swactor/crates/datastream/src/endpoint.rs

460 lines
14 KiB
Rust
Raw Normal View History

//! Process-local datastream endpoint.
//!
//! This is the sidecar a process creates next to a swactor runtime. It is not
//! stored in core runtime state: producers get a cheap [`DatastreamProducer`],
//! frames are ordered by the local [`Mux`], and subscribers receive future
//! deliveries through bounded queues. With no subscribers, draining is a
//! bitbucket.
use std::collections::BTreeMap;
use std::sync::atomic::{AtomicU64, Ordering};
use std::sync::{Arc, Mutex, mpsc};
use serde::Serialize;
use swactor::process_observer::ProcessOutputObserver;
use swactor::stats::{ActorSnapshot, StatsHook};
use crate::frame::{ChannelId, Position, StreamId};
use crate::mux::Mux;
use crate::record::Record;
use crate::transport::Delivery;
const DEFAULT_MUX_CAPACITY: usize = 4096;
const DEFAULT_SUBSCRIBER_CAPACITY: usize = 1024;
const DEFAULT_STATS_CHANNEL: &str = "runtime.actors";
/// Stable handle identifying a local datastream subscription.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct SubscriptionId(pub u64);
/// Drain statistics for one endpoint tick.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub struct EndpointTick {
/// Frames drained from the endpoint mux.
pub drained: usize,
/// Delivery copies successfully enqueued to subscribers.
pub delivered: usize,
/// Delivery copies dropped because a subscriber queue was full or closed.
pub dropped_for_subscribers: usize,
/// Number of subscribers present when the batch was published.
pub subscribers: usize,
}
/// Snapshot of one subscriber's local fanout state.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SubscriberSnapshot {
pub id: SubscriptionId,
pub name: String,
pub dropped: u64,
}
/// Bounded future-frame subscription.
pub struct DatastreamSubscription {
id: SubscriptionId,
name: String,
rx: mpsc::Receiver<Delivery>,
}
impl DatastreamSubscription {
pub fn id(&self) -> SubscriptionId {
self.id
}
pub fn name(&self) -> &str {
&self.name
}
pub fn try_recv(&self) -> Result<Delivery, mpsc::TryRecvError> {
self.rx.try_recv()
}
pub fn recv(&self) -> Result<Delivery, mpsc::RecvError> {
self.rx.recv()
}
pub fn recv_timeout(
&self,
timeout: std::time::Duration,
) -> Result<Delivery, mpsc::RecvTimeoutError> {
self.rx.recv_timeout(timeout)
}
pub fn drain_available(&self) -> Vec<Delivery> {
let mut out = Vec::new();
while let Ok(delivery) = self.rx.try_recv() {
out.push(delivery);
}
out
}
}
struct SubscriberSlot {
name: String,
tx: mpsc::SyncSender<Delivery>,
dropped: u64,
}
struct FanoutState {
next_id: u64,
subscribers: BTreeMap<SubscriptionId, SubscriberSlot>,
}
/// Local delivery fanout used by endpoints and collectors.
///
/// Publishing with zero subscribers is intentionally a drop. This gives the
/// default runtime behavior requested for v0: telemetry can be produced and
/// drained without retaining historical frames, and subscriptions receive only
/// future deliveries.
