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
// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.

//! Transformation based on pushing demand information about columns toward sources.

use itertools::Itertools;
use std::collections::{BTreeMap, BTreeSet};

use mz_expr::{
    AggregateExpr, AggregateFunc, Id, JoinInputMapper, MirRelationExpr, MirScalarExpr,
    RECURSION_LIMIT,
};
use mz_ore::stack::{CheckedRecursion, RecursionGuard};
use mz_repr::{Datum, Row};

use crate::TransformCtx;

/// Drive demand from the root through operators.
///
/// This transform alerts operators to their columns that influence the
/// ultimate output of the expression, and gives them permission to swap
/// other columns for dummy values. As part of this, operators should not
/// actually use any of these dummy values, lest they run-time error.
///
/// This transformation primarily informs the `Join` operator, which can
/// simplify its intermediate state when it knows that certain columns are
/// not observed in its output. Internal arrangements need not maintain
/// columns that are no longer required in the join pipeline, which are
/// those columns not required by the output nor any further equalities.
///
/// Nowadays, this transform is mostly obsoleted by `ProjectionPushdown`.
/// However, I know of one thing that it still does that `ProjectionPushdown`
/// doesn't do (might possibly be more such things):
/// if you have something like
// ```
//     Project (#0, #1)
//       Join on=(#0 = #1)
// ```
// then this is turned into
// ```
//     Project (#0, #0)
//       Join on=(#0 = #1)
// ```
// This can be beneficial for projecting out some columns earlier inside a complex join (by the LIR
// planning), and then recovering them after the join (if needed) by duplicating existing columns.
#[derive(Debug)]
pub struct Demand {
    recursion_guard: RecursionGuard,
}

impl Default for Demand {
    fn default() -> Demand {
        Demand {
            recursion_guard: RecursionGuard::with_limit(RECURSION_LIMIT),
        }
    }
}

impl CheckedRecursion for Demand {
    fn recursion_guard(&self) -> &RecursionGuard {
        &self.recursion_guard
    }
}

impl crate::Transform for Demand {
    #[mz_ore::instrument(
        target = "optimizer",
        level = "debug",
        fields(path.segment = "demand")
    )]
    fn transform(
        &self,
        relation: &mut MirRelationExpr,
        _: &mut TransformCtx,
    ) -> Result<(), crate::TransformError> {
        let result = self.action(
            relation,
            (0..relation.arity()).collect(),
            &mut BTreeMap::new(),
        );
        mz_repr::explain::trace_plan(&*relation);
        result
    }
}

impl Demand {
    /// Columns to be produced.
    fn action(
        &self,
        relation: &mut MirRelationExpr,
        mut columns: BTreeSet<usize>,
        gets: &mut BTreeMap<Id, BTreeSet<usize>>,
    ) -> Result<(), crate::TransformError> {
        self.checked_recur(|_| {
            // A valid relation type is only needed for Maps, but we can't borrow
            // the relation in the corresponding branch of the match statement, since
            // it is already borrowed mutably.
            let relation_type = if matches!(relation, MirRelationExpr::Map { .. }) {
                Some(relation.typ())
            } else {
                None
            };
            match relation {
                MirRelationExpr::Constant { .. } => {
                    // Nothing clever to do with constants, that I can think of.
                    Ok(())
                }
                MirRelationExpr::Get { id, .. } => {
                    gets.entry(*id)
                        .or_insert_with(BTreeSet::new)
                        .extend(columns);
                    Ok(())
                }
                MirRelationExpr::Let { id, value, body } => {
                    // Let harvests any requirements of get from its body,
                    // and pushes the union of the requirements at its value.
                    let id = Id::Local(*id);
                    let prior = gets.insert(id, BTreeSet::new());
                    assert!(prior.is_none()); // no shadowing
                    self.action(body, columns, gets)?;
                    let needs = gets.remove(&id).expect("existing gets entry");
                    if let Some(prior) = prior {
                        gets.insert(id, prior);
                    }

                    self.action(value, needs, gets)
                }
                MirRelationExpr::LetRec {
                    ids,
                    values,
                    limits: _,
                    body,
                } => {
                    let ids_used_across_iterations = MirRelationExpr::recursive_ids(ids, values)
                        .iter()
                        .map(|id| Id::Local(*id))
                        .collect::<BTreeSet<_>>();
                    let ids = ids.iter().map(|id| Id::Local(*id)).collect_vec();
                    for id in ids.iter() {
                        let prior = gets.insert(id.clone(), BTreeSet::new());
                        assert!(prior.is_none()); // no shadowing
                    }
                    self.action(body, columns, gets)?;
                    for (id, value) in ids.iter().rev().zip_eq(values.iter_mut().rev()) {
                        let needs = if !ids_used_across_iterations.contains(id) {
                            gets.remove(id).expect("existing gets entry")
                        } else {
                            // Remove, but ignore the collected needs
                            gets.remove(id).expect("existing gets entry");
                            // Instead of using `gets`, we'll say we need all columns for a
                            // recursive id
                            (0..value.arity()).collect::<BTreeSet<_>>()
                        };
                        self.action(value, needs, gets)?;
                    }
                    Ok(())
                }
                MirRelationExpr::Project { input, outputs } => self.action(
                    input,
                    columns.into_iter().map(|c| outputs[c]).collect(),
                    gets,
                ),
                MirRelationExpr::Map { input, scalars } => {
                    let relation_type = relation_type.as_ref().unwrap();
                    let arity = input.arity();
                    // contains columns whose supports have yet to be explored
                    let mut new_columns = columns.clone();
                    new_columns.retain(|c| *c >= arity);
                    while !new_columns.is_empty() {
                        // explore supports
                        new_columns = new_columns
                            .iter()
                            .flat_map(|c| scalars[*c - arity].support())
                            .filter(|c| !columns.contains(c))
                            .collect();
                        // add those columns to the seen list
                        columns.extend(new_columns.clone());
                        new_columns.retain(|c| *c >= arity);
                    }

