mz_transform/analysis.rs
1// Copyright Materialize, Inc. and contributors. All rights reserved.
2//
3// Use of this software is governed by the Business Source License
4// included in the LICENSE file.
5//
6// As of the Change Date specified in that file, in accordance with
7// the Business Source License, use of this software will be governed
8// by the Apache License, Version 2.0.
9
10//! Traits and types for reusable expression analysis
11
12pub mod equivalences;
13pub mod monotonic;
14
15use mz_expr::MirRelationExpr;
16
17pub use arity::Arity;
18pub use cardinality::Cardinality;
19pub use column_names::{ColumnName, ColumnNames};
20pub use common::{Derived, DerivedBuilder, DerivedView};
21pub use explain::annotate_plan;
22pub use non_negative::NonNegative;
23pub use subtree::SubtreeSize;
24pub use types::RelationType;
25pub use unique_keys::UniqueKeys;
26
27/// An analysis that can be applied bottom-up to a `MirRelationExpr`.
28pub trait Analysis: 'static {
29 /// The type of value this analysis associates with an expression.
30 type Value: std::fmt::Debug;
31 /// Announce any dependencies this analysis has on other analyses.
32 ///
33 /// The method should invoke `builder.require::<Foo>()` for each other
34 /// analysis `Foo` this analysis depends upon.
35 fn announce_dependencies(_builder: &mut DerivedBuilder) {}
36 /// The analysis value derived for an expression, given other analysis results.
37 ///
38 /// The other analysis results include the results of this analysis for all children,
39 /// in `results`, and the results of other analyses this analysis has expressed a
40 /// dependence upon, in `depends`, for children and the expression itself.
41 /// The analysis results for `Self` can only be found in `results`, and are not
42 /// available in `depends`.
43 ///
44 /// Implementors of this method must defensively check references into `results`, as
45 /// it may be invoked on `LetRec` bindings that have not yet been populated. It is up
46 /// to the analysis what to do in that case, but conservative behavior is recommended.
47 ///
48 /// The `index` indicates the post-order index for the expression, for use in finding
49 /// the corresponding information in `results` and `depends`.
50 ///
51 /// The returned result will be associated with this expression for this analysis,
52 /// and the analyses will continue.
53 fn derive(
54 &self,
55 expr: &MirRelationExpr,
56 index: usize,
57 results: &[Self::Value],
58 depends: &Derived,
59 ) -> Self::Value;
60
61 /// When available, provide a lattice interface to allow optimistic recursion.
62 ///
63 /// Providing a non-`None` output indicates that the analysis intends re-iteration.
64 fn lattice() -> Option<Box<dyn Lattice<Self::Value>>> {
65 None
66 }
67}
68
69/// Lattice methods for repeated analysis
70pub trait Lattice<T> {
71 /// An element greater than all other elements.
72 fn top(&self) -> T;
73 /// Set `a` to the greatest lower bound of `a` and `b`, and indicate if `a` changed as a result.
74 fn meet_assign(&self, a: &mut T, b: T) -> bool;
75}
76
77/// Types common across multiple analyses
78pub mod common {
79
80 use std::any::{Any, TypeId};
81 use std::collections::BTreeMap;
82
83 use mz_expr::LocalId;
84 use mz_expr::MirRelationExpr;
85 use mz_ore::assert_none;
86 use mz_repr::optimize::OptimizerFeatures;
87
88 use super::Analysis;
89 use super::subtree::SubtreeSize;
90
91 /// Container for analysis state and binding context.
92 #[derive(Default)]
93 #[allow(missing_debug_implementations)]
94 pub struct Derived {
95 /// A record of active analyses and their results, indexed by their type id.
96 analyses: BTreeMap<TypeId, Box<dyn AnalysisBundle>>,
97 /// Analyses ordered where each depends only on strictly prior analyses.
98 order: Vec<TypeId>,
99 /// Map from local identifier to result offset for analysis values.
100 bindings: BTreeMap<LocalId, usize>,
101 }
102
103 impl Derived {
104 /// Return the analysis results derived so far.
105 ///
106 /// # Panics
107 ///
108 /// This method panics if `A` was not installed as a required analysis.
109 pub fn results<A: Analysis>(&self) -> &[A::Value] {
110 let type_id = TypeId::of::<Bundle<A>>();
111 if let Some(bundle) = self.analyses.get(&type_id) {
112 if let Some(bundle) = bundle.as_any().downcast_ref::<Bundle<A>>() {
113 return &bundle.results[..];
114 }
115 }
116 panic!("Analysis {:?} missing", std::any::type_name::<A>());
117 }
118 /// Bindings from local identifiers to result offsets for analysis values.
119 pub fn bindings(&self) -> &BTreeMap<LocalId, usize> {
120 &self.bindings
121 }
122 /// Result offsets for the state of a various number of children of the current expression.
123 ///
124 /// The integers are the zero-offset locations in the `SubtreeSize` analysis. The order of
125 /// the children is descending, from last child to first, because of how the information is
126 /// laid out, and the non-reversibility of the look-ups.
127 ///
128 /// It is an error to call this method with more children than expression has.
129 pub fn children_of_rev<'a>(
130 &'a self,
131 start: usize,
132 count: usize,
133 ) -> impl Iterator<Item = usize> + 'a {
134 let sizes = self.results::<SubtreeSize>();
135 let offset = 1;
136 (0..count).scan(offset, move |offset, _| {
137 let result = start - *offset;
138 *offset += sizes[result];
139 Some(result)
140 })
141 }
142
143 /// Recast the derived data as a view that can be subdivided into views over child state.
144 pub fn as_view<'a>(&'a self) -> DerivedView<'a> {
145 DerivedView {
146 derived: self,
147 lower: 0,
148 upper: self.results::<SubtreeSize>().len(),
149 }
150 }
151 }
152
153 /// The subset of a `Derived` corresponding to an expression and its children.
154 ///
155 /// Specifically, bounds an interval `[lower, upper)` that ends with the state
156 /// of an expression, at `upper-1`, and is preceded by the state of descendents.