pub struct DeliveryFanout {
default_capacity: usize,
state: Mutex<FanoutState>,
}
impl DeliveryFanout {
pub fn new(default_capacity: usize) -> Self {
Self {
default_capacity: default_capacity.max(1),
state: Mutex::new(FanoutState {
next_id: 1,
subscribers: BTreeMap::new(),
}),
}
}
pub fn subscribe_all(&self, name: impl Into<String>) -> DatastreamSubscription {
self.subscribe_all_with_capacity(name, self.default_capacity)
}
pub fn subscribe_all_with_capacity(
&self,
name: impl Into<String>,
capacity: usize,
) -> DatastreamSubscription {
let name = name.into();
let (tx, rx) = mpsc::sync_channel(capacity.max(1));
let mut state = self.state.lock().expect("datastream fanout poisoned");
let id = SubscriptionId(state.next_id);
state.next_id = state.next_id.wrapping_add(1).max(1);
state.subscribers.insert(
id,
SubscriberSlot {
name: name.clone(),
tx,
dropped: 0,
},
);
DatastreamSubscription { id, name, rx }
}
pub fn subscriber_count(&self) -> usize {
self.state
.lock()
.expect("datastream fanout poisoned")
.subscribers
.len()
}
pub fn subscriber_snapshots(&self) -> Vec<SubscriberSnapshot> {
self.state
.lock()
.expect("datastream fanout poisoned")
.subscribers
.iter()
.map(|(id, slot)| SubscriberSnapshot {
id: *id,
name: slot.name.clone(),
dropped: slot.dropped,
})
.collect()
}
pub fn publish(&self, delivery: Delivery) -> EndpointTick {
self.publish_batch(std::iter::once(delivery))
}
pub fn publish_batch(&self, deliveries: impl IntoIterator<Item = Delivery>) -> EndpointTick {
let deliveries: Vec<Delivery> = deliveries.into_iter().collect();
if deliveries.is_empty() {
return EndpointTick::default();
}
let mut state = self.state.lock().expect("datastream fanout poisoned");
let subscribers = state.subscribers.len();
if subscribers == 0 {
return EndpointTick {
drained: deliveries.len(),
subscribers: 0,
..EndpointTick::default()
};
}
let mut delivered = 0;
let mut dropped = 0;
let mut disconnected = Vec::new();
for (id, slot) in state.subscribers.iter_mut() {
for delivery in &deliveries {
match slot.tx.try_send(delivery.clone()) {
Ok(()) => delivered += 1,
Err(mpsc::TrySendError::Full(_)) => {
slot.dropped = slot.dropped.saturating_add(1);
dropped += 1;
}
Err(mpsc::TrySendError::Disconnected(_)) => {
slot.dropped = slot.dropped.saturating_add(1);
dropped += 1;
disconnected.push(*id);
break;
}
}
}
}
for id in disconnected {
state.subscribers.remove(&id);
}
EndpointTick {
drained: deliveries.len(),
delivered,
dropped_for_subscribers: dropped,
subscribers,
}
}
}
/// Process-local datastream endpoint.
pub struct DatastreamEndpoint {
stream: StreamId,
mux: Arc<Mux>,
fanout: DeliveryFanout,
drained: AtomicU64,
bitbucketed: AtomicU64,
}
impl DatastreamEndpoint {
pub fn new(stream: StreamId) -> Self {
Self::with_capacity(stream, DEFAULT_MUX_CAPACITY, DEFAULT_SUBSCRIBER_CAPACITY)
}
pub fn with_capacity(
stream: StreamId,
mux_capacity: usize,
subscriber_capacity: usize,
) -> Self {
let mux = Arc::new(Mux::new(stream.clone(), mux_capacity.max(1)));
Self {
stream,
mux,
fanout: DeliveryFanout::new(subscriber_capacity),
drained: AtomicU64::new(0),
bitbucketed: AtomicU64::new(0),
}
}
pub fn stream_id(&self) -> &StreamId {
&self.stream
}
pub fn mux(&self) -> &Arc<Mux> {
&self.mux
}
/// Enable or disable optional sidecar timing samples for newly submitted
/// frames from this endpoint's producers.
pub fn set_frame_timing_enabled(&self, enabled: bool) {
self.mux.set_frame_timing_enabled(enabled);
}
/// Whether this endpoint currently emits sidecar frame timing samples.
pub fn frame_timing_enabled(&self) -> bool {
self.mux.frame_timing_enabled()
}
pub fn producer(&self) -> DatastreamProducer {
DatastreamProducer {
mux: Arc::clone(&self.mux),
}
}
pub fn subscribe_all(&self, name: impl Into<String>) -> DatastreamSubscription {
self.fanout.subscribe_all(name)
}
pub fn subscribe_all_with_capacity(
&self,
name: impl Into<String>,
capacity: usize,
) -> DatastreamSubscription {
self.fanout.subscribe_all_with_capacity(name, capacity)
}
pub fn subscriber_count(&self) -> usize {
self.fanout.subscriber_count()
}
pub fn subscriber_snapshots(&self) -> Vec<SubscriberSnapshot> {
self.fanout.subscriber_snapshots()
}
/// Drain the mux and fan out the resulting future deliveries.
pub fn tick(&self) -> EndpointTick {
let frames = self.mux.drain();
if frames.is_empty() {
return EndpointTick::default();
}
let drained = frames.len();
self.drained.fetch_add(drained as u64, Ordering::Relaxed);
let deliveries = frames
.into_iter()
.map(|frame| Delivery::new(self.stream.clone(), frame));
let tick = self.fanout.publish_batch(deliveries);
if tick.subscribers == 0 {
self.bitbucketed
.fetch_add(drained as u64, Ordering::Relaxed);
}
tick
}
pub fn assigned(&self) -> u64 {
self.mux.assigned()
}
pub fn mux_dropped(&self) -> u64 {
self.mux.dropped()
}
pub fn drained(&self) -> u64 {
self.drained.load(Ordering::Relaxed)
}
pub fn bitbucketed(&self) -> u64 {
self.bitbucketed.load(Ordering::Relaxed)
}
}
/// Cloneable producer handle for code that emits telemetry.
