(Ok(item)) => {
                                    if let Some(selected) =
                                        self.state.select_partition(&item, &mut topology_cache)?
                                    {
                                        push_partition_routed(&mut routed, selected, item);
                                    }
                                }
                                Some(Err(error)) => {
                                    terminal = Some(Err(error));
                                    break;
                                }
                                None => {
                                    terminal = Some(Ok(()));
                                    break;
                                }
                            }
                        }
                        for (selected, batch) in routed.iter_mut() {
                            if !batch.is_empty() {
                                self.state
                                    .enqueue_partition_batch(Arc::clone(selected), batch)?;
                            }
                        }
                        match terminal {
                            Some(Err(StreamError::Cancelled)) => {
                                self.state.fail(StreamError::Cancelled);
                                guard.disarm();
                                return Err(StreamError::Cancelled);
                            }
                            Some(Err(error)) => {
                                self.state.fail(error.clone());
                                guard.disarm();
                                return Err(error);
                            }
                            Some(Ok(())) => {
                                self.state.complete();
                                guard.disarm();
                                return Ok(NotUsed);
                            }
                            None => continue,
                        }
                    }

                    match self.input.next() {
                        Some(Ok(item)) => {
                            if let Some(selected) =
                                self.state.select_partition(&item, &mut topology_cache)?
                            {
                                self.state.enqueue_partition(selected, item)?;
                            }
                        }
                        Some(Err(error)) => {
                            self.state.fail(error.clone());
                            guard.disarm();
                            return Err(error);
                        }
                        None => {
                            self.state.complete();
                            guard.disarm();
                            return Ok(NotUsed);
                        }
                    }
                }
            }
        }
    }
}

struct FanOutConsumerStream<T> {
    state: Arc<FanOutHubShared<T>>,
    lane: Arc<FanOutConsumerLane<T>>,
    local: Option<std::vec::IntoIter<T>>,
    detached: bool,
}

impl<T: Clone + Send + 'static> Iterator for FanOutConsumerStream<T> {
    type Item = StreamResult<T>;

    fn next(&mut self) -> Option<Self::Item> {
        if self.lane.failed.load(Ordering::Acquire) {
            self.local = None;
        } else if let Some(local) = &mut self.local {
            if let Some(item) = local.next() {
                return Some(Ok(item));
            }
            self.local = None;
        }

        let mut state = self.lane.state.lock().expect("fan-out lane poisoned");
        loop {
            if let Some(FanOutTerminal::Failed(error)) = state.terminal.clone() {
                return Some(Err(error));
            }
            if let Some(batch) = state.chunks.pop_front() {
                state.queued = state.queued.saturating_sub(batch.len());
                self.lane.queued.store(state.queued, Ordering::Release);
                if matches!(self.state.mode, FanOutMode::Partition) {
                    self.lane.condvar.notify_one();
                }
                self.state.notify_producer_transition();
                drop(state);
                let mut batch = batch.into_iter();
                let first = batch.next().expect("fan-out drained non-empty lane batch");
                if batch.len() > 0 {
                    self.local = Some(batch);
                }
                return Some(Ok(first));
            }
            if matches!(state.terminal, Some(FanOutTerminal::Completed)) {
                return None;
            }
            let (guard, _) = self
                .lane
                .condvar
                .wait_timeout(state, fan_out_wait_timeout())
                .expect("fan-out lane poisoned while waiting");
            state = guard;
        }
    }
}

impl<T> Drop for FanOutConsumerStream<T> {
    fn drop(&mut self) {
        if !self.detached {
            self.state.remove_consumer(self.lane.id());
            self.detached = true;
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::testkit::{TestSink, TestSource};
    use crate::{Keep, Materializer, Sink, Source};
    use std::{
        panic::{self, AssertUnwindSafe},
        sync::{
            Arc,
            atomic::{AtomicUsize, Ordering},
        },
        thread,
        time::{Duration, Instant},
    };

    #[test]
    fn merge_hub_accepts_dynamic_producers_and_drains() {
        let materializer = Materializer::new();
        let ((hub_sink, control), completion) = MergeHub::source_with_draining::<i32>(4)
            .to_mat(Sink::collect(), Keep::both)
            .run_with_materializer(&materializer)
            .expect("merge hub materializes");

        hub_sink
            .clone()
            .run_with(Source::from_iter([1, 2, 3]))
            .expect("first producer attaches");
        hub_sink
            .run_with(Source::from_iter([4, 5]))
            .expect("second producer attaches");
        control.drain_and_complete();
        let mut result = completion.wait().expect("merge hub completes");
        result.sort_unstable();
        assert_eq!(result, vec![1, 2, 3, 4, 5]);
    }