157 ///
158 /// This is best thought of as a node in a tree rather
159 #[allow(missing_debug_implementations)]
160 #[derive(Copy, Clone)]
161 pub struct DerivedView<'a> {
162 derived: &'a Derived,
163 lower: usize,
164 upper: usize,
165 }
166
167 impl<'a> DerivedView<'a> {
168 /// The value associated with the expression.
169 pub fn value<A: Analysis>(&self) -> Option<&'a A::Value> {
170 self.results::<A>().last()
171 }
172
173 /// The post-order traversal index for the expression.
174 ///
175 /// This can be used to index into the full set of results, as provided by an
176 /// instance of `Derived`.
177 pub fn index(&self) -> usize {
178 self.upper - 1
179 }
180
181 /// The value bound to an identifier, if it has been derived.
182 ///
183 /// There are several reasons the value could not be derived, and this method
184 /// does not distinguish between them.
185 pub fn bound<A: Analysis>(&self, id: LocalId) -> Option<&'a A::Value> {
186 self.derived
187 .bindings
188 .get(&id)
189 .and_then(|index| self.derived.results::<A>().get(*index))
190 }
191
192 /// The results for expression and its children.
193 ///
194 /// The results for the expression itself will be the last element.
195 ///
196 /// # Panics
197 ///
198 /// This method panics if `A` was not installed as a required analysis.
199 pub fn results<A: Analysis>(&self) -> &'a [A::Value] {
200 &self.derived.results::<A>()[self.lower..self.upper]
201 }
202
203 /// Bindings from local identifiers to result offsets for analysis values.
204 ///
205 /// This method returns all bindings, which may include bindings not in scope for
206 /// the expression and its children; they should be ignored.
207 pub fn bindings(&self) -> &'a BTreeMap<LocalId, usize> {
208 self.derived.bindings()
209 }
210
211 /// Subviews over `self` corresponding to the children of the expression, in reverse order.
212 ///
213 /// These views should disjointly cover the same interval as `self`, except for the last element
214 /// which corresponds to the expression itself.
215 ///
216 /// The number of produced items should exactly match the number of children, which need not
217 /// be provided as an argument. This relies on the well-formedness of the view, which should
218 /// exhaust itself just as it enumerates its last (the first) child view.
219 pub fn children_rev(&self) -> impl Iterator<Item = DerivedView<'a>> + 'a {
220 // This logic is copy/paste from `Derived::children_of_rev` but it was annoying to layer
221 // it over the output of that function, and perhaps clearer to rewrite in any case.
222
223 // Discard the last element (the size of the expression's subtree).
224 // Repeatedly read out the last element, then peel off that many elements.
225 // Each extracted slice corresponds to a child of the current expression.
226 // We should end cleanly with an empty slice, otherwise there is an issue.
227 let sizes = self.results::<SubtreeSize>();
228 let sizes = &sizes[..sizes.len() - 1];
229
230 let offset = self.lower;
231 let derived = self.derived;
232 (0..).scan(sizes, move |sizes, _| {
233 if let Some(size) = sizes.last() {
234 *sizes = &sizes[..sizes.len() - size];
235 Some(Self {
236 derived,
237 lower: offset + sizes.len(),
238 upper: offset + sizes.len() + size,
239 })
240 } else {
241 None
242 }
243 })
244 }
245
246 /// A convenience method for the view over the expressions last child.
247 ///
248 /// This method is appropriate to call on expressions with multiple children,
249 /// and in particular for `Let` and `LetRecv` variants where the body is the
250 /// last child.
251 ///
252 /// It is an error to call this on a view for an expression with no children.
253 pub fn last_child(&self) -> DerivedView<'a> {
254 self.children_rev().next().unwrap()
255 }
256 }
257
258 /// A builder wrapper to accumulate announced dependencies and construct default state.
259 #[allow(missing_debug_implementations)]
260 pub struct DerivedBuilder<'a> {
261 result: Derived,
262 features: &'a OptimizerFeatures,
263 }
264
265 impl<'a> DerivedBuilder<'a> {
266 /// Create a new [`DerivedBuilder`] parameterized by [`OptimizerFeatures`].
267 pub fn new(features: &'a OptimizerFeatures) -> Self {
268 // The default builder should include `SubtreeSize` to facilitate navigation.
269 let mut builder = DerivedBuilder {
270 result: Derived::default(),
271 features,
272 };
273 builder.require(SubtreeSize);
274 builder
275 }
276 }
277
278 impl<'a> DerivedBuilder<'a> {
279 /// Announces a dependence on an analysis `A`.
280 ///
281 /// This ensures that `A` will be performed, and before any analysis that
282 /// invokes this method.
283 pub fn require<A: Analysis>(&mut self, analysis: A) {
284 // The method recursively descends through required analyses, first
285 // installing each in `result.analyses` and second in `result.order`.
286 // The first is an obligation, and serves as an indication that we have
287 // found a cycle in dependencies.
288 let type_id = TypeId::of::<Bundle<A>>();
289 if !self.result.order.contains(&type_id) {
290 // If we have not sequenced `type_id` but have a bundle, it means
291 // we are in the process of fulfilling its requirements: a cycle.
292 if self.result.analyses.contains_key(&type_id) {
293 panic!("Cyclic dependency detected: {}", std::any::type_name::<A>());
294 }
295 // Insert the analysis bundle first, so that we can detect cycles.
296 self.result.analyses.insert(
297 type_id,
298 Box::new(Bundle::<A> {
299 analysis,
300 results: Vec::new(),
301 fuel: 100,
302 allow_optimistic: self.features.enable_letrec_fixpoint_analysis,
303 }),
304 );
305 A::announce_dependencies(self);
306 // All dependencies are successfully sequenced; sequence `type_id`.
307 self.result.order.push(type_id);
308 }
309 }
310 /// Complete the building: perform analyses and return the resulting `Derivation`.
311 pub fn visit(mut self, expr: &MirRelationExpr) -> Derived {
312 // A stack of expressions to process (`Ok`) and let bindings to fill (`Err`).
