Skip to main content

spate_core/pipeline/
mod.rs

1//! Pipeline runtime: pinned driver threads, the source controller, and
2//! process assembly.
3//!
4//! Thread anatomy (see `docs/DESIGN.md` § Process anatomy):
5//!
6//! ```text
7//! main (run())          controller (std thread)      driver 0..N (std threads)
8//!   metrics/admin/io  ──  owns Source + Checkpointer ── own lanes + chain
9//!   joins everything      poll_events / commit tick     poll → push_batch → route
10//!                         pause/resume application      backpressure ticks
11//! ```
12//!
13//! Communication: the controller sends [`ThreadControl`] messages to
14//! drivers (lane assignment, drain barriers); drivers send [`DriverEvent`]
15//! requests back (pause/resume — only the controller touches the
16//! [`Source`](crate::source::Source) — and fatal reports). All channels are
17//! unbounded crossbeam channels: control traffic is rare and must never
18//! block a poll loop.
19//!
20//! Shutdown (also the full-revocation path, per DESIGN.md § Shutdown):
21//! SIGTERM → controller stops event polling and sends `Shutdown` to every
22//! driver → each driver flushes its chain, drops its lanes, and arrives at
23//! the barrier → main joins driver threads (dropping chains closes the
24//! shard queues) → the sink drains under the remaining deadline → the
25//! controller runs a final drain + commit + `flush_commits` → the process
26//! reports an [`ExitReport`]. A sink that cannot flush by the deadline is
27//! abandoned loudly; unacknowledged offsets are never committed, so the
28//! data replays after restart (at-least-once).
29
30mod builder;
31mod controller;
32mod driver;
33mod runtime;
34
35pub use builder::{BuildError, ChainCtx, Pipeline, PipelineError, SinkOptions};
36pub use runtime::{PipelineRuntime, RuntimeOptions, ShutdownHandle, StartError, metrics_settings};
37
38use crate::error::FatalError;
39use crate::record::PartitionId;
40use crate::sink::ShardQueues;
41use crate::source::{DrainBarrier, LaneId};
42use std::time::Instant;
43
44/// Control messages the controller sends to a driver thread.
45pub(crate) enum ThreadControl<L> {
46    /// Take ownership of a newly assigned lane.
47    AddLane(L),
48    /// Stop and drop the listed lanes (revocation): flush the chain, then
49    /// arrive at `barrier` once per stopped lane before `deadline`.
50    StopLanes {
51        lanes: Vec<LaneId>,
52        barrier: DrainBarrier,
53        deadline: Instant,
54    },
55    /// Flush the chain's partial terminal state now, without waiting out
56    /// the idle-flush lull
57    /// ([`SourceEvent::CommitReady`](crate::source::SourceEvent::CommitReady)).
58    ///
59    /// A lane that reached genuine end-of-input has no more data coming, so
60    /// its tail would otherwise sit in the chain until `idle_flush` elapses
61    /// — unacknowledged, and therefore blocking the completion of whatever
62    /// unit of work it belongs to. Best-effort and unsynchronised: the
63    /// controller does not wait, and a blocked chain simply retries on the
64    /// ordinary lull check.
65    FlushNow,
66    /// Drop the listed lanes without flushing or synchronising
67    /// ([`SourceEvent::LanesRetired`](crate::source::SourceEvent::LanesRetired)):
68    /// their input is fully delivered, acknowledged *and* committed, so
69    /// nothing of theirs can sit unflushed in the chain. Pure bookkeeping —
70    /// no barrier, no deadline, no `flush_until`. Forcing a flush here would
71    /// fragment sink batches and park the thread once per completed unit of
72    /// work; sources with anything in flight must use
73    /// [`ThreadControl::StopLanes`] instead.
74    DropLanes { lanes: Vec<LaneId> },
75    /// Stop everything (shutdown): flush the chain, drop all lanes, arrive
76    /// once at `barrier`, and exit the thread.
77    Shutdown {
78        barrier: DrainBarrier,
79        deadline: Instant,
80    },
81}
82
83/// Requests and reports a driver thread sends the controller.
84#[derive(Debug)]
85pub(crate) enum DriverEvent {
86    /// Backpressure tripped: pause these lanes at the source.
87    PauseLanes { lanes: Vec<LaneId> },
88    /// Backpressure cleared: resume these lanes.
89    ResumeLanes { lanes: Vec<LaneId> },
90    /// The chain failed or panicked; the pipeline must stop.
91    Fatal { thread: usize, error: FatalError },
92}
93
94/// What the sink reported when draining at shutdown — the sink layer's
95/// [`DrainReport`](crate::sink::DrainReport), re-exported so assemblies
96/// hand `SinkPool::drain`'s result straight through.
97pub use crate::sink::DrainReport;
98
99/// The sink half the runtime drives: the shared shard-queue handle plus a
100/// drain hook invoked once at shutdown with the remaining drain budget.