    #[test]
    fn merge_hub_producer_error_fails_downstream_consumer() {
        let materializer = Materializer::new();
        let (hub_sink, sink) = MergeHub::source::<i32>(4)
            .to_mat(TestSink::probe(), Keep::both)
            .run_with_materializer(&materializer)
            .expect("merge hub materializes");

        let producer_ok = TestSource::probe::<i32>()
            .to_mat(hub_sink.clone(), Keep::left)
            .run_with_materializer(&materializer)
            .expect("successful producer attaches");
        let producer_fail = TestSource::probe::<i32>()
            .to_mat(hub_sink, Keep::left)
            .run_with_materializer(&materializer)
            .expect("failing producer attaches");

        sink.request(1);
        assert_eq!(producer_ok.expect_request(), 1);
        producer_ok.send_next(1);
        sink.assert_next(1);

        producer_fail.send_error(StreamError::Failed("producer failed".to_owned()));
        sink.request(1);
        assert_eq!(
            sink.expect_error(),
            StreamError::Failed("producer failed".to_owned())
        );
    }

    #[test]
    fn broadcast_hub_backpressures_slowest_consumer() {
        let materializer = Materializer::new();
        let (publisher, hub_source) = TestSource::probe::<i32>()
            .to_mat(BroadcastHub::sink(1), Keep::both)
            .run_with_materializer(&materializer)
            .expect("broadcast hub materializes");

        let sink_a = hub_source
            .source()
            .run_with(TestSink::probe())
            .expect("first consumer materializes");
        let sink_b = hub_source
            .source()
            .run_with(TestSink::probe())
            .expect("second consumer materializes");

        sink_a.request(1);
        sink_b.request(1);
        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(1);
        sink_a.assert_next(1);
        sink_b.assert_next(1);

        sink_a.request(1);
        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(2);
        sink_a.assert_next(2);
        sink_b.expect_no_message(Duration::from_millis(250));

        sink_a.request(1);
        sink_a.expect_no_message(Duration::from_millis(250));

        sink_b.request(1);
        sink_b.assert_next(2);
        assert_eq!(publisher.expect_request(), 1);
    }

    #[test]
    fn broadcast_hub_late_consumer_sees_only_late_elements() {
        let materializer = Materializer::new();
        let (publisher, hub_source) = TestSource::probe::<i32>()
            .to_mat(BroadcastHub::sink(2), Keep::both)
            .run_with_materializer(&materializer)
            .expect("broadcast hub materializes");

        let sink_a = hub_source
            .source()
            .run_with(TestSink::probe())
            .expect("first consumer materializes");
        sink_a.request(1);
        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(1);
        sink_a.assert_next(1);

        let sink_b = hub_source
            .source()
            .run_with(TestSink::probe())
            .expect("late consumer materializes");
        sink_a.request(1);
        sink_b.request(1);
        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(2);
        sink_a.assert_next(2);
        sink_b.assert_next(2);

        publisher.send_complete();
        sink_a.request(1);
        sink_b.request(1);
        sink_a.expect_complete();
        sink_b.expect_complete();
    }

    #[test]
    fn partition_hub_routes_elements_to_selected_consumers() {
        let materializer = Materializer::new();
        let hub = Source::from_iter([0, 1, 2, 3])
            .run_with_materializer(
                PartitionHub::sink(
                    |info, item| {
                        let idx = (*item as usize) % info.size();
                        info.consumer_id_by_idx(idx) as isize
                    },
                    2,
                    8,
                ),
                &materializer,
            )
            .expect("partition hub materializes");

        let sink_a = hub
            .source()
            .run_with(TestSink::probe())
            .expect("first consumer materializes");
        let sink_b = hub
            .source()
            .run_with(TestSink::probe())
            .expect("second consumer materializes");

        sink_a.request(2);
        sink_b.request(2);
        sink_a.assert_next_n([0, 2]);
        sink_b.assert_next_n([1, 3]);
    }