#[derive(Clone)]
pub struct DatastreamProducer {
mux: Arc<Mux>,
}
impl DatastreamProducer {
pub fn stream_id(&self) -> &StreamId {
self.mux.stream_id()
}
pub fn submit_record<R: Record>(&self, record: &R) -> Position {
self.mux.submit(R::channel(), record.encode())
}
pub fn submit_text(&self, channel: impl Into<ChannelId>, text: impl AsRef<[u8]>) -> Position {
self.mux.submit(channel, text.as_ref().to_vec())
}
pub fn submit_bytes(&self, channel: impl Into<ChannelId>, bytes: Vec<u8>) -> Position {
self.mux.submit(channel, bytes)
}
/// Enable or disable optional sidecar timing samples for this producer's mux.
pub fn set_frame_timing_enabled(&self, enabled: bool) {
self.mux.set_frame_timing_enabled(enabled);
}
/// Whether this producer's mux currently emits sidecar frame timing samples.
pub fn frame_timing_enabled(&self) -> bool {
self.mux.frame_timing_enabled()
}
pub fn process_observer_with<F>(&self, channel_for: F) -> Arc<dyn ProcessOutputObserver>
where
F: Fn(&str, bool) -> ChannelId + Send + Sync + 'static,
{
Arc::new(DatastreamProcessObserver {
producer: self.clone(),
channel_for: Arc::new(channel_for),
})
}
pub fn stats_hook(&self) -> Arc<dyn StatsHook> {
self.stats_hook_on(DEFAULT_STATS_CHANNEL)
}
pub fn stats_hook_on(&self, channel: impl Into<ChannelId>) -> Arc<dyn StatsHook> {
Arc::new(DatastreamStatsHook {
producer: self.clone(),
channel: channel.into(),
})
}
}
/// Managed-process output observer that submits stdout/stderr chunks as frames.
pub struct DatastreamProcessObserver {
producer: DatastreamProducer,
channel_for: Arc<dyn Fn(&str, bool) -> ChannelId + Send + Sync>,
}
impl ProcessOutputObserver for DatastreamProcessObserver {
fn on_output(&self, label: &str, is_stderr: bool, data: &[u8]) {
self.producer
.submit_bytes((self.channel_for)(label, is_stderr), data.to_vec());
}
}
/// Runtime stats hook that submits one JSON record per productive worker tick.
pub struct DatastreamStatsHook {
producer: DatastreamProducer,
channel: ChannelId,
}
impl StatsHook for DatastreamStatsHook {
fn on_tick(&self, worker_id: usize, snapshots: &[ActorSnapshot]) {
let payload = RuntimeActorStatsRecord::from_snapshots(worker_id, snapshots);
let bytes = serde_json::to_vec(&payload).expect("runtime stats record serializes");
self.producer.submit_bytes(self.channel.clone(), bytes);
}
}
#[derive(Debug, Serialize)]
struct RuntimeActorStatsRecord<'a> {
worker_id: usize,
actors: Vec<RuntimeActorSnapshotRecord<'a>>,
}
impl<'a> RuntimeActorStatsRecord<'a> {
fn from_snapshots(worker_id: usize, snapshots: &'a [ActorSnapshot]) -> Self {
Self {
worker_id,
actors: snapshots
.iter()
.map(|snapshot| RuntimeActorSnapshotRecord {
address: snapshot.address.to_string(),
mailbox_depth: snapshot.mailbox_depth,
last_msg_type: snapshot.last_msg_type,
messages_processed: snapshot.messages_processed,
poisoned: snapshot.poisoned,
message_type_counts: snapshot
.message_type_counts
.iter()
.map(|(ty, count)| RuntimeMessageTypeCount { ty, count: *count })
.collect(),
})
.collect(),
}
}
}
#[derive(Debug, Serialize)]
struct RuntimeActorSnapshotRecord<'a> {
address: String,
mailbox_depth: usize,
last_msg_type: Option<&'static str>,
messages_processed: u64,
poisoned: bool,
message_type_counts: Vec<RuntimeMessageTypeCount<'a>>,
}
#[derive(Debug, Serialize)]
struct RuntimeMessageTypeCount<'a> {
ty: &'a str,
count: u64,
}