differential_dataflow/operators/

join.rs

//! Match pairs of records based on a key.
//!
//! The various `join` implementations require that the units of each collection can be multiplied, and that
//! the multiplication distributes over addition. That is, we will repeatedly evaluate (a + b) * c as (a * c)
//! + (b * c), and if this is not equal to the former term, little is known about the actual output.
use std::cmp::Ordering;

use timely::{Accountable, ContainerBuilder};
use timely::container::PushInto;
use timely::order::PartialOrder;
use timely::progress::Timestamp;
use timely::dataflow::Stream;
use timely::dataflow::operators::generic::{Operator, OutputBuilderSession, Session};
use timely::dataflow::channels::pact::Pipeline;
use timely::dataflow::operators::Capability;

use crate::lattice::Lattice;
use crate::operators::arrange::Arranged;
use crate::trace::{BatchReader, Cursor};
use crate::operators::ValueHistory;

use crate::trace::TraceReader;

/// The session passed to join closures.
pub type JoinSession<'a, 'b, T, CB, CT> = Session<'a, 'b, T, EffortBuilder<CB>, CT>;

/// A container builder that tracks the length of outputs to estimate the effort of join closures.
#[derive(Default, Debug)]
pub struct EffortBuilder<CB>(pub std::cell::Cell<usize>, pub CB);

impl<CB: ContainerBuilder> timely::container::ContainerBuilder for EffortBuilder<CB> {
    type Container = CB::Container;

    #[inline]
    fn extract(&mut self) -> Option<&mut Self::Container> {
        let extracted = self.1.extract();
        self.0.replace(self.0.take() + extracted.as_ref().map_or(0, |e| e.record_count() as usize));
        extracted
    }

    #[inline]
    fn finish(&mut self) -> Option<&mut Self::Container> {
        let finished = self.1.finish();
        self.0.replace(self.0.take() + finished.as_ref().map_or(0, |e| e.record_count() as usize));
        finished
    }
}

impl<CB: PushInto<D>, D> PushInto<D> for EffortBuilder<CB> {
    #[inline]
    fn push_into(&mut self, item: D) {
        self.1.push_into(item);
    }
}

/// An equijoin of two traces, sharing a common key type.
///
/// This method exists to provide join functionality without opinions on the specific input types, keys and values,
/// that should be presented. The two traces here can have arbitrary key and value types, which can be unsized and
/// even potentially unrelated to the input collection data. Importantly, the key and value types could be generic
/// associated types (GATs) of the traces, and we would seemingly struggle to frame these types as trait arguments.
///
/// The implementation produces a caller-specified container. Implementations can use [`AsCollection`] to wrap the
/// output stream in a collection.
///
/// The "correctness" of this method depends heavily on the behavior of the supplied `result` function.
///
/// [`AsCollection`]: crate::collection::AsCollection
pub fn join_traces<'scope, Tr1, Tr2, L, CB>(arranged1: Arranged<'scope, Tr1>, arranged2: Arranged<'scope, Tr2>, mut result: L) -> Stream<'scope, Tr1::Time, CB::Container>
where
    Tr1: TraceReader+Clone+'static,
    Tr2: for<'a> TraceReader<Key<'a>=Tr1::Key<'a>, Time = Tr1::Time>+Clone+'static,
    L: FnMut(Tr1::Key<'_>,Tr1::Val<'_>,Tr2::Val<'_>,&Tr1::Time,&Tr1::Diff,&Tr2::Diff,&mut JoinSession<Tr1::Time, CB, Capability<Tr1::Time>>)+'static,
    CB: ContainerBuilder,
{
    // Rename traces for symmetry from here on out.
    let mut trace1 = arranged1.trace.clone();
    let mut trace2 = arranged2.trace.clone();

    let scope = arranged1.stream.scope();
    arranged1.stream.binary_frontier(arranged2.stream, Pipeline, Pipeline, "Join", move |capability, info| {

        // Acquire an activator to reschedule the operator when it has unfinished work.
        use timely::scheduling::Activator;
        let activations = scope.activations().clone();
        let activator = Activator::new(info.address, activations);

        // Our initial invariants are that for each trace, physical compaction is less or equal the trace's upper bound.
        // These invariants ensure that we can reference observed batch frontiers from `_start_upper` onward, as long as
        // we maintain our physical compaction capabilities appropriately. These assertions are tested as we load up the
        // initial work for the two traces, and before the operator is constructed.