313 let mut todo = vec![Ok(expr)];
314 // Expressions in reverse post-order: each expression, followed by its children in reverse order.
315 // We will reverse this to get the post order, but must form it in reverse.
316 let mut rev_post_order = Vec::new();
317 while let Some(command) = todo.pop() {
318 match command {
319 // An expression to visit.
320 Ok(expr) => {
321 match expr {
322 MirRelationExpr::Let { id, value, body } => {
323 todo.push(Ok(value));
324 todo.push(Err(*id));
325 todo.push(Ok(body));
326 }
327 MirRelationExpr::LetRec {
328 ids, values, body, ..
329 } => {
330 for (id, value) in ids.iter().zip(values) {
331 todo.push(Ok(value));
332 todo.push(Err(*id));
333 }
334 todo.push(Ok(body));
335 }
336 _ => {
337 todo.extend(expr.children().map(Ok));
338 }
339 }
340 rev_post_order.push(expr);
341 }
342 // A local id to install
343 Err(local_id) => {
344 // Capture the *remaining* work, which we'll need to flip around.
345 let prior = self.result.bindings.insert(local_id, rev_post_order.len());
346 assert_none!(prior, "Shadowing not allowed");
347 }
348 }
349 }
350 // Flip the offsets now that we know a length.
351 for value in self.result.bindings.values_mut() {
352 *value = rev_post_order.len() - *value - 1;
353 }
354 // Visit the pre-order in reverse order: post-order.
355 rev_post_order.reverse();
356
357 // Apply each analysis to `expr` in order.
358 for id in self.result.order.iter() {
359 if let Some(mut bundle) = self.result.analyses.remove(id) {
360 bundle.analyse(&rev_post_order[..], &self.result);
361 self.result.analyses.insert(*id, bundle);
362 }
363 }
364
365 self.result
366 }
367 }
368
369 /// An abstraction for an analysis and associated state.
370 trait AnalysisBundle: Any {
371 /// Populates internal state for all of `exprs`.
372 ///
373 /// Result indicates whether new information was produced for `exprs.last()`.
374 fn analyse(&mut self, exprs: &[&MirRelationExpr], depends: &Derived) -> bool;
375 /// Upcasts `self` to a `&dyn Any`.
376 ///
377 /// NOTE: This is required until <https://github.com/rust-lang/rfcs/issues/2765> is fixed
378 fn as_any(&self) -> &dyn std::any::Any;
379 }
380
381 /// A wrapper for analysis state.
382 struct Bundle<A: Analysis> {
383 /// The algorithm instance used to derive the results.
384 analysis: A,
385 /// A vector of results.
386 results: Vec<A::Value>,
387 /// Counts down with each `LetRec` re-iteration, to avoid unbounded effort.
388 /// Should it reach zero, the analysis should discard its results and restart as if pessimistic.
389 fuel: usize,
390 /// Allow optimistic analysis for `A` (otherwise we always do pesimistic
391 /// analysis, even if a [`crate::analysis::Lattice`] is available for `A`).
392 allow_optimistic: bool,
393 }
394
395 impl<A: Analysis> AnalysisBundle for Bundle<A> {
396 fn analyse(&mut self, exprs: &[&MirRelationExpr], depends: &Derived) -> bool {
397 self.results.clear();
398 // Attempt optimistic analysis, and if that fails go pessimistic.
399 let update = A::lattice()
400 .filter(|_| self.allow_optimistic)
401 .and_then(|lattice| {
402 for _ in exprs.iter() {
403 self.results.push(lattice.top());
404 }
405 self.analyse_optimistic(exprs, 0, exprs.len(), depends, &*lattice)
406 .ok()
407 })
408 .unwrap_or_else(|| {
409 self.results.clear();
410 self.analyse_pessimistic(exprs, depends)
411 });
412 assert_eq!(self.results.len(), exprs.len());
413 update
414 }
415 fn as_any(&self) -> &dyn std::any::Any {
416 self
417 }
418 }
419
420 impl<A: Analysis> Bundle<A> {
421 /// Analysis that starts optimistically but is only correct at a fixed point.
422 ///
423 /// Will fail out to `analyse_pessimistic` if the lattice is missing, or `self.fuel` is exhausted.
424 /// When successful, the result indicates whether new information was produced for `exprs.last()`.
425 fn analyse_optimistic(
426 &mut self,
427 exprs: &[&MirRelationExpr],
428 lower: usize,
429 upper: usize,
430 depends: &Derived,
431 lattice: &dyn crate::analysis::Lattice<A::Value>,
432 ) -> Result<bool, ()> {
433 // Repeatedly re-evaluate the whole tree bottom up, until no changes of fuel spent.
434 let mut changed = true;
435 while changed {
436 changed = false;
437
438 // Bail out if we have done too many `LetRec` passes in this analysis.
439 self.fuel -= 1;
440 if self.fuel == 0 {
441 return Err(());
442 }
443
444 // Track if repetitions may be required, to avoid iteration if they are not.
445 let mut is_recursive = false;
446 // Update each derived value and track if any have changed.
447 for index in lower..upper {
448 let value = self.derive(exprs[index], index, depends);
449 changed = lattice.meet_assign(&mut self.results[index], value) || changed;
450 if let MirRelationExpr::LetRec { .. } = &exprs[index] {
451 is_recursive = true;
452 }
453 }
454
455 // Un-set the potential loop if there was no recursion.
456 if !is_recursive {
457 changed = false;
458 }
459 }
460 Ok(true)
461 }
462
463 /// Analysis that starts conservatively and can be stopped at any point.
464 ///
465 /// Result indicates whether new information was produced for `exprs.last()`.
466 fn analyse_pessimistic(&mut self, exprs: &[&MirRelationExpr], depends: &Derived) -> bool {
467 // TODO: consider making iterative, from some `bottom()` up using `join_assign()`.
