Demonstrate lazy scheduling (#754)

frankmcsherry · web-flow · commit 09d0b87dae40 · 2026-03-13T09:33:49.000-04:00
* Demonstrate lazy scheduling

* Add FrontierInterest with IfCapability variant

* Move FrontierInterest to be per-input

* Remove default notify_me impl; streamline setting the value
diff --git a/timely/examples/event_driven.rs b/timely/examples/event_driven.rs
@@ -1,4 +1,5 @@
-use timely::dataflow::operators::{Input, Probe};
+use timely::dataflow::Scope;
+use timely::dataflow::operators::{Input, Probe, Enter, Leave};
 use timely::dataflow::operators::vec::Map;
 
 fn main() {
@@ -20,10 +21,14 @@ fn main() {
         // create a new input, exchange data, and inspect its output
         for _dataflow in 0 .. dataflows {
             worker.dataflow(|scope| {
-                let (input, mut stream) = scope.new_input();
-                for _step in 0 .. length {
-                    stream = stream.map(|x: ()| x);
-                }
+                let (input, stream) = scope.new_input();
+                let stream = scope.region(|inner| {
+                    let mut stream = stream.enter(inner);
+                    for _step in 0 .. length {
+                        stream = stream.map(|x: ()| x);
+                    }
+                    stream.leave()
+                });
                 let (probe, _stream) = stream.probe();
                 inputs.push(input);
                 probes.push(probe);
@@ -43,7 +48,7 @@ fn main() {
                 worker.step();
                 steps += 1;
             }
-            println!("{:?}\tround {} complete in {} steps", timer.elapsed(), round, steps);
+            if round % 1000 == 0 { println!("{:?}\tround {} complete in {} steps", timer.elapsed(), round, steps); }
         }
 
     }).unwrap();
diff --git a/timely/src/dataflow/operators/core/concat.rs b/timely/src/dataflow/operators/core/concat.rs
@@ -63,10 +63,10 @@ impl<G: Scope, C: Container> Concatenate<G, C> for G {
         // create an operator builder.
         use crate::dataflow::operators::generic::builder_rc::OperatorBuilder;
         let mut builder = OperatorBuilder::new("Concatenate".to_string(), self.clone());
-        builder.set_notify(false);
 
         // create new input handles for each input stream.
         let mut handles = sources.into_iter().map(|s| builder.new_input(s, Pipeline)).collect::<Vec<_>>();
+        for i in 0 .. handles.len() { builder.set_notify_for(i, crate::progress::operate::FrontierInterest::Never); }
 
         // create one output handle for the concatenated results.
         let (mut output, result) = builder.new_output();
diff --git a/timely/src/dataflow/operators/core/feedback.rs b/timely/src/dataflow/operators/core/feedback.rs
@@ -73,7 +73,6 @@ impl<G: Scope> Feedback<G> for G {
     fn feedback<C: Container>(&mut self, summary: <G::Timestamp as Timestamp>::Summary) -> (Handle<G, C>, Stream<G, C>) {
 
         let mut builder = OperatorBuilder::new("Feedback".to_owned(), self.clone());
-        builder.set_notify(false);
         let (output, stream) = builder.new_output();
 
         (Handle { builder, summary, output }, stream)
@@ -118,6 +117,7 @@ impl<G: Scope, C: Container> ConnectLoop<G, C> for Stream<G, C> {
         let mut output = handle.output;
 
         let mut input = builder.new_input_connection(self, Pipeline, [(0, Antichain::from_elem(summary.clone()))]);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
 
         builder.build(move |_capability| move |_frontier| {
             let mut output = output.activate();
diff --git a/timely/src/dataflow/operators/core/input.rs b/timely/src/dataflow/operators/core/input.rs
@@ -199,6 +199,7 @@ impl<T:Timestamp> Schedule for Operator<T> {
     }
 }
 
+use crate::progress::operate::FrontierInterest;
 impl<T:Timestamp> Operate<T> for Operator<T> {
 
     fn inputs(&self) -> usize { 0 }
@@ -209,7 +210,7 @@ impl<T:Timestamp> Operate<T> for Operator<T> {
         (Vec::new(), Rc::clone(&self.shared_progress), self)
     }
 