        // Acknowledged frontier for each input.
        // These two are used exclusively to track batch boundaries on which we may want/need to call `cursor_through`.
        // They will drive our physical compaction of each trace, and we want to maintain at all times that each is beyond
        // the physical compaction frontier of their corresponding trace.
        // Should we ever *drop* a trace, these are 1. much harder to maintain correctly, but 2. no longer used.
        use timely::progress::frontier::Antichain;
        let mut acknowledged1 = Antichain::from_elem(Tr1::Time::minimum());
        let mut acknowledged2 = Antichain::from_elem(Tr1::Time::minimum());

        // deferred work of batches from each input.
        let mut todo1 = std::collections::VecDeque::new();
        let mut todo2 = std::collections::VecDeque::new();

        // We'll unload the initial batches here, to put ourselves in a less non-deterministic state to start.
        trace1.map_batches(|batch1| {
            acknowledged1.clone_from(batch1.upper());
            // No `todo1` work here, because we haven't accepted anything into `batches2` yet.
            // It is effectively "empty", because we choose to drain `trace1` before `trace2`.
            // Once we start streaming batches in, we will need to respond to new batches from
            // `input1` with logic that would have otherwise been here. Check out the next loop
            // for the structure.
        });
        // At this point, `ack1` should exactly equal `trace1.read_upper()`, as they are both determined by
        // iterating through batches and capturing the upper bound. This is a great moment to assert that
        // `trace1`'s physical compaction frontier is before the frontier of completed times in `trace1`.
        // TODO: in the case that this does not hold, instead "upgrade" the physical compaction frontier.
        assert!(PartialOrder::less_equal(&trace1.get_physical_compaction(), &acknowledged1.borrow()));

        // We capture batch2 cursors first and establish work second to avoid taking a `RefCell` lock
        // on both traces at the same time, as they could be the same trace and this would panic.
        let mut batch2_cursors = Vec::new();
        trace2.map_batches(|batch2| {
            acknowledged2.clone_from(batch2.upper());
            batch2_cursors.push((batch2.cursor(), batch2.clone()));
        });
        // At this point, `ack2` should exactly equal `trace2.read_upper()`, as they are both determined by
        // iterating through batches and capturing the upper bound. This is a great moment to assert that
        // `trace2`'s physical compaction frontier is before the frontier of completed times in `trace2`.
        // TODO: in the case that this does not hold, instead "upgrade" the physical compaction frontier.
        assert!(PartialOrder::less_equal(&trace2.get_physical_compaction(), &acknowledged2.borrow()));

        // Load up deferred work using trace2 cursors and batches captured just above.
        for (batch2_cursor, batch2) in batch2_cursors.into_iter() {
            // It is safe to ask for `ack1` because we have confirmed it to be in advance of `distinguish_since`.
            let (trace1_cursor, trace1_storage) = trace1.cursor_through(acknowledged1.borrow()).unwrap();
            // We could downgrade the capability here, but doing so is a bit complicated mathematically.
            // TODO: downgrade the capability by searching out the one time in `batch2.lower()` and not
            // in `batch2.upper()`. Only necessary for non-empty batches, as empty batches may not have
            // that property.
            todo2.push_back(Deferred::new(trace1_cursor, trace1_storage, batch2_cursor, batch2.clone(), capability.clone()));
        }

        // Droppable handles to shared trace data structures.
        let mut trace1_option = Some(trace1);
        let mut trace2_option = Some(trace2);

        move |(input1, frontier1), (input2, frontier2), output| {

            // 1. Consuming input.
            //
            // The join computation repeatedly accepts batches of updates from each of its inputs.
            //
            // For each accepted batch, it prepares a work-item to join the batch against previously "accepted"
            // updates from its other input. It is important to track which updates have been accepted, because
            // we use a shared trace and there may be updates present that are in advance of this accepted bound.
            //
            // Batches are accepted: 1. in bulk at start-up (above), 2. as we observe them in the input stream,
            // and 3. if the trace can confirm a region of empty space directly following our accepted bound.
            // This last case is a consequence of our inability to transmit empty batches, as they may be formed
            // in the absence of timely dataflow capabilities.