468 self.results.clear();
469 for (index, expr) in exprs.iter().enumerate() {
470 self.results.push(self.derive(expr, index, depends));
471 }
472 true
473 }
474
475 #[inline]
476 fn derive(&self, expr: &MirRelationExpr, index: usize, depends: &Derived) -> A::Value {
477 self.analysis
478 .derive(expr, index, &self.results[..], depends)
479 }
480 }
481}
482
483/// Expression subtree sizes
484///
485/// This analysis counts the number of expressions in each subtree, and is most useful
486/// for navigating the results of other analyses that are offset by subtree sizes.
487pub mod subtree {
488
489 use super::{Analysis, Derived};
490 use mz_expr::MirRelationExpr;
491
492 /// Analysis that determines the size in child expressions of relation expressions.
493 #[derive(Debug)]
494 pub struct SubtreeSize;
495
496 impl Analysis for SubtreeSize {
497 type Value = usize;
498
499 fn derive(
500 &self,
501 expr: &MirRelationExpr,
502 index: usize,
503 results: &[Self::Value],
504 _depends: &Derived,
505 ) -> Self::Value {
506 match expr {
507 MirRelationExpr::Constant { .. } | MirRelationExpr::Get { .. } => 1,
508 _ => {
509 let mut offset = 1;
510 for _ in expr.children() {
511 offset += results[index - offset];
512 }
513 offset
514 }
515 }
516 }
517 }
518}
519
520/// Expression arities
521mod arity {
522
523 use super::{Analysis, Derived};
524 use mz_expr::MirRelationExpr;
525
526 /// Analysis that determines the number of columns of relation expressions.
527 #[derive(Debug)]
528 pub struct Arity;
529
530 impl Analysis for Arity {
531 type Value = usize;
532
533 fn derive(
534 &self,
535 expr: &MirRelationExpr,
536 index: usize,
537 results: &[Self::Value],
538 depends: &Derived,
539 ) -> Self::Value {
540 let mut offsets = depends
541 .children_of_rev(index, expr.children().count())
542 .map(|child| results[child])
543 .collect::<Vec<_>>();
544 offsets.reverse();
545 expr.arity_with_input_arities(offsets.into_iter())
546 }
547 }
548}
549
550/// Expression types
551mod types {
552
553 use super::{Analysis, Derived, Lattice};
554 use mz_expr::MirRelationExpr;
555 use mz_repr::ColumnType;
556
557 /// Analysis that determines the type of relation expressions.
558 ///
559 /// The value is `Some` when it discovers column types, and `None` in the case that
560 /// it has discovered no constraining information on the column types. The `None`
561 /// variant should only occur in the course of iteration, and should not be revealed
562 /// as an output of the analysis. One can `unwrap()` the result, and if it errors then
563 /// either the expression is malformed or the analysis has a bug.
564 ///
565 /// The analysis will panic if an expression is not well typed (i.e. if `try_col_with_input_cols`
566 /// returns an error).
567 #[derive(Debug)]
568 pub struct RelationType;
569
570 impl Analysis for RelationType {
571 type Value = Option<Vec<ColumnType>>;
572
573 fn derive(
574 &self,
575 expr: &MirRelationExpr,
576 index: usize,
577 results: &[Self::Value],
578 depends: &Derived,
579 ) -> Self::Value {
580 let offsets = depends
581 .children_of_rev(index, expr.children().count())
582 .map(|child| &results[child])
583 .collect::<Vec<_>>();
584
585 // For most expressions we'll apply `try_col_with_input_cols`, but for `Get` expressions
586 // we'll want to combine what we know (iteratively) with the stated `Get::typ`.
587 match expr {
588 MirRelationExpr::Get {
589 id: mz_expr::Id::Local(i),
590 typ,
591 ..
592 } => {
593 let mut result = typ.column_types.clone();
594 if let Some(o) = depends.bindings().get(i) {
595 if let Some(t) = results.get(*o) {
596 if let Some(rec_typ) = t {
597 // Reconcile nullability statements.
598 // Unclear if we should trust `typ`.
599 assert_eq!(result.len(), rec_typ.len());
600 result.clone_from(rec_typ);
601 for (res, col) in result.iter_mut().zip(typ.column_types.iter()) {
602 if !col.nullable {
603 res.nullable = false;
604 }
605 }
606 } else {
607 // Our `None` information indicates that we are optimistically
608 // assuming the best, including that all columns are non-null.
609 // This should only happen in the first visit to a `Get` expr.
610 // Use `typ`, but flatten nullability.
611 for col in result.iter_mut() {
612 col.nullable = false;
613 }
614 }
615 }
616 }
617 Some(result)
618 }
619 _ => {
620 // Every expression with inputs should have non-`None` inputs at this point.
621 let input_cols = offsets.into_iter().rev().map(|o| {
622 o.as_ref()
623 .expect("RelationType analysis discovered type-less expression")
624 });
625 Some(expr.try_col_with_input_cols(input_cols).unwrap())
626 }
627 }
628 }
629
630 fn lattice() -> Option<Box<dyn Lattice<Self::Value>>> {
631 Some(Box::new(RTLattice))
632 }
633 }
634
635 struct RTLattice;
636
637 impl Lattice<Option<Vec<ColumnType>>> for RTLattice {
638 fn top(&self) -> Option<Vec<ColumnType>> {
639 None
640 }
641 fn meet_assign(&self, a: &mut Option<Vec<ColumnType>>, b: Option<Vec<ColumnType>>) -> bool {
642 match (a, b) {
643 (_, None) => false,
644 (Some(a), Some(b)) => {
645 let mut changed = false;
646 assert_eq!(a.len(), b.len());
647 for (at, bt) in a.iter_mut().zip(b.iter()) {
648 assert_eq!(at.scalar_type, bt.scalar_type);
649 if !at.nullable && bt.nullable {
650 at.nullable = true;
651 changed = true;
652 }
653 }
654 changed
655 }
656 (a, b) => {
657 *a = b;
658 true
659 }
660 }
661 }
662 }
663}
664
665/// Expression unique keys
666mod unique_keys {
667
668 use super::arity::Arity;
669 use super::{Analysis, Derived, DerivedBuilder, Lattice};
670 use mz_expr::MirRelationExpr;
671
672 /// Analysis that determines the unique keys of relation expressions.