                    // Replace un-read expressions with literals to prevent evaluation.
                    for (index, scalar) in scalars.iter_mut().enumerate() {
                        if !columns.contains(&(arity + index)) {
                            // Leave literals as they are, to benefit explain.
                            if !scalar.is_literal() {
                                let typ = relation_type.column_types[arity + index].clone();
                                *scalar = MirScalarExpr::Literal(
                                    Ok(Row::pack_slice(&[Datum::Dummy])),
                                    typ,
                                );
                            }
                        }
                    }

                    columns.retain(|c| *c < arity);
                    self.action(input, columns, gets)
                }
                MirRelationExpr::FlatMap {
                    input,
                    func: _,
                    exprs,
                } => {
                    // A FlatMap which returns zero rows acts like a filter
                    // so we always need to execute it
                    for expr in exprs {
                        expr.support_into(&mut columns);
                    }
                    columns.retain(|c| *c < input.arity());
                    self.action(input, columns, gets)
                }
                MirRelationExpr::Filter { input, predicates } => {
                    for predicate in predicates {
                        predicate.support_into(&mut columns)
                    }
                    self.action(input, columns, gets)
                }
                MirRelationExpr::Join {
                    inputs,
                    equivalences,
                    implementation: _,
                } => {
                    let input_mapper = JoinInputMapper::new(inputs);

                    // Each produced column that is equivalent to a prior column should be remapped
                    // so that upstream uses depend only on the first column, simplifying the demand
                    // analysis. In principle we could choose any representative, if it turns out
                    // that some other column would have been more helpful, but we don't have a great
                    // reason to do that at the moment.
                    let mut permutation: Vec<usize> = (0..input_mapper.total_columns()).collect();
                    for equivalence in equivalences.iter() {
                        let mut first_column = None;
                        for expr in equivalence.iter() {
                            if let MirScalarExpr::Column(c) = expr {
                                if let Some(prior) = &first_column {
                                    permutation[*c] = *prior;
                                } else {
                                    first_column = Some(*c);
                                }
                            }
                        }
                    }

                    let should_permute = columns.iter().any(|c| permutation[*c] != *c);

                    // Each equivalence class imposes internal demand for columns.
                    for equivalence in equivalences.iter() {
                        for expr in equivalence.iter() {
                            expr.support_into(&mut columns);
                        }
                    }

                    // Populate child demands from external and internal demands.
                    let new_columns = input_mapper.split_column_set_by_input(columns.iter());

                    // Recursively indicate the requirements.
                    for (input, columns) in inputs.iter_mut().zip(new_columns) {
                        self.action(input, columns, gets)?;
                    }

                    // Install a permutation if any demanded column is not the
                    // canonical column.
                    if should_permute {
                        *relation = relation.take_dangerous().project(permutation);
                    }

                    Ok(())
                }
                MirRelationExpr::Reduce {
                    input,
                    group_key,
                    aggregates,
                    monotonic: _,
                    expected_group_size: _,
                } => {
                    let mut new_columns = BTreeSet::new();
                    // Group keys determine aggregation granularity and are
                    // each crucial in determining aggregates and even the
                    // multiplicities of other keys.
                    for k in group_key.iter() {
                        k.support_into(&mut new_columns)
                    }
                    for column in columns.iter() {
                        // No obvious requirements on aggregate columns.
                        // A "non-empty" requirement, I guess?
                        if *column >= group_key.len() {
                            aggregates[*column - group_key.len()]
                                .expr
                                .support_into(&mut new_columns);
                        }
                    }

                    // Replace un-demanded aggregations with dummies.
                    let input_type = input.typ();
                    for index in (0..aggregates.len()).rev() {
                        if !columns.contains(&(group_key.len() + index)) {
                            let typ = aggregates[index].typ(&input_type.column_types);
                            aggregates[index] = AggregateExpr {
                                func: AggregateFunc::Dummy,
                                expr: MirScalarExpr::literal_ok(Datum::Dummy, typ.scalar_type),
                                distinct: false,
                            };
                        }
                    }

                    self.action(input, new_columns, gets)
                }
                MirRelationExpr::TopK {
                    input,
                    group_key,
                    order_key,
                    limit,
                    ..
                } => {
                    // Group and order keys and limit must be retained, as they
                    // define which rows are retained.
                    columns.extend(group_key.iter().cloned());
                    columns.extend(order_key.iter().map(|o| o.column));
                    if let Some(limit) = limit {
                        // Strictly speaking not needed because the
                        // `limit` support should be a subset of the
                        // `group_key` support, but we don't want to
                        // take this for granted here.
                        limit.support_into(&mut columns)
                    }
                    self.action(input, columns, gets)
                }
                MirRelationExpr::Negate { input } => self.action(input, columns, gets),
                MirRelationExpr::Threshold { input } => {
                    // Threshold requires all columns, as collapsing any distinct values
                    // has the potential to change how it thresholds counts. This could
                    // be improved with reasoning about distinctness or non-negativity.
                    let arity = input.arity();
                    self.action(input, (0..arity).collect(), gets)
                }
                MirRelationExpr::Union { base, inputs } => {
                    self.action(base, columns.clone(), gets)?;
                    for input in inputs {
                        self.action(input, columns.clone(), gets)?;
                    }
                    Ok(())
                }
                MirRelationExpr::ArrangeBy { input, keys } => {
                    for key_set in keys {
                        for key in key_set {
                            key.support_into(&mut columns);
                        }
                    }
                    self.action(input, columns, gets)
                }
            }
        })
    }
}