            // Drain input 1, prepare work.
            input1.for_each(|capability, data| {
                // This test *should* always pass, as we only drop a trace in response to the other input emptying.
                if let Some(ref mut trace2) = trace2_option {
                    let capability = capability.retain(0);
                    for batch1 in data.drain(..) {
                        // Ignore any pre-loaded data.
                        if PartialOrder::less_equal(&acknowledged1, batch1.lower()) {
                            if !batch1.is_empty() {
                                // It is safe to ask for `ack2` as we validated that it was at least `get_physical_compaction()`
                                // at start-up, and have held back physical compaction ever since.
                                let (trace2_cursor, trace2_storage) = trace2.cursor_through(acknowledged2.borrow()).unwrap();
                                let batch1_cursor = batch1.cursor();
                                todo1.push_back(Deferred::new(trace2_cursor, trace2_storage, batch1_cursor, batch1.clone(), capability.clone()));
                            }

                            // To update `acknowledged1` we might presume that `batch1.lower` should equal it, but we
                            // may have skipped over empty batches. Still, the batches are in-order, and we should be
                            // able to just assume the most recent `batch1.upper`
                            debug_assert!(PartialOrder::less_equal(&acknowledged1, batch1.upper()));
                            acknowledged1.clone_from(batch1.upper());
                        }
                    }
                }
                else { panic!("`trace2_option` dropped before `input1` emptied!"); }
            });

            // Drain input 2, prepare work.
            input2.for_each(|capability, data| {
                // This test *should* always pass, as we only drop a trace in response to the other input emptying.
                if let Some(ref mut trace1) = trace1_option {
                    let capability = capability.retain(0);
                    for batch2 in data.drain(..) {
                        // Ignore any pre-loaded data.
                        if PartialOrder::less_equal(&acknowledged2, batch2.lower()) {
                            if !batch2.is_empty() {
                                // It is safe to ask for `ack1` as we validated that it was at least `get_physical_compaction()`
                                // at start-up, and have held back physical compaction ever since.
                                let (trace1_cursor, trace1_storage) = trace1.cursor_through(acknowledged1.borrow()).unwrap();
                                let batch2_cursor = batch2.cursor();
                                todo2.push_back(Deferred::new(trace1_cursor, trace1_storage, batch2_cursor, batch2.clone(), capability.clone()));
                            }

                            // To update `acknowledged2` we might presume that `batch2.lower` should equal it, but we
                            // may have skipped over empty batches. Still, the batches are in-order, and we should be
                            // able to just assume the most recent `batch2.upper`
                            debug_assert!(PartialOrder::less_equal(&acknowledged2, batch2.upper()));
                            acknowledged2.clone_from(batch2.upper());
                        }
                    }
                }
                else { panic!("`trace1_option` dropped before `input2` emptied!"); }
            });

            // Advance acknowledged frontiers through any empty regions that we may not receive as batches.
            if let Some(trace1) = trace1_option.as_mut() {
                trace1.advance_upper(&mut acknowledged1);
            }
            if let Some(trace2) = trace2_option.as_mut() {
                trace2.advance_upper(&mut acknowledged2);
            }

            // 2. Join computation.
            //
            // For each of the inputs, we do some amount of work (measured in terms of number
            // of output records produced). This is meant to yield control to allow downstream
            // operators to consume and reduce the output, but it it also means to provide some
            // degree of responsiveness. There is a potential risk here that if we fall behind
            // then the increasing queues hold back physical compaction of the underlying traces
            // which results in unintentionally quadratic processing time (each batch of either
            // input must scan all batches from the other input).

            // Perform some amount of outstanding work.
            let mut fuel = 1_000_000;
            while !todo1.is_empty() && fuel > 0 {
                todo1.front_mut().unwrap().work(
                    output,
                    |k,v2,v1,t,r2,r1,c| result(k,v1,v2,t,r1,r2,c),
                    &mut fuel
                );
                if !todo1.front().unwrap().work_remains() { todo1.pop_front(); }
            }

            // Perform some amount of outstanding work.
            let mut fuel = 1_000_000;
            while !todo2.is_empty() && fuel > 0 {
                todo2.front_mut().unwrap().work(
                    output,
                    |k,v1,v2,t,r1,r2,c| result(k,v1,v2,t,r1,r2,c),
                    &mut fuel
                );
                if !todo2.front().unwrap().work_remains() { todo2.pop_front(); }
            }

            // Re-activate operator if work remains.
            if !todo1.is_empty() || !todo2.is_empty() {
                activator.activate();
            }

            // 3. Trace maintenance.
            //
            // Importantly, we use `input.frontier()` here rather than `acknowledged` to track
            // the progress of an input, because should we ever drop one of the traces we will
            // lose the ability to extract information from anything other than the input.
            // For example, if we dropped `trace2` we would not be able to use `advance_upper`
            // to keep `acknowledged2` up to date wrt empty batches, and would hold back logical
            // compaction of `trace1`.

