palimpsest_dataflow/operators/
consolidate.rs

1//! Aggregates the weights of equal records into at most one record.
2//!
3//! As differential dataflow streams are unordered and taken to be the accumulation of all records,
4//! no semantic change happens via `consolidate`. However, there is a practical difference between
5//! a collection that aggregates down to zero records, and one that actually has no records. The
6//! underlying system can more clearly see that no work must be done in the later case, and we can
7//! drop out of, e.g. iterative computations.
8
9use timely::dataflow::Scope;
10
11use crate::consolidation::ConsolidatingContainerBuilder;
12use crate::difference::Semigroup;
13use crate::{ExchangeData, Hashable, VecCollection};
14
15use crate::lattice::Lattice;
16use crate::trace::{Batcher, Builder};
17use crate::Data;
18
19/// Methods which require data be arrangeable.
20impl<G, D, R> VecCollection<G, D, R>
21where
22    G: Scope<Timestamp: Data + Lattice>,
23    D: ExchangeData + Hashable,
24    R: Semigroup + ExchangeData,
25{
26    /// Aggregates the weights of equal records into at most one record.
27    ///
28    /// This method uses the type `D`'s `hashed()` method to partition the data. The data are
29    /// accumulated in place, each held back until their timestamp has completed.
30    ///
31    /// # Examples
32    ///
33    /// ```
34    /// use palimpsest_dataflow::input::Input;
35    ///
36    /// ::timely::example(|scope| {
37    ///
38    ///     let x = scope.new_collection_from(1 .. 10u32).1;
39    ///
40    ///     x.negate()
41    ///      .concat(&x)
42    ///      .consolidate() // <-- ensures cancellation occurs
43    ///      .assert_empty();
44    /// });
45    /// ```
46    pub fn consolidate(&self) -> Self {
47        use crate::trace::implementations::{KeyBatcher, KeyBuilder, KeySpine};
48        self.consolidate_named::<KeyBatcher<_, _, _>, KeyBuilder<_, _, _>, KeySpine<_, _, _>, _>(
49            "Consolidate",
50            |key, &()| key.clone(),
51        )
52    }
53
54    /// As `consolidate` but with the ability to name the operator, specify the trace type,
55    /// and provide the function `reify` to produce owned keys and values..
56    pub fn consolidate_named<Ba, Bu, Tr, F>(&self, name: &str, reify: F) -> Self
57    where
58        Ba: Batcher<Input = Vec<((D, ()), G::Timestamp, R)>, Time = G::Timestamp> + 'static,
59        Tr: for<'a> crate::trace::Trace<Time = G::Timestamp, Diff = R> + 'static,
60        Bu: Builder<Time = Tr::Time, Input = Ba::Output, Output = Tr::Batch>,
61        F: Fn(Tr::Key<'_>, Tr::Val<'_>) -> D + 'static,
62    {
63        use crate::operators::arrange::arrangement::Arrange;
64        self.map(|k| (k, ()))
65            .arrange_named::<Ba, Bu, Tr>(name)
66            .as_collection(reify)
67    }
68
69    /// Aggregates the weights of equal records.
70    ///
71    /// Unlike `consolidate`, this method does not exchange data and does not
72    /// ensure that at most one copy of each `(data, time)` pair exists in the
73    /// results. Instead, it acts on each batch of data and collapses equivalent
74    /// `(data, time)` pairs found therein, suppressing any that accumulate to
75    /// zero.
76    ///
77    /// # Examples
78    ///
79    /// ```
80    /// use palimpsest_dataflow::input::Input;
81    ///
82    /// ::timely::example(|scope| {
83    ///
84    ///     let x = scope.new_collection_from(1 .. 10u32).1;
85    ///
86    ///     // nothing to assert, as no particular guarantees.
87    ///     x.negate()
88    ///      .concat(&x)
89    ///      .consolidate_stream();
90    /// });
91    /// ```
92    pub fn consolidate_stream(&self) -> Self {
93        use crate::collection::AsCollection;
94        use timely::dataflow::channels::pact::Pipeline;
95        use timely::dataflow::operators::Operator;
96
97        self.inner
98            .unary::<ConsolidatingContainerBuilder<_>, _, _, _>(
99                Pipeline,
100                "ConsolidateStream",
101                |_cap, _info| {
102                    move |input, output| {
103                        input.for_each(|time, data| {
104                            output
105                                .session_with_builder(&time)
106                                .give_iterator(data.drain(..));
107                        })
108                    }
109                },
110            )
111            .as_collection()
112    }
113}
palimpsest_dataflow/operators/consolidate.rs

palimpsest_dataflow/operators/
consolidate.rs