673 ///
674 /// The analysis value is a `Vec<Vec<usize>>`, which should be interpreted as a list
675 /// of sets of column identifiers, each set of which has the property that there is at
676 /// most one instance of each assignment of values to those columns.
677 ///
678 /// The sets are minimal, in that any superset of another set is removed from the list.
679 /// Any superset of unique key columns are also unique key columns.
680 #[derive(Debug)]
681 pub struct UniqueKeys;
682
683 impl Analysis for UniqueKeys {
684 type Value = Vec<Vec<usize>>;
685
686 fn announce_dependencies(builder: &mut DerivedBuilder) {
687 builder.require(Arity);
688 }
689
690 fn derive(
691 &self,
692 expr: &MirRelationExpr,
693 index: usize,
694 results: &[Self::Value],
695 depends: &Derived,
696 ) -> Self::Value {
697 let mut offsets = depends
698 .children_of_rev(index, expr.children().count())
699 .collect::<Vec<_>>();
700 offsets.reverse();
701
702 match expr {
703 MirRelationExpr::Get {
704 id: mz_expr::Id::Local(i),
705 typ,
706 ..
707 } => {
708 // We have information from `typ` and from the analysis.
709 // We should "join" them, unioning and reducing the keys.
710 let mut keys = typ.keys.clone();
711 if let Some(o) = depends.bindings().get(i) {
712 if let Some(ks) = results.get(*o) {
713 for k in ks.iter() {
714 antichain_insert(&mut keys, k.clone());
715 }
716 keys.extend(ks.iter().cloned());
717 keys.sort();
718 keys.dedup();
719 }
720 }
721 keys
722 }
723 _ => {
724 let arity = depends.results::<Arity>();
725 expr.keys_with_input_keys(
726 offsets.iter().map(|o| arity[*o]),
727 offsets.iter().map(|o| &results[*o]),
728 )
729 }
730 }
731 }
732
733 fn lattice() -> Option<Box<dyn Lattice<Self::Value>>> {
734 Some(Box::new(UKLattice))
735 }
736 }
737
738 fn antichain_insert(into: &mut Vec<Vec<usize>>, item: Vec<usize>) {
739 // Insert only if there is not a dominating element of `into`.
740 if into.iter().all(|key| !key.iter().all(|k| item.contains(k))) {
741 into.retain(|key| !key.iter().all(|k| item.contains(k)));
742 into.push(item);
743 }
744 }
745
746 /// Lattice for sets of columns that define a unique key.
747 ///
748 /// An element `Vec<Vec<usize>>` describes all sets of columns `Vec<usize>` that are a
749 /// superset of some set of columns in the lattice element.
750 struct UKLattice;
751
752 impl Lattice<Vec<Vec<usize>>> for UKLattice {
753 fn top(&self) -> Vec<Vec<usize>> {
754 vec![vec![]]
755 }
756 fn meet_assign(&self, a: &mut Vec<Vec<usize>>, b: Vec<Vec<usize>>) -> bool {
757 a.sort();
758 a.dedup();
759 let mut c = Vec::new();
760 for cols_a in a.iter_mut() {
761 cols_a.sort();
762 cols_a.dedup();
763 for cols_b in b.iter() {
764 let mut cols_c = cols_a.iter().chain(cols_b).cloned().collect::<Vec<_>>();
765 cols_c.sort();
766 cols_c.dedup();
767 antichain_insert(&mut c, cols_c);
768 }
769 }
770 c.sort();
771 c.dedup();
772 std::mem::swap(a, &mut c);
773 a != &mut c
774 }
775 }
776}
777
778/// Determines if accumulated frequences can be negative.
779///
780/// This analysis assumes that globally identified collection have the property, and it is
781/// incorrect to apply it to expressions that reference external collections that may have
782/// negative accumulations.
783mod non_negative {
784
785 use super::{Analysis, Derived, Lattice};
786 use crate::analysis::common_lattice::BoolLattice;
787 use mz_expr::{Id, MirRelationExpr};
788
789 /// Analysis that determines if all accumulations at all times are non-negative.
790 ///
791 /// The analysis assumes that `Id::Global` references only refer to non-negative collections.
792 #[derive(Debug)]
793 pub struct NonNegative;
794
795 impl Analysis for NonNegative {
796 type Value = bool;
797
798 fn derive(
799 &self,
800 expr: &MirRelationExpr,
801 index: usize,
802 results: &[Self::Value],
803 depends: &Derived,
804 ) -> Self::Value {
805 match expr {
806 MirRelationExpr::Constant { rows, .. } => rows
807 .as_ref()
808 .map(|r| r.iter().all(|(_, diff)| *diff >= mz_repr::Diff::ZERO))
809 .unwrap_or(true),
810 MirRelationExpr::Get { id, .. } => match id {
811 Id::Local(id) => {
812 let index = *depends
813 .bindings()
814 .get(id)
815 .expect("Dependency info not found");
816 *results.get(index).unwrap_or(&false)
817 }
818 Id::Global(_) => true,
819 },
820 // Negate must be false unless input is "non-positive".
821 MirRelationExpr::Negate { .. } => false,
822 // Threshold ensures non-negativity.
823 MirRelationExpr::Threshold { .. } => true,
824 // Reduce errors on negative input.
825 MirRelationExpr::Reduce { .. } => true,
826 MirRelationExpr::Join { .. } => {
827 // If all inputs are non-negative, the join is non-negative.
828 depends
829 .children_of_rev(index, expr.children().count())
830 .all(|off| results[off])
831 }
832 MirRelationExpr::Union { base, inputs } => {
833 // If all inputs are non-negative, the union is non-negative.
834 let all_non_negative = depends
835 .children_of_rev(index, expr.children().count())
836 .all(|off| results[off]);
837
838 if all_non_negative {
839 return true;
840 }
841
842 // We look for the pattern `Union { base, Negate(Subset(base)) }`.