-    fn notify_me(&self) -> bool { false }
+    fn notify_me(&self) -> &[FrontierInterest] { &[] }
 }
 
 
diff --git a/timely/src/dataflow/operators/core/ok_err.rs b/timely/src/dataflow/operators/core/ok_err.rs
@@ -52,9 +52,9 @@ impl<S: Scope, C: Container + DrainContainer> OkErr<S, C> for Stream<S, C> {
         L: FnMut(C::Item<'_>) -> Result<D1,D2>+'static
     {
         let mut builder = OperatorBuilder::new("OkErr".to_owned(), self.scope());
-        builder.set_notify(false);
 
         let mut input = builder.new_input(self, Pipeline);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
         let (output1, stream1) = builder.new_output();
         let (output2, stream2) = builder.new_output();
 
diff --git a/timely/src/dataflow/operators/core/partition.rs b/timely/src/dataflow/operators/core/partition.rs
@@ -41,9 +41,9 @@ impl<G: Scope, C: Container + DrainContainer> Partition<G, C> for Stream<G, C> {
         F: FnMut(C::Item<'_>) -> (u64, D2) + 'static,
     {
         let mut builder = OperatorBuilder::new("Partition".to_owned(), self.scope());
-        builder.set_notify(false);
 
         let mut input = builder.new_input(self, Pipeline);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
         let mut outputs = Vec::with_capacity(parts as usize);
         let mut streams = Vec::with_capacity(parts as usize);
 
diff --git a/timely/src/dataflow/operators/core/unordered_input.rs b/timely/src/dataflow/operators/core/unordered_input.rs
@@ -129,6 +129,7 @@ impl<T:Timestamp> Schedule for UnorderedOperator<T> {
     }
 }
 
+use crate::progress::operate::FrontierInterest;
 impl<T:Timestamp> Operate<T> for UnorderedOperator<T> {
     fn inputs(&self) -> usize { 0 }
     fn outputs(&self) -> usize { 1 }
@@ -140,7 +141,7 @@ impl<T:Timestamp> Operate<T> for UnorderedOperator<T> {
         (Vec::new(), Rc::clone(&self.shared_progress), self)
     }
 
-    fn notify_me(&self) -> bool { false }
+    fn notify_me(&self) -> &[FrontierInterest] { &[] }
 }
 
 /// A handle to an input [Stream], used to introduce data to a timely dataflow computation.
diff --git a/timely/src/dataflow/operators/generic/builder_raw.rs b/timely/src/dataflow/operators/generic/builder_raw.rs
@@ -12,7 +12,7 @@ use crate::scheduling::{Schedule, Activations};
 
 use crate::progress::{Source, Target};
 use crate::progress::{Timestamp, Operate, operate::SharedProgress, Antichain};
-use crate::progress::operate::{Connectivity, PortConnectivity};
+use crate::progress::operate::{FrontierInterest, Connectivity, PortConnectivity};
 use crate::Container;
 use crate::dataflow::{Stream, Scope};
 use crate::dataflow::channels::pushers::Tee;
@@ -23,7 +23,7 @@ use crate::dataflow::operators::generic::operator_info::OperatorInfo;
 #[derive(Debug)]
 pub struct OperatorShape {
     name: String,   // A meaningful name for the operator.
-    notify: bool,   // Does the operator require progress notifications.
+    notify: Vec<FrontierInterest>,   // Per-input frontier interest.
     peers: usize,   // The total number of workers in the computation. Needed to initialize pointstamp counts with the correct magnitude.
     inputs: usize,  // The number of input ports.
     outputs: usize, // The number of output ports.
@@ -34,7 +34,7 @@ impl OperatorShape {
     fn new(name: String, peers: usize) -> Self {
         OperatorShape {
             name,
-            notify: true,
+            notify: Vec::new(),
             peers,
             inputs: 0,
             outputs: 0,
@@ -88,8 +88,10 @@ impl<G: Scope> OperatorBuilder<G> {
     /// Return a reference to the operator's shape
     pub fn shape(&self) -> &OperatorShape { &self.shape }
 
-    /// Indicates whether the operator requires frontier information.
-    pub fn set_notify(&mut self, notify: bool) { self.shape.notify = notify; }
+    /// Sets frontier interest for a specific input.
+    pub fn set_notify_for(&mut self, input: usize, notify: FrontierInterest) {
+        self.shape.notify[input] = notify;
+    }
 
