polars_pipe/pipeline/dispatcher/
mod.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
use std::cell::RefCell;
use std::fmt::{Debug, Formatter};
use std::rc::Rc;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Arc, Mutex};

use polars_core::error::PolarsResult;
use polars_core::utils::accumulate_dataframes_vertical_unchecked;
use polars_core::POOL;
use polars_expr::state::ExecutionState;
use polars_utils::sync::SyncPtr;
use rayon::prelude::*;

use crate::executors::sources::DataFrameSource;
use crate::operators::{
    DataChunk, FinalizedSink, OperatorResult, PExecutionContext, Sink, SinkResult, Source,
    SourceResult,
};
use crate::pipeline::dispatcher::drive_operator::{par_flush, par_process_chunks};
mod drive_operator;
use super::*;

pub(super) struct ThreadedSink {
    /// A sink split per thread.
    pub sinks: Vec<Box<dyn Sink>>,
    /// when that hits 0, the sink will finalize
    pub shared_count: Rc<RefCell<u32>>,
    initial_shared_count: u32,
    /// - offset in the operators vec
    ///   at that point the sink should be called.
    ///   the pipeline will first call the operators on that point and then
    ///   push the result in the sink.
    pub operator_end: usize,
}

impl ThreadedSink {
    pub fn new(sink: Box<dyn Sink>, shared_count: Rc<RefCell<u32>>, operator_end: usize) -> Self {
        let n_threads = morsels_per_sink();
        let sinks = (0..n_threads).map(|i| sink.split(i)).collect();
        let initial_shared_count = *shared_count.borrow();
        ThreadedSink {
            sinks,
            initial_shared_count,
            shared_count,
            operator_end,
        }
    }

    // Only the first node of a shared sink should recurse. The others should return.
    fn allow_recursion(&self) -> bool {
        self.initial_shared_count == *self.shared_count.borrow()
    }
}

/// A pipeline consists of:
///
/// - 1. One or more sources.
///         Sources get pulled and their data is pushed into operators.
/// - 2. Zero or more operators.
///         The operators simply pass through data, modifying it as they need.
///         Operators can work on batches and don't need all data in scope to
///         succeed.
///         Think for example on multiply a few columns, or applying a predicate.
///         Operators can shrink the batches: filter
///         Grow the batches: explode/ unpivot
///         Keep them the same size: element-wise operations
///         The probe side of join operations is also an operator.
///
///
/// - 3. One or more sinks
///         A sink needs all data in scope to finalize a pipeline branch.
///         Think of sorts, preparing a build phase of a join, group_by + aggregations.
///
/// This struct will have the SOS (source, operators, sinks) of its own pipeline branch, but also
/// the SOS of other branches. The SOS are stored data oriented and the sinks have an offset that
/// indicates the last operator node before that specific sink. We only store the `end offset` and
/// keep track of the starting operator during execution.
///
/// Pipelines branches are shared with other pipeline branches at the join/union nodes.
/// # JOIN
/// Consider this tree:
///         out
///       /
///     /\
///    1  2
///
/// And let's consider that branch 2 runs first. It will run until the join node where it will sink
/// into a build table. Once that is done it will replace the build-phase placeholder operator in
/// branch 1. Branch one can then run completely until out.
pub struct PipeLine {
    /// All the sources of this pipeline
    sources: Vec<Box<dyn Source>>,
    /// All the operators of this pipeline. Some may be placeholders that will be replaced during
    /// execution
    operators: Vec<ThreadedOperator>,
    /// - offset in the operators vec
    ///   at that point the sink should be called.
    ///   the pipeline will first call the operators on that point and then
    ///   push the result in the sink.
    /// - shared_count
    ///     when that hits 0, the sink will finalize
    /// - node of the sink
    sinks: Vec<ThreadedSink>,
    /// Log runtime info to stderr
    verbose: bool,
}

impl PipeLine {
    #[allow(clippy::type_complexity)]
    pub(super) fn new(
        sources: Vec<Box<dyn Source>>,
        operators: Vec<PhysOperator>,
        sinks: Vec<ThreadedSink>,
        verbose: bool,
    ) -> PipeLine {
        // we don't use the power of two partition size here
        // we only do that in the sinks itself.
        let n_threads = morsels_per_sink();

