Trait timely::dataflow::scopes::Scope

pub trait Scope: ScopeParent {
    // Required methods
    fn name(&self) -> String;
    fn addr(&self) -> Rc<[usize]>;
    fn addr_for_child(&self, index: usize) -> Rc<[usize]>;
    fn add_edge(&self, source: Source, target: Target);
    fn allocate_operator_index(&mut self) -> usize;
    fn add_operator_with_indices(
        &mut self,
        operator: Box<dyn Operate<Self::Timestamp>>,
        local: usize,
        global: usize,
    );
    fn scoped<T, R, F>(&mut self, name: &str, func: F) -> R
       where T: Timestamp + Refines<<Self as ScopeParent>::Timestamp>,
             F: FnOnce(&mut Child<'_, Self, T>) -> R;

    // Provided methods
    fn add_operator(
        &mut self,
        operator: Box<dyn Operate<Self::Timestamp>>,
    ) -> usize { ... }
    fn add_operator_with_index(
        &mut self,
        operator: Box<dyn Operate<Self::Timestamp>>,
        index: usize,
    ) { ... }
    fn iterative<T, R, F>(&mut self, func: F) -> R
       where T: Timestamp,
             F: FnOnce(&mut Child<'_, Self, Product<<Self as ScopeParent>::Timestamp, T>>) -> R { ... }
    fn region<R, F>(&mut self, func: F) -> R
       where F: FnOnce(&mut Child<'_, Self, <Self as ScopeParent>::Timestamp>) -> R { ... }
    fn region_named<R, F>(&mut self, name: &str, func: F) -> R
       where F: FnOnce(&mut Child<'_, Self, <Self as ScopeParent>::Timestamp>) -> R { ... }
}

The fundamental operations required to add and connect operators in a timely dataflow graph.

Importantly, this is often a shared object, backed by a Rc<RefCell<>> wrapper. Each method takes a shared reference, but can be thought of as first calling .clone() and then calling the method. Each method does not hold the RefCell’s borrow, and should prevent accidental panics.

Required Methods

fn name(&self) -> String

A useful name describing the scope.

fn addr(&self) -> Rc<[usize]>

A sequence of scope identifiers describing the path from the worker root to this scope.

fn addr_for_child(&self, index: usize) -> Rc<[usize]>

A sequence of scope identifiers describing the path from the worker root to the child indicated by index.

fn add_edge(&self, source: Source, target: Target)

Connects a source of data with a target of the data. This only links the two for the purposes of tracking progress, rather than effect any data movement itself.

fn allocate_operator_index(&mut self) -> usize

Allocates a new scope-local operator index.

This method is meant for use with add_operator_with_index, which accepts a scope-local operator index allocated with this method. This method does cause the scope to expect that an operator will be added, and it is an error not to eventually add such an operator.

fn add_operator_with_indices( &mut self, operator: Box<dyn Operate<Self::Timestamp>>, local: usize, global: usize, )

Adds a child Operate to the builder’s scope using supplied indices.

The two indices are the scope-local operator index, and a worker-unique index used for e.g. logging.

fn scoped<T, R, F>(&mut self, name: &str, func: F) -> R

where T: Timestamp + Refines<<Self as ScopeParent>::Timestamp>, F: FnOnce(&mut Child<'_, Self, T>) -> R,

Creates a dataflow subgraph.

This method allows the user to create a nested scope with any timestamp that “refines” the enclosing timestamp (informally: extends it in a reversible way).

This is most commonly used to create new iterative contexts, and the provided method iterative for this task demonstrates the use of this method.

Examples

use timely::dataflow::Scope;
use timely::dataflow::operators::{Input, Enter, Leave};
use timely::order::Product;

timely::execute_from_args(std::env::args(), |worker| {
    // must specify types as nothing else drives inference.
    let input = worker.dataflow::<u64,_,_>(|child1| {
        let (input, stream) = child1.new_input::<String>();
        let output = child1.scoped::<Product<u64,u32>,_,_>("ScopeName", |child2| {
            stream.enter(child2).leave()
        });
        input
    });
});

Provided Methods

fn add_operator(&mut self, operator: Box<dyn Operate<Self::Timestamp>>) -> usize

Adds a child Operate to the builder’s scope. Returns the new child’s index.

fn add_operator_with_index( &mut self, operator: Box<dyn Operate<Self::Timestamp>>, index: usize, )

Adds a child Operate to the builder’s scope using a supplied index.

This is used internally when there is a gap between allocate a child identifier and adding the child, as happens in subgraph creation.

fn iterative<T, R, F>(&mut self, func: F) -> R

where T: Timestamp, F: FnOnce(&mut Child<'_, Self, Product<<Self as ScopeParent>::Timestamp, T>>) -> R,

Creates a iterative dataflow subgraph.

This method is a specialization of scoped which uses the Product timestamp combinator, suitable for iterative computations in which iterative development at some time cannot influence prior iterations at a future time.

Examples

use timely::dataflow::Scope;
use timely::dataflow::operators::{Input, Enter, Leave};

timely::execute_from_args(std::env::args(), |worker| {
    // must specify types as nothing else drives inference.
    let input = worker.dataflow::<u64,_,_>(|child1| {
        let (input, stream) = child1.new_input::<String>();
        let output = child1.iterative::<u32,_,_>(|child2| {
            stream.enter(child2).leave()
        });
        input
    });
});

fn region<R, F>(&mut self, func: F) -> R

where F: FnOnce(&mut Child<'_, Self, <Self as ScopeParent>::Timestamp>) -> R,

Creates a dataflow region with the same timestamp.

This method is a specialization of scoped which uses the same timestamp as the containing scope. It is used mainly to group regions of a dataflow computation, and provides some computational benefits by abstracting the specifics of the region.

Examples

use timely::dataflow::Scope;
use timely::dataflow::operators::{Input, Enter, Leave};

timely::execute_from_args(std::env::args(), |worker| {
    // must specify types as nothing else drives inference.
    let input = worker.dataflow::<u64,_,_>(|child1| {
        let (input, stream) = child1.new_input::<String>();
        let output = child1.region(|child2| {
            stream.enter(child2).leave()
        });
        input
    });
});

fn region_named<R, F>(&mut self, name: &str, func: F) -> R

where F: FnOnce(&mut Child<'_, Self, <Self as ScopeParent>::Timestamp>) -> R,

Creates a dataflow region with the same timestamp.

This method is a specialization of scoped which uses the same timestamp as the containing scope. It is used mainly to group regions of a dataflow computation, and provides some computational benefits by abstracting the specifics of the region.

This variant allows you to specify a name for the region, which can be read out in the timely logging streams.

Examples

use timely::dataflow::Scope;
use timely::dataflow::operators::{Input, Enter, Leave};

timely::execute_from_args(std::env::args(), |worker| {
    // must specify types as nothing else drives inference.
    let input = worker.dataflow::<u64,_,_>(|child1| {
        let (input, stream) = child1.new_input::<String>();
        let output = child1.region_named("region", |child2| {
            stream.enter(child2).leave()
        });
        input
    });
});

Object Safety

This trait is not object safe.

Implementors

impl<'a, G, T> Scope for Child<'a, G, T>

where G: ScopeParent, T: Timestamp + Refines<G::Timestamp>,