     /// Adds a new input to a generic operator builder, returning the `Pull` implementor to use.
     pub fn new_input<C: Container, P>(&mut self, stream: Stream<G, C>, pact: P) -> P::Puller
@@ -113,6 +115,7 @@ impl<G: Scope> OperatorBuilder<G> {
         stream.connect_to(target, sender, channel_id);
 
         self.shape.inputs += 1;
+        self.shape.notify.push(FrontierInterest::Always);
         let connectivity: PortConnectivity<_> = connection.into_iter().collect();
         assert!(connectivity.iter_ports().all(|(o,_)| o < self.shape.outputs));
         self.summary.push(connectivity);
@@ -220,5 +223,5 @@ where
         (self.summary.clone(), Rc::clone(&self.shared_progress), self)
     }
 
-    fn notify_me(&self) -> bool { self.shape.notify }
+    fn notify_me(&self) -> &[FrontierInterest] { &self.shape.notify }
 }
diff --git a/timely/src/dataflow/operators/generic/builder_rc.rs b/timely/src/dataflow/operators/generic/builder_rc.rs
@@ -18,7 +18,7 @@ use crate::dataflow::operators::capability::Capability;
 use crate::dataflow::operators::generic::handles::{InputHandleCore, new_input_handle};
 use crate::dataflow::operators::generic::operator_info::OperatorInfo;
 use crate::dataflow::operators::generic::builder_raw::OperatorShape;
-use crate::progress::operate::PortConnectivity;
+use crate::progress::operate::{FrontierInterest, PortConnectivity};
 
 use super::builder_raw::OperatorBuilder as OperatorBuilderRaw;
 
@@ -48,8 +48,10 @@ impl<G: Scope> OperatorBuilder<G> {
         }
     }
 
-    /// Indicates whether the operator requires frontier information.
-    pub fn set_notify(&mut self, notify: bool) { self.builder.set_notify(notify); }
+    /// Sets frontier interest for a specific input.
+    pub fn set_notify_for(&mut self, input: usize, notify: FrontierInterest) {
+        self.builder.set_notify_for(input, notify);
+    }
 
     /// Adds a new input to a generic operator builder, returning the `Pull` implementor to use.
     pub fn new_input<C: Container, P>(&mut self, stream: Stream<G, C>, pact: P) -> InputHandleCore<G::Timestamp, C, P::Puller>
diff --git a/timely/src/dataflow/operators/generic/operator.rs b/timely/src/dataflow/operators/generic/operator.rs
@@ -373,9 +373,9 @@ impl<G: Scope, C1: Container> Operator<G, C1> for Stream<G, C1> {
         let operator_info = builder.operator_info();
 
         let mut input = builder.new_input(self, pact);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
         let (output, stream) = builder.new_output();
         let mut output = OutputBuilder::from(output);
-        builder.set_notify(false);
 
         builder.build(move |mut capabilities| {
             // `capabilities` should be a single-element vector.
@@ -461,9 +461,10 @@ impl<G: Scope, C1: Container> Operator<G, C1> for Stream<G, C1> {
 
         let mut input1 = builder.new_input(self, pact1);
         let mut input2 = builder.new_input(other, pact2);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
+        builder.set_notify_for(1, crate::progress::operate::FrontierInterest::Never);
         let (output, stream) = builder.new_output();
         let mut output = OutputBuilder::from(output);
-        builder.set_notify(false);
 
         builder.build(move |mut capabilities| {
             // `capabilities` should be a single-element vector.
@@ -547,7 +548,6 @@ where
 
     let (output, stream) = builder.new_output();
     let mut output = OutputBuilder::from(output);
-    builder.set_notify(false);
 
     builder.build(move |mut capabilities| {
         // `capabilities` should be a single-element vector.
diff --git a/timely/src/dataflow/operators/vec/branch.rs b/timely/src/dataflow/operators/vec/branch.rs
@@ -41,9 +41,9 @@ impl<S: Scope, D: 'static> Branch<S, D> for StreamVec<S, D> {
         condition: impl Fn(&S::Timestamp, &D) -> bool + 'static,
     ) -> (StreamVec<S, D>, StreamVec<S, D>) {
         let mut builder = OperatorBuilder::new("Branch".to_owned(), self.scope());
-        builder.set_notify(false);
 
         let mut input = builder.new_input(self, Pipeline);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
         let (output1, stream1) = builder.new_output();
         let (output2, stream2) = builder.new_output();
 