        // We split so that every thread gets an operator
        // every index maps to a chain of operators than can be pushed as a pipeline for one thread
        let operators = (0..n_threads)
            .map(|i| {
                operators
                    .iter()
                    .map(|op| op.get_ref().split(i).into())
                    .collect()
            })
            .collect();

        PipeLine {
            sources,
            operators,
            sinks,
            verbose,
        }
    }

    /// Create a pipeline only consisting of a single branch that always finishes with a sink
    pub(crate) fn new_simple(
        sources: Vec<Box<dyn Source>>,
        operators: Vec<PhysOperator>,
        sink: Box<dyn Sink>,
        verbose: bool,
    ) -> Self {
        let operators_len = operators.len();
        Self::new(
            sources,
            operators,
            vec![ThreadedSink::new(
                sink,
                Rc::new(RefCell::new(1)),
                operators_len,
            )],
            verbose,
        )
    }

    /// Replace the current sources with a [`DataFrameSource`].
    fn set_df_as_sources(&mut self, df: DataFrame) {
        let src = Box::new(DataFrameSource::from_df(df)) as Box<dyn Source>;
        self.set_sources(src)
    }

    /// Replace the current sources.
    fn set_sources(&mut self, src: Box<dyn Source>) {
        self.sources.clear();
        self.sources.push(src);
    }

    fn run_pipeline_no_finalize(
        &mut self,
        ec: &PExecutionContext,
        pipelines: &mut Vec<PipeLine>,
    ) -> PolarsResult<(u32, Box<dyn Sink>)> {
        let mut out = None;
        let mut operator_start = 0;
        let last_i = self.sinks.len() - 1;

        // For unions we typically first want to push all pipelines
        // into the union sink before we call `finalize`
        // however if the sink is finished early, (for instance a `head`)
        // we don't want to run the rest of the pipelines and we finalize early
        let mut sink_finished = false;

        for (i, mut sink) in std::mem::take(&mut self.sinks).into_iter().enumerate() {
            for src in &mut std::mem::take(&mut self.sources) {
                let mut next_batches = src.get_batches(ec)?;

                let must_flush: AtomicBool = AtomicBool::new(false);
                while let SourceResult::GotMoreData(chunks) = next_batches {
                    // Every batches iteration we check if we must continue.
                    ec.execution_state.should_stop()?;

                    let (sink_result, next_batches2) = par_process_chunks(
                        chunks,
                        &mut sink.sinks,
                        ec,
                        &mut self.operators,
                        operator_start,
                        sink.operator_end,
                        src,
                        &must_flush,
                    )?;
                    next_batches = next_batches2;

                    if let Some(SinkResult::Finished) = sink_result {
                        sink_finished = true;
                        break;
                    }
                }
                if !sink_finished && must_flush.load(Ordering::Relaxed) {
                    par_flush(
                        &mut sink.sinks,
                        ec,
                        &mut self.operators,
                        operator_start,
                        sink.operator_end,
                    );
                }
            }

            // Before we reduce we also check if we should continue.
            ec.execution_state.should_stop()?;
            let allow_recursion = sink.allow_recursion();

            // The sinks have taken all chunks thread locally, now we reduce them into a single
            // result sink.
            let mut reduced_sink = POOL
                .install(|| {
                    sink.sinks.into_par_iter().reduce_with(|mut a, mut b| {
                        a.combine(&mut *b);
                        a
                    })
                })
                .unwrap();
            operator_start = sink.operator_end;

            let mut shared_sink_count = {
                let mut shared_sink_count = sink.shared_count.borrow_mut();
                *shared_sink_count -= 1;
                *shared_sink_count
            };

            // Prevent very deep recursion. Only the outer callee can pop and run.
            if allow_recursion {
                while shared_sink_count > 0 && !sink_finished {
                    let mut pipeline = pipelines.pop().unwrap();
                    let (count, mut sink) = pipeline.run_pipeline_no_finalize(ec, pipelines)?;
                    // This branch is hit when we have a Union of joins.
                    // The build side must be converted into an operator and replaced in the next pipeline.