101///
102/// Built by the sink layer (`SinkPool`) or by tests; the runtime is
103/// deliberately ignorant of worker internals.
104pub struct SinkRuntime {
105    /// Sending side of the per-shard chunk queues, one entry per installed
106    /// sink (the runtime only uses capacity introspection for the
107    /// backpressure resume gate; the chain's terminal stage holds clones).
108    /// A single-sink pipeline has one entry.
109    pub queues: Vec<ShardQueues>,
110    /// Drain the sinks: flush what's pending within the budget, fail the
111    /// acknowledgements of anything abandoned, and report. For multi-sink
112    /// pipelines this is the composed hook that drains every sink.
113    pub drain: SinkDrainFn,
114    /// Optional connectivity probe (e.g. `SinkPool::probe_all`). The
115    /// runtime probes at startup and then periodically, driving the
116    /// sinks-connected half of `/readyz`. Without a probe the flag is set
117    /// unconditionally.
118    pub probe: Option<SinkProbeFn>,
119}
120
121impl std::fmt::Debug for SinkRuntime {
122    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
123        f.debug_struct("SinkRuntime")
124            .field("queues", &self.queues)
125            .finish_non_exhaustive()
126    }
127}
128
129/// Boxed sink drain hook — defined next to the sink layer that produces
130/// it, re-exported here where the runtime consumes it.
131pub use crate::sink::{SinkDrainFn, SinkProbeFn};
132
133/// Terminal state of a pipeline run.
134#[derive(Clone, Debug, PartialEq, Eq)]
135#[non_exhaustive]
136pub enum ExitState {
137    /// Drained and committed cleanly: a requested shutdown (SIGTERM or
138    /// programmatic), or a bounded source reporting
139    /// [`SourceEvent::Drained`](crate::source::SourceEvent::Drained) after
140    /// exhausting its input. For the drained case this additionally
141    /// guarantees every batch was acknowledged and the final watermark
142    /// commit persisted (anything less exits `Failed`).
143    Completed,
144    /// A fatal error stopped the pipeline; the process should exit
145    /// non-zero.
146    Failed(FatalErrorReport),
147}
148
149/// Owned copy of the fatal error carried in the exit report.
150#[derive(Clone, Debug, PartialEq, Eq)]
151#[non_exhaustive]
152pub struct FatalErrorReport {
153    /// Component that failed.
154    pub component: String,
155    /// Human-readable cause.
156    pub reason: String,
157}
158
159impl std::fmt::Display for FatalErrorReport {
160    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
161        write!(f, "pipeline failed in {}: {}", self.component, self.reason)
162    }
163}
164
165impl std::error::Error for FatalErrorReport {}
166
167/// Outcome of [`PipelineRuntime::run`].
168#[derive(Debug)]
169#[non_exhaustive]
170pub struct ExitReport {
171    /// How the run ended.
172    pub state: ExitState,
173    /// The sink's drain report (absent when the sink drain hook could not
174    /// run, e.g. the I/O runtime was already gone).
175    pub sink_drain: Option<DrainReport>,
176    /// The last committed watermark per partition, as reported by the
177    /// final commit.
178    pub final_watermarks: Vec<(PartitionId, i64)>,
179}
180
181impl ExitReport {
182    /// Log the outcome — state, sink drain, final watermarks — at the level
183    /// matching it (`info` for a clean exit, `error` for a failure).
184    pub fn log(&self) {
185        match &self.state {
186            ExitState::Completed => tracing::info!(
187                state = ?self.state,
188                drain = ?self.sink_drain,
189                watermarks = ?self.final_watermarks,
190                "pipeline finished"
191            ),
192            ExitState::Failed(failure) => tracing::error!(
193                component = %failure.component,
194                reason = %failure.reason,
195                drain = ?self.sink_drain,
196                watermarks = ?self.final_watermarks,
197                "pipeline failed"
198            ),
199        }
200    }
201
202    /// The process exit code this outcome maps to: `0` for a clean exit,
203    /// `1` for a failure.
204    #[must_use]
205    pub fn exit_code(&self) -> i32 {
206        match self.state {
207            ExitState::Completed => 0,
208            ExitState::Failed(_) => 1,
209        }
210    }
211
212    /// The report as a `Result`, so a `main` can `?` a failed run: a clean
213    /// exit passes the report through, a failure returns the
214    /// [`FatalErrorReport`] as the error.
215    pub fn ok(self) -> Result<ExitReport, FatalErrorReport> {
216        match &self.state {
217            ExitState::Completed => Ok(self),
218            ExitState::Failed(failure) => Err(failure.clone()),
219        }
220    }
221}
222
223#[cfg(all(test, not(loom)))]
224pub(crate) mod fakes;
225#[cfg(all(test, not(loom)))]
226mod tests;