843 // TODO: take some care to ensure that union fusion does not introduce a regression.
844 if inputs.len() == 1 {
845 if let MirRelationExpr::Negate { input } = &inputs[0] {
846 // If `base` is non-negative, and `is_superset_of(base, input)`, return true.
847 // TODO: this is not correct until we have `is_superset_of` validate non-negativity
848 // as it goes, but it matches the current implementation.
849 let mut children = depends.children_of_rev(index, 2);
850 let _negate = children.next().unwrap();
851 let base_id = children.next().unwrap();
852 debug_assert_eq!(children.next(), None);
853 if results[base_id] && is_superset_of(&*base, &*input) {
854 return true;
855 }
856 }
857 }
858
859 false
860 }
861 _ => results[index - 1],
862 }
863 }
864
865 fn lattice() -> Option<Box<dyn Lattice<Self::Value>>> {
866 Some(Box::new(BoolLattice))
867 }
868 }
869
870 /// Returns true only if `rhs.negate().union(lhs)` contains only non-negative multiplicities
871 /// once consolidated.
872 ///
873 /// Informally, this happens when `rhs` is a multiset subset of `lhs`, meaning the multiplicity
874 /// of any record in `rhs` is at most the multiplicity of the same record in `lhs`.
875 ///
876 /// This method recursively descends each of `lhs` and `rhs` and performs a great many equality
877 /// tests, which has the potential to be quadratic. We should consider restricting its attention
878 /// to `Get` identifiers, under the premise that equal AST nodes would necessarily be identified
879 /// by common subexpression elimination. This requires care around recursively bound identifiers.
880 ///
881 /// These rules are .. somewhat arbitrary, and likely reflect observed opportunities. For example,
882 /// while we do relate `distinct(filter(A)) <= distinct(A)`, we do not relate `distinct(A) <= A`.
883 /// Further thoughts about the class of optimizations, and whether there should be more or fewer,
884 /// can be found here: <https://github.com/MaterializeInc/database-issues/issues/4044>.
885 fn is_superset_of(mut lhs: &MirRelationExpr, mut rhs: &MirRelationExpr) -> bool {
886 // This implementation is iterative.
887 // Before converting this implementation to recursive (e.g. to improve its accuracy)
888 // make sure to use the `CheckedRecursion` struct to avoid blowing the stack.
889 while lhs != rhs {
890 match rhs {
891 MirRelationExpr::Filter { input, .. } => rhs = &**input,
892 MirRelationExpr::TopK { input, .. } => rhs = &**input,
893 // Descend in both sides if the current roots are
894 // projections with the same `outputs` vector.
895 MirRelationExpr::Project {
896 input: rhs_input,
897 outputs: rhs_outputs,
898 } => match lhs {
899 MirRelationExpr::Project {
900 input: lhs_input,
901 outputs: lhs_outputs,
902 } if lhs_outputs == rhs_outputs => {
903 rhs = &**rhs_input;
904 lhs = &**lhs_input;
905 }
906 _ => return false,
907 },
908 // Descend in both sides if the current roots are reduces with empty aggregates
909 // on the same set of keys (that is, a distinct operation on those keys).
910 MirRelationExpr::Reduce {
911 input: rhs_input,
912 group_key: rhs_group_key,
913 aggregates: rhs_aggregates,
914 monotonic: _,
915 expected_group_size: _,
916 } if rhs_aggregates.is_empty() => match lhs {
917 MirRelationExpr::Reduce {
918 input: lhs_input,
919 group_key: lhs_group_key,
920 aggregates: lhs_aggregates,
921 monotonic: _,
922 expected_group_size: _,
923 } if lhs_aggregates.is_empty() && lhs_group_key == rhs_group_key => {
924 rhs = &**rhs_input;
925 lhs = &**lhs_input;
926 }
927 _ => return false,
928 },
929 _ => {
930 // TODO: Imagine more complex reasoning here!
931 return false;
932 }
933 }
934 }
935 true
936 }
937}
938
939mod column_names {
940 use std::ops::Range;
941
942 use super::Analysis;
943 use mz_expr::{AggregateFunc, Id, MirRelationExpr, MirScalarExpr};
944 use mz_repr::GlobalId;
945 use mz_repr::explain::ExprHumanizer;
946
947 /// An abstract type denoting an inferred column name.
948 #[derive(Debug, Clone)]
949 pub enum ColumnName {
950 /// A column with name inferred to be equal to a GlobalId schema column.
951 Global(GlobalId, usize),
952 /// An anonymous expression named after the top-level function name.
953 Aggregate(AggregateFunc, Box<ColumnName>),
954 /// An column with an unknown name.
955 Unknown,
956 }
957
958 impl ColumnName {
959 /// Return `true` iff this the variant is not unknown.
960 pub fn is_known(&self) -> bool {
961 matches!(self, Self::Global(..) | Self::Aggregate(..))
962 }
963
964 /// Humanize the column to a [`String`], returns an empty [`String`] for
965 /// unknown columns.
966 pub fn humanize(&self, humanizer: &dyn ExprHumanizer) -> String {
967 match self {
968 Self::Global(id, c) => humanizer.humanize_column(*id, *c).unwrap_or_default(),
969 Self::Aggregate(func, expr) => {
970 let func = func.name();
971 let expr = expr.humanize(humanizer);
972 if expr.is_empty() {
973 String::from(func)
974 } else {
975 format!("{func}_{expr}")
976 }
977 }
978 Self::Unknown => String::new(),
979 }
980 }
981 }
982
983 /// Compute the column types of each subtree of a [MirRelationExpr] from the
984 /// bottom-up.
985 #[derive(Debug)]
986 pub struct ColumnNames;
987
988 impl ColumnNames {
989 /// fallback schema consisting of ordinal column names: #0, #1, ...
990 fn anonymous(range: Range<usize>) -> impl Iterator<Item = ColumnName> {
991 range.map(|_| ColumnName::Unknown)
992 }
993
994 /// fallback schema consisting of ordinal column names: #0, #1, ...