                    // Check either:
                    // 1. There can be a union source that sinks into a single join:
                    //      scan_parquet(*) -> join B
                    // 2. There can be a union of joins
                    //      C - JOIN A, B
                    //      concat (A, B, C)
                    //
                    // So to ensure that we don't finalize we check
                    // - They are not both join builds
                    // - If they are both join builds, check they are note the same build, otherwise
                    //   we must call the `combine` branch.
                    if sink.is_join_build()
                        && (!reduced_sink.is_join_build() || (sink.node() != reduced_sink.node()))
                    {
                        let FinalizedSink::Operator = sink.finalize(ec)? else {
                            unreachable!()
                        };
                    } else {
                        reduced_sink.combine(sink.as_mut());
                        shared_sink_count = count;
                    }
                }
            }

            if i != last_i {
                let sink_result = reduced_sink.finalize(ec)?;
                match sink_result {
                    // turn this sink an a new source
                    FinalizedSink::Finished(df) => self.set_df_as_sources(df),
                    FinalizedSink::Source(src) => self.set_sources(src),
                    // should not happen
                    FinalizedSink::Operator => {
                        unreachable!()
                    },
                }
            } else {
                out = Some((shared_sink_count, reduced_sink))
            }
        }
        Ok(out.unwrap())
    }

    /// Run a single pipeline branch.
    /// This pulls data from the sources and pushes it into the operators which run on a different
    /// thread and finalize in a sink.
    ///
    /// The sink can be finished, but can also become a new source and then rinse and repeat.
    pub fn run_pipeline(
        &mut self,
        ec: &PExecutionContext,
        pipelines: &mut Vec<PipeLine>,
    ) -> PolarsResult<Option<FinalizedSink>> {
        let (sink_shared_count, mut reduced_sink) = self.run_pipeline_no_finalize(ec, pipelines)?;
        assert_eq!(sink_shared_count, 0);

        let finalized_reduced_sink = reduced_sink.finalize(ec)?;
        Ok(Some(finalized_reduced_sink))
    }
}

/// Executes all branches and replaces operators and sinks during execution to ensure
/// we materialize.
pub fn execute_pipeline(
    state: ExecutionState,
    mut pipelines: Vec<PipeLine>,
) -> PolarsResult<DataFrame> {
    let mut pipeline = pipelines.pop().unwrap();
    let ec = PExecutionContext::new(state, pipeline.verbose);

    let mut sink_out = pipeline.run_pipeline(&ec, &mut pipelines)?;
    loop {
        match &mut sink_out {
            None => {
                let mut pipeline = pipelines.pop().unwrap();
                sink_out = pipeline.run_pipeline(&ec, &mut pipelines)?;
            },
            Some(FinalizedSink::Finished(df)) => return Ok(std::mem::take(df)),
            Some(FinalizedSink::Source(src)) => return consume_source(&mut **src, &ec),

            //
            //  1/\
            //   2/\
            //     3\
            // the left hand side of the join has finished and now is an operator
            // we replace the dummy node in the right hand side pipeline with this
            // operator and then we run the pipeline rinse and repeat
            // until the final right hand side pipeline ran
            Some(FinalizedSink::Operator) => {
                // we unwrap, because the latest pipeline should not return an Operator
                let mut pipeline = pipelines.pop().unwrap();

                sink_out = pipeline.run_pipeline(&ec, &mut pipelines)?;
            },
        }
    }
}

impl Debug for PipeLine {
    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        let mut fmt = String::new();
        let mut start = 0usize;
        fmt.push_str(self.sources[0].fmt());
        for sink in &self.sinks {
            fmt.push_str(" -> ");
            // take operators of a single thread
            let ops = &self.operators[0];
            // slice the pipeline
            let ops = &ops[start..sink.operator_end];
            for op in ops {
                fmt.push_str(op.get_ref().fmt());
                fmt.push_str(" -> ")
            }
            start = sink.operator_end;
            fmt.push_str(sink.sinks[0].fmt())
        }
        write!(f, "{fmt}")
    }
}

/// Take a source and materialize it into a [`DataFrame`].
fn consume_source(src: &mut dyn Source, context: &PExecutionContext) -> PolarsResult<DataFrame> {
    let mut frames = Vec::with_capacity(32);

    while let SourceResult::GotMoreData(batch) = src.get_batches(context)? {
        frames.extend(batch.into_iter().map(|chunk| chunk.data))
    }
    Ok(accumulate_dataframes_vertical_unchecked(frames))
}

unsafe impl Send for PipeLine {}
unsafe impl Sync for PipeLine {}