@@ -102,9 +102,9 @@ pub trait BranchWhen<T>: Sized {
 impl<S: Scope, C: Container> BranchWhen<S::Timestamp> for Stream<S, C> {
     fn branch_when(self, condition: impl Fn(&S::Timestamp) -> bool + 'static) -> (Self, Self) {
         let mut builder = OperatorBuilder::new("Branch".to_owned(), self.scope());
-        builder.set_notify(false);
 
         let mut input = builder.new_input(self, Pipeline);
+        builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);
         let (output1, stream1) = builder.new_output();
         let (output2, stream2) = builder.new_output();
 
diff --git a/timely/src/progress/operate.rs b/timely/src/progress/operate.rs
@@ -55,12 +55,23 @@ pub trait Operate<T: Timestamp> {
     /// safely "create" capabilities without basing them on other, prior capabilities.
     fn initialize(self: Box<Self>) -> (Connectivity<T::Summary>, Rc<RefCell<SharedProgress<T>>>, Box<dyn Schedule>);
 
-    /// Indicates if the operator should be invoked on the basis of input frontier transitions.
+    /// Indicates for each input whether the operator should be invoked when that input's frontier changes.
     ///
-    /// This value is conservatively set to `true`, but operators that know they are oblivious to
-    /// frontier information can indicate this with `false`, and they will not be scheduled on the
-    /// basis of their input frontiers changing.
-    fn notify_me(&self) -> bool { true }
+    /// Returns a `Vec<FrontierInterest>` with one entry per input. Each entry describes whether
+    /// frontier changes on that input should cause the operator to be scheduled. The conservative
+    /// default is `Always` for each input.
+    fn notify_me(&self) -> &[FrontierInterest];// { &vec![FrontierInterest::Always; self.inputs()] }
+}
+
+/// The ways in which an operator can express interest in activation when an input frontier changes.
+#[derive(Ord, PartialOrd, Eq, PartialEq, Copy, Clone, Debug)]
+pub enum FrontierInterest {
+    /// Never interested in frontier changes, as for example the `map()` and `filter()` operators.
+    Never,
+    /// Interested when the operator holds capabilities.
+    IfCapability,
+    /// Always interested in frontier changes, as for example the `probe()` and `capture()` operators.
+    Always,
 }
 
 /// Operator internal connectivity, from inputs to outputs.
diff --git a/timely/src/progress/reachability.rs b/timely/src/progress/reachability.rs
@@ -38,6 +38,7 @@
 //! let mut results =
 //! tracker
 //!     .pushed()
+//!     .0
 //!     .drain()
 //!     .filter(|((location, time), delta)| location.is_target())
 //!     .collect::<Vec<_>>();
@@ -60,6 +61,7 @@
 //! let mut results =
 //! tracker
 //!     .pushed()
+//!     .0
 //!     .drain()
 //!     .filter(|((location, time), delta)| location.is_target())
 //!     .collect::<Vec<_>>();
@@ -413,6 +415,8 @@ pub struct PerOperator<T: Timestamp> {
     pub targets: Vec<PortInformation<T>>,
     /// Port information for each source.
     pub sources: Vec<PortInformation<T>>,
+    /// Sum across outputs of capabilities.
+    pub cap_counts: i64,
 }
 
 impl<T: Timestamp> PerOperator<T> {
@@ -421,6 +425,7 @@ impl<T: Timestamp> PerOperator<T> {
         PerOperator {
             targets: vec![PortInformation::new(); inputs],
             sources: vec![PortInformation::new(); outputs],
+            cap_counts: 0,
         }
     }
 }
@@ -618,13 +623,15 @@ impl<T:Timestamp> Tracker<T> {
         while let Some(((source, _), _)) = source_changes.peek() {
 
             let source = *source;
-            let operator = &mut self.per_operator[source.node].sources[source.port];
+            let operator = &mut self.per_operator[source.node];
+            let op_source = &mut operator.sources[source.port];
             let source_updates = source_changes.peeking_take_while(|((s, _),_)| s == &source).map(|((_,time),diff)| (time,diff));
-            let changes = operator.pointstamps.update_iter(source_updates);
+            let changes = op_source.pointstamps.update_iter(source_updates);
 
             for (time, diff) in changes {
                 self.total_counts += diff;
-                for (output, summaries) in operator.output_summaries.iter_ports() {
+                operator.cap_counts += diff;
+                for (output, summaries) in op_source.output_summaries.iter_ports() {
                     let output_changes = &mut self.output_changes[output];
                     summaries
                         .elements()
@@ -708,8 +715,8 @@ impl<T:Timestamp> Tracker<T> {
     }
 