995 fn extend_with_scalars(column_names: &mut Vec<ColumnName>, scalars: &Vec<MirScalarExpr>) {
996 for scalar in scalars {
997 column_names.push(match scalar {
998 MirScalarExpr::Column(c) => column_names[*c].clone(),
999 _ => ColumnName::Unknown,
1000 });
1001 }
1002 }
1003 }
1004
1005 impl Analysis for ColumnNames {
1006 type Value = Vec<ColumnName>;
1007
1008 fn derive(
1009 &self,
1010 expr: &MirRelationExpr,
1011 index: usize,
1012 results: &[Self::Value],
1013 depends: &crate::analysis::Derived,
1014 ) -> Self::Value {
1015 use MirRelationExpr::*;
1016
1017 match expr {
1018 Constant { rows: _, typ } => {
1019 // Fallback to an anonymous schema for constants.
1020 ColumnNames::anonymous(0..typ.arity()).collect()
1021 }
1022 Get {
1023 id: Id::Global(id),
1024 typ,
1025 access_strategy: _,
1026 } => {
1027 // Emit ColumnName::Global instances for each column in the
1028 // `Get` type. Those can be resolved to real names later when an
1029 // ExpressionHumanizer is available.
1030 (0..typ.columns().len())
1031 .map(|c| ColumnName::Global(*id, c))
1032 .collect()
1033 }
1034 Get {
1035 id: Id::Local(id),
1036 typ,
1037 access_strategy: _,
1038 } => {
1039 let index_child = *depends.bindings().get(id).expect("id in scope");
1040 if index_child < results.len() {
1041 results[index_child].clone()
1042 } else {
1043 // Possible because we infer LetRec bindings in order. This
1044 // can be improved by introducing a fixpoint loop in the
1045 // Env<A>::schedule_tasks LetRec handling block.
1046 ColumnNames::anonymous(0..typ.arity()).collect()
1047 }
1048 }
1049 Let {
1050 id: _,
1051 value: _,
1052 body: _,
1053 } => {
1054 // Return the column names of the `body`.
1055 results[index - 1].clone()
1056 }
1057 LetRec {
1058 ids: _,
1059 values: _,
1060 limits: _,
1061 body: _,
1062 } => {
1063 // Return the column names of the `body`.
1064 results[index - 1].clone()
1065 }
1066 Project { input: _, outputs } => {
1067 // Permute the column names of the input.
1068 let input_column_names = &results[index - 1];
1069 let mut column_names = vec![];
1070 for col in outputs {
1071 column_names.push(input_column_names[*col].clone());
1072 }
1073 column_names
1074 }
1075 Map { input: _, scalars } => {
1076 // Extend the column names of the input with anonymous columns.
1077 let mut column_names = results[index - 1].clone();
1078 Self::extend_with_scalars(&mut column_names, scalars);
1079 column_names
1080 }
1081 FlatMap {
1082 input: _,
1083 func,
1084 exprs: _,
1085 } => {
1086 // Extend the column names of the input with anonymous columns.
1087 let mut column_names = results[index - 1].clone();
1088 let func_output_start = column_names.len();
1089 let func_output_end = column_names.len() + func.output_arity();
1090 column_names.extend(Self::anonymous(func_output_start..func_output_end));
1091 column_names
1092 }
1093 Filter {
1094 input: _,
1095 predicates: _,
1096 } => {
1097 // Return the column names of the `input`.
1098 results[index - 1].clone()
1099 }
1100 Join {
1101 inputs: _,
1102 equivalences: _,
1103 implementation: _,
1104 } => {
1105 let mut input_results = depends
1106 .children_of_rev(index, expr.children().count())
1107 .map(|child| &results[child])
1108 .collect::<Vec<_>>();
1109 input_results.reverse();
1110
1111 let mut column_names = vec![];
1112 for input_column_names in input_results {
1113 column_names.extend(input_column_names.iter().cloned());
1114 }
1115 column_names
1116 }
1117 Reduce {
1118 input: _,
1119 group_key,
1120 aggregates,
1121 monotonic: _,
1122 expected_group_size: _,
1123 } => {
1124 // We clone and extend the input vector and then remove the part
1125 // associated with the input at the end.
1126 let mut column_names = results[index - 1].clone();
1127 let input_arity = column_names.len();
1128
1129 // Infer the group key part.
1130 Self::extend_with_scalars(&mut column_names, group_key);
1131 // Infer the aggregates part.
1132 for aggregate in aggregates.iter() {
1133 // The inferred name will consist of (1) the aggregate
1134 // function name and (2) the aggregate expression (iff
1135 // it is a simple column reference).
1136 let func = aggregate.func.clone();
1137 let expr = match aggregate.expr.as_column() {
1138 Some(c) => column_names.get(c).unwrap_or(&ColumnName::Unknown).clone(),
1139 None => ColumnName::Unknown,
1140 };
1141 column_names.push(ColumnName::Aggregate(func, Box::new(expr)));
1142 }
1143 // Remove the prefix associated with the input
1144 column_names.drain(0..input_arity);
1145
1146 column_names
1147 }
1148 TopK {
1149 input: _,
1150 group_key: _,
1151 order_key: _,
1152 limit: _,
1153 offset: _,
1154 monotonic: _,
1155 expected_group_size: _,
1156 } => {
1157 // Return the column names of the `input`.
1158 results[index - 1].clone()
1159 }
1160 Negate { input: _ } => {
1161 // Return the column names of the `input`.
1162 results[index - 1].clone()
1163 }
1164 Threshold { input: _ } => {
1165 // Return the column names of the `input`.
1166 results[index - 1].clone()
1167 }
1168 Union { base: _, inputs: _ } => {
1169 // Use the first non-empty column across all inputs.