            // Maintain `trace1`. Drop if `input2` is empty, or advance based on future needs.
            if let Some(trace1) = trace1_option.as_mut() {
                if frontier2.is_empty() { trace1_option = None; }
                else {
                    // Allow `trace1` to compact logically up to the frontier we may yet receive,
                    // in the opposing input (`input2`). All `input2` times will be beyond this
                    // frontier, and joined times only need to be accurate when advanced to it.
                    trace1.set_logical_compaction(frontier2.frontier());
                    // Allow `trace1` to compact physically up to the upper bound of batches we
                    // have received in its input (`input1`). We will not require a cursor that
                    // is not beyond this bound.
                    trace1.set_physical_compaction(acknowledged1.borrow());
                }
            }

            // Maintain `trace2`. Drop if `input1` is empty, or advance based on future needs.
            if let Some(trace2) = trace2_option.as_mut() {
                if frontier1.is_empty() { trace2_option = None;}
                else {
                    // Allow `trace2` to compact logically up to the frontier we may yet receive,
                    // in the opposing input (`input1`). All `input1` times will be beyond this
                    // frontier, and joined times only need to be accurate when advanced to it.
                    trace2.set_logical_compaction(frontier1.frontier());
                    // Allow `trace2` to compact physically up to the upper bound of batches we
                    // have received in its input (`input2`). We will not require a cursor that
                    // is not beyond this bound.
                    trace2.set_physical_compaction(acknowledged2.borrow());
                }
            }
        }
    })
}


/// Deferred join computation.
///
/// The structure wraps cursors which allow us to play out join computation at whatever rate we like.
/// This allows us to avoid producing and buffering massive amounts of data, without giving the timely
/// dataflow system a chance to run operators that can consume and aggregate the data.
struct Deferred<T, C1, C2>
where
    T: Timestamp+Lattice+Ord,
    C1: Cursor<Time=T>,
    C2: for<'a> Cursor<Key<'a>=C1::Key<'a>, Time=T>,
{
    trace: C1,
    trace_storage: C1::Storage,
    batch: C2,
    batch_storage: C2::Storage,
    capability: Capability<T>,
    done: bool,
}

impl<T, C1, C2> Deferred<T, C1, C2>
where
    C1: Cursor<Time=T>,
    C2: for<'a> Cursor<Key<'a>=C1::Key<'a>, Time=T>,
    T: Timestamp+Lattice+Ord,
{
    fn new(trace: C1, trace_storage: C1::Storage, batch: C2, batch_storage: C2::Storage, capability: Capability<T>) -> Self {
        Deferred {
            trace,
            trace_storage,
            batch,
            batch_storage,
            capability,
            done: false,
        }
    }

    fn work_remains(&self) -> bool {
        !self.done
    }

    /// Process keys until at least `fuel` output tuples produced, or the work is exhausted.
    #[inline(never)]
    fn work<L, CB: ContainerBuilder>(&mut self, output: &mut OutputBuilderSession<T, EffortBuilder<CB>>, mut logic: L, fuel: &mut usize)
    where
        L: for<'a> FnMut(C1::Key<'a>, C1::Val<'a>, C2::Val<'a>, &T, &C1::Diff, &C2::Diff, &mut JoinSession<T, CB, Capability<T>>),
    {

        let meet = self.capability.time();

        let mut effort = 0;
        let mut session = output.session_with_builder(&self.capability);

        let trace_storage = &self.trace_storage;
        let batch_storage = &self.batch_storage;

        let trace = &mut self.trace;
        let batch = &mut self.batch;

        let mut thinker = JoinThinker::new();

        while let (Some(batch_key), Some(trace_key), true) = (batch.get_key(batch_storage), trace.get_key(trace_storage), effort < *fuel) {

            match trace_key.cmp(&batch_key) {
                Ordering::Less => trace.seek_key(trace_storage, batch_key),
                Ordering::Greater => batch.seek_key(batch_storage, trace_key),
                Ordering::Equal => {

                    thinker.history1.edits.load(trace, trace_storage, |time| {
                        let mut time = C1::owned_time(time);
                        time.join_assign(meet);
                        time
                    });
                    thinker.history2.edits.load(batch, batch_storage, |time| C2::owned_time(time));