     /// A mutable reference to the pushed results of changes.
-    pub fn pushed(&mut self) -> &mut ChangeBatch<(Location, T)> {
-        &mut self.pushed_changes
+    pub fn pushed(&mut self) -> (&mut ChangeBatch<(Location, T)>, &[PerOperator<T>]) {
+        (&mut self.pushed_changes, &self.per_operator)
     }
 
     /// Reveals per-operator frontier state.
diff --git a/timely/src/progress/subgraph.rs b/timely/src/progress/subgraph.rs

Original file line number	Diff line number	Diff line change
`@@ -199,6 +199,7 @@ impl<T:Timestamp> Schedule for Operator<T> {`
`199`	`199`	`}`
`200`	`200`	`}`
`201`	`201`
	`202`	`+use crate::progress::operate::FrontierInterest;`
`202`	`203`	`impl<T:Timestamp> Operate<T> for Operator<T> {`
`203`	`204`
`204`	`205`	`fn inputs(&self) -> usize { 0 }`
`@@ -209,7 +210,7 @@ impl<T:Timestamp> Operate<T> for Operator<T> {`
`209`	`210`	`(Vec::new(), Rc::clone(&self.shared_progress), self)`
`210`	`211`	`}`
`211`	`212`
`212`		`- fn notify_me(&self) -> bool { false }`
	`213`	`+ fn notify_me(&self) -> &[FrontierInterest] { &[] }`
`213`	`214`	`}`
`214`	`215`
`215`	`216`
Original file line number	Diff line number	Diff line change
`@@ -52,9 +52,9 @@ impl<S: Scope, C: Container + DrainContainer> OkErr<S, C> for Stream<S, C> {`
`52`	`52`	`L: FnMut(C::Item<'_>) -> Result<D1,D2>+'static`
`53`	`53`	`{`
`54`	`54`	`let mut builder = OperatorBuilder::new("OkErr".to_owned(), self.scope());`
`55`		`- builder.set_notify(false);`
`56`	`55`
`57`	`56`	`let mut input = builder.new_input(self, Pipeline);`
	`57`	`+ builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);`
`58`	`58`	`let (output1, stream1) = builder.new_output();`
`59`	`59`	`let (output2, stream2) = builder.new_output();`
`60`	`60`
Original file line number	Diff line number	Diff line change
`@@ -41,9 +41,9 @@ impl<G: Scope, C: Container + DrainContainer> Partition<G, C> for Stream<G, C> {`
`41`	`41`	`F: FnMut(C::Item<'_>) -> (u64, D2) + 'static,`
`42`	`42`	`{`
`43`	`43`	`let mut builder = OperatorBuilder::new("Partition".to_owned(), self.scope());`
`44`		`- builder.set_notify(false);`
`45`	`44`
`46`	`45`	`let mut input = builder.new_input(self, Pipeline);`
	`46`	`+ builder.set_notify_for(0, crate::progress::operate::FrontierInterest::Never);`
`47`	`47`	`let mut outputs = Vec::with_capacity(parts as usize);`
`48`	`48`	`let mut streams = Vec::with_capacity(parts as usize);`
`49`	`49`
Original file line number	Diff line number	Diff line change
`@@ -129,6 +129,7 @@ impl<T:Timestamp> Schedule for UnorderedOperator<T> {`
`129`	`129`	`}`
`130`	`130`	`}`
`131`	`131`
	`132`	`+use crate::progress::operate::FrontierInterest;`
`132`	`133`	`impl<T:Timestamp> Operate<T> for UnorderedOperator<T> {`
`133`	`134`	`fn inputs(&self) -> usize { 0 }`
`134`	`135`	`fn outputs(&self) -> usize { 1 }`
`@@ -140,7 +141,7 @@ impl<T:Timestamp> Operate<T> for UnorderedOperator<T> {`
`140`	`141`	`(Vec::new(), Rc::clone(&self.shared_progress), self)`
`141`	`142`	`}`
`142`	`143`
`143`		`- fn notify_me(&self) -> bool { false }`
	`144`	`+ fn notify_me(&self) -> &[FrontierInterest] { &[] }`
`144`	`145`	`}`
`145`	`146`
`146`	`147`	`/// A handle to an input [Stream], used to introduce data to a timely dataflow computation.`