    #[test]
    fn partition_hub_evaluates_stateful_partitioner_once_per_blocked_element() {
        let materializer = Materializer::new();
        let partition_calls = Arc::new(AtomicUsize::new(0));
        let partition_calls_for_hub = Arc::clone(&partition_calls);

        let (publisher, hub) = TestSource::probe::<i32>()
            .to_mat(
                PartitionHub::sink(
                    move |info, _item| {
                        partition_calls_for_hub.fetch_add(1, Ordering::SeqCst);
                        info.consumer_id_by_idx(0) as isize
                    },
                    1,
                    1,
                ),
                Keep::both,
            )
            .run_with_materializer(&materializer)
            .expect("partition hub materializes");

        let sink = hub
            .source()
            .run_with(TestSink::probe())
            .expect("consumer materializes");

        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(1);
        wait_for_partition_calls(&partition_calls, 1);

        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(2);
        sink.expect_no_message(Duration::from_millis(250));
        wait_for_partition_calls(&partition_calls, 2);

        sink.request(1);
        sink.assert_next(1);
        sink.request(1);
        sink.assert_next(2);
        assert_eq!(partition_calls.load(Ordering::SeqCst), 2);
    }

    #[test]
    fn partition_hub_invalidates_cached_topology_on_consumer_churn() {
        let materializer = Materializer::new();
        let (publisher, hub) = TestSource::probe::<i32>()
            .to_mat(
                PartitionHub::sink(
                    |info, _item| info.consumer_id_by_idx(info.size() - 1) as isize,
                    1,
                    8,
                ),
                Keep::both,
            )
            .run_with_materializer(&materializer)
            .expect("partition hub materializes");

        let sink_a = hub
            .source()
            .run_with(TestSink::probe())
            .expect("first consumer materializes");
        sink_a.request(1);
        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(1);
        sink_a.assert_next(1);
        drop(sink_a);

        let sink_b = hub
            .source()
            .run_with(TestSink::probe())
            .expect("replacement consumer materializes");
        sink_b.request(1);
        assert_eq!(publisher.expect_request(), 1);
        publisher.send_next(2);
        sink_b.assert_next(2);

        publisher.send_complete();
        sink_b.request(1);
        sink_b.expect_complete();
    }

    #[test]
    fn broadcast_hub_drains_local_chunk_before_completion() {
        let materializer = Materializer::new();
        let hub = Source::from_iter([1, 2, 3])
            .run_with_materializer(BroadcastHub::sink_starting_after(1, 8), &materializer)
            .expect("broadcast hub materializes");

        let sink = hub
            .source()
            .run_with(TestSink::probe())
            .expect("consumer materializes");

        sink.request(1);
        sink.assert_next(1);
        sink.request(3);
        sink.assert_next_n([2, 3]);
        sink.expect_complete();
    }

    #[test]
    fn broadcast_hub_panicking_upstream_fails_consumers() {
        let materializer = Materializer::new();
        let hub = Source::from_fn_iter(|| {
            let mut yielded = false;
            std::iter::from_fn(move || {
                if !yielded {
                    yielded = true;
                    Some(1)
                } else {
                    panic!("boom");
                }
            })
        })
        .run_with_materializer(BroadcastHub::sink_starting_after(1, 8), &materializer)
        .expect("broadcast hub materializes");

        let sink = hub
            .source()
            .run_with(TestSink::probe())
            .expect("consumer materializes");

        sink.request(2);
        match panic::catch_unwind(AssertUnwindSafe(|| sink.expect_error())) {
            Ok(error) => assert_eq!(
                error,
                StreamError::Failed("fan-out hub producer panicked".to_owned())
            ),
            Err(payload) => {
                assert_eq!(
                    panic_message(payload),
                    "expected stream error, got next element"
                );
                sink.request(1);
                assert_eq!(
                    sink.expect_error(),
                    StreamError::Failed("fan-out hub producer panicked".to_owned())
                );
            }
        }
    }

    fn panic_message(payload: Box<dyn std::any::Any + Send>) -> String {
        match payload.downcast::<String>() {
            Ok(message) => *message,
            Err(payload) => match payload.downcast::<&'static str>() {
                Ok(message) => (*message).to_owned(),
                Err(_) => "<non-string panic payload>".to_owned(),
            },
        }
    }

    fn wait_for_partition_calls(counter: &AtomicUsize, expected: usize) {
        let deadline = Instant::now() + Duration::from_secs(1);
        while Instant::now() < deadline {
            if counter.load(Ordering::SeqCst) == expected {
                return;
            }
            thread::sleep(Duration::from_millis(5));
        }
        assert_eq!(counter.load(Ordering::SeqCst), expected);
    }
}