                    // populate `temp` with the results in the best way we know how.
                    thinker.think(|v1,v2,t,r1,r2| {
                        logic(batch_key, v1, v2, &t, r1, r2, &mut session);
                    });

                    // TODO: Effort isn't perfectly tracked as we might still have some data in the
                    // session at the moment it's dropped.
                    effort += session.builder().0.take();
                    batch.step_key(batch_storage);
                    trace.step_key(trace_storage);

                    thinker.history1.clear();
                    thinker.history2.clear();
                }
            }
        }
        self.done = !batch.key_valid(batch_storage) || !trace.key_valid(trace_storage);

        if effort > *fuel { *fuel = 0; }
        else              { *fuel -= effort; }
    }
}

struct JoinThinker<'a, C1, C2>
where
    C1: Cursor,
    C2: Cursor<Time = C1::Time>,
{
    pub history1: ValueHistory<'a, C1>,
    pub history2: ValueHistory<'a, C2>,
}

impl<'a, C1, C2> JoinThinker<'a, C1, C2>
where
    C1: Cursor,
    C2: Cursor<Time = C1::Time>,
{
    fn new() -> Self {
        JoinThinker {
            history1: ValueHistory::new(),
            history2: ValueHistory::new(),
        }
    }

    fn think<F: FnMut(C1::Val<'a>,C2::Val<'a>,C1::Time,&C1::Diff,&C2::Diff)>(&mut self, mut results: F) {

        // for reasonably sized edits, do the dead-simple thing.
        if self.history1.edits.len() < 10 || self.history2.edits.len() < 10 {
            self.history1.edits.map(|v1, t1, d1| {
                self.history2.edits.map(|v2, t2, d2| {
                    results(v1, v2, t1.join(t2), d1, d2);
                })
            })
        }
        else {

            let mut replay1 = self.history1.replay();
            let mut replay2 = self.history2.replay();

            // TODO: It seems like there is probably a good deal of redundant `advance_buffer_by`
            //       in here. If a time is ever repeated, for example, the call will be identical
            //       and accomplish nothing. If only a single record has been added, it may not
            //       be worth the time to collapse (advance, re-sort) the data when a linear scan
            //       is sufficient.

            while !replay1.is_done() && !replay2.is_done() {

                if replay1.time().unwrap().cmp(replay2.time().unwrap()) == ::std::cmp::Ordering::Less {
                    replay2.advance_buffer_by(replay1.meet().unwrap());
                    for &((val2, ref time2), ref diff2) in replay2.buffer().iter() {
                        let (val1, time1, diff1) = replay1.edit().unwrap();
                        results(val1, val2, time1.join(time2), diff1, diff2);
                    }
                    replay1.step();
                }
                else {
                    replay1.advance_buffer_by(replay2.meet().unwrap());
                    for &((val1, ref time1), ref diff1) in replay1.buffer().iter() {
                        let (val2, time2, diff2) = replay2.edit().unwrap();
                        results(val1, val2, time1.join(time2), diff1, diff2);
                    }
                    replay2.step();
                }
            }

            while !replay1.is_done() {
                replay2.advance_buffer_by(replay1.meet().unwrap());
                for &((val2, ref time2), ref diff2) in replay2.buffer().iter() {
                    let (val1, time1, diff1) = replay1.edit().unwrap();
                    results(val1, val2, time1.join(time2), diff1, diff2);
                }
                replay1.step();
            }
            while !replay2.is_done() {
                replay1.advance_buffer_by(replay2.meet().unwrap());
                for &((val1, ref time1), ref diff1) in replay1.buffer().iter() {
                    let (val2, time2, diff2) = replay2.edit().unwrap();
                    results(val1, val2, time1.join(time2), diff1, diff2);
                }
                replay2.step();
            }
        }
    }
}