1170 let mut column_names = vec![];
1171
1172 let mut inputs_results = depends
1173 .children_of_rev(index, expr.children().count())
1174 .map(|child| &results[child])
1175 .collect::<Vec<_>>();
1176
1177 let base_results = inputs_results.pop().unwrap();
1178 inputs_results.reverse();
1179
1180 for (i, mut column_name) in base_results.iter().cloned().enumerate() {
1181 for input_results in inputs_results.iter() {
1182 if !column_name.is_known() && input_results[i].is_known() {
1183 column_name = input_results[i].clone();
1184 break;
1185 }
1186 }
1187 column_names.push(column_name);
1188 }
1189
1190 column_names
1191 }
1192 ArrangeBy { input: _, keys: _ } => {
1193 // Return the column names of the `input`.
1194 results[index - 1].clone()
1195 }
1196 }
1197 }
1198 }
1199}
1200
1201mod explain {
1202 //! Derived Analysis framework and definitions.
1203
1204 use std::collections::BTreeMap;
1205
1206 use mz_expr::MirRelationExpr;
1207 use mz_expr::explain::{ExplainContext, HumanizedExplain, HumanizerMode};
1208 use mz_ore::stack::RecursionLimitError;
1209 use mz_repr::explain::{Analyses, AnnotatedPlan};
1210
1211 use crate::analysis::equivalences::{Equivalences, HumanizedEquivalenceClasses};
1212
1213 // Analyses should have shortened paths when exported.
1214 use super::DerivedBuilder;
1215
1216 impl<'c> From<&ExplainContext<'c>> for DerivedBuilder<'c> {
1217 fn from(context: &ExplainContext<'c>) -> DerivedBuilder<'c> {
1218 let mut builder = DerivedBuilder::new(context.features);
1219 if context.config.subtree_size {
1220 builder.require(super::SubtreeSize);
1221 }
1222 if context.config.non_negative {
1223 builder.require(super::NonNegative);
1224 }
1225 if context.config.types {
1226 builder.require(super::RelationType);
1227 }
1228 if context.config.arity {
1229 builder.require(super::Arity);
1230 }
1231 if context.config.keys {
1232 builder.require(super::UniqueKeys);
1233 }
1234 if context.config.cardinality {
1235 builder.require(super::Cardinality::with_stats(
1236 context.cardinality_stats.clone(),
1237 ));
1238 }
1239 if context.config.column_names || context.config.humanized_exprs {
1240 builder.require(super::ColumnNames);
1241 }
1242 if context.config.equivalences {
1243 builder.require(Equivalences);
1244 }
1245 builder
1246 }
1247 }
1248
1249 /// Produce an [`AnnotatedPlan`] wrapping the given [`MirRelationExpr`] along
1250 /// with [`Analyses`] derived from the given context configuration.
1251 pub fn annotate_plan<'a>(
1252 plan: &'a MirRelationExpr,
1253 context: &'a ExplainContext,
1254 ) -> Result<AnnotatedPlan<'a, MirRelationExpr>, RecursionLimitError> {
1255 let mut annotations = BTreeMap::<&MirRelationExpr, Analyses>::default();
1256 let config = context.config;
1257
1258 // We want to annotate the plan with analyses in the following cases:
1259 // 1. An Analysis was explicitly requested in the ExplainConfig.
1260 // 2. Humanized expressions were requested in the ExplainConfig (in which
1261 // case we need to derive the ColumnNames Analysis).
1262 if config.requires_analyses() || config.humanized_exprs {
1263 // get the annotation keys
1264 let subtree_refs = plan.post_order_vec();
1265 // get the annotation values
1266 let builder = DerivedBuilder::from(context);
1267 let derived = builder.visit(plan);
1268
1269 if config.subtree_size {
1270 for (expr, subtree_size) in std::iter::zip(
1271 subtree_refs.iter(),
1272 derived.results::<super::SubtreeSize>().into_iter(),
1273 ) {
1274 let analyses = annotations.entry(expr).or_default();
1275 analyses.subtree_size = Some(*subtree_size);
1276 }
1277 }
1278 if config.non_negative {
1279 for (expr, non_negative) in std::iter::zip(
1280 subtree_refs.iter(),
1281 derived.results::<super::NonNegative>().into_iter(),
1282 ) {
1283 let analyses = annotations.entry(expr).or_default();
1284 analyses.non_negative = Some(*non_negative);
1285 }
1286 }
1287
1288 if config.arity {
1289 for (expr, arity) in std::iter::zip(
1290 subtree_refs.iter(),
1291 derived.results::<super::Arity>().into_iter(),
1292 ) {
1293 let analyses = annotations.entry(expr).or_default();
1294 analyses.arity = Some(*arity);
1295 }
1296 }
1297
1298 if config.types {
1299 for (expr, types) in std::iter::zip(
1300 subtree_refs.iter(),
1301 derived.results::<super::RelationType>().into_iter(),
1302 ) {
1303 let analyses = annotations.entry(expr).or_default();
1304 analyses.types = Some(types.clone());
1305 }
1306 }
1307
1308 if config.keys {
1309 for (expr, keys) in std::iter::zip(
1310 subtree_refs.iter(),
1311 derived.results::<super::UniqueKeys>().into_iter(),
1312 ) {
1313 let analyses = annotations.entry(expr).or_default();
1314 analyses.keys = Some(keys.clone());
1315 }
1316 }
1317
1318 if config.cardinality {
1319 for (expr, card) in std::iter::zip(
1320 subtree_refs.iter(),
1321 derived.results::<super::Cardinality>().into_iter(),
1322 ) {
1323 let analyses = annotations.entry(expr).or_default();
1324 analyses.cardinality = Some(card.to_string());
1325 }
1326 }
1327
1328 if config.column_names || config.humanized_exprs {
1329 for (expr, column_names) in std::iter::zip(
1330 subtree_refs.iter(),
1331 derived.results::<super::ColumnNames>().into_iter(),
1332 ) {
1333 let analyses = annotations.entry(expr).or_default();
1334 let value = column_names
1335 .iter()
1336 .map(|column_name| column_name.humanize(context.humanizer))
1337 .collect();
1338 analyses.column_names = Some(value);
1339 }
1340 }
1341
1342 if config.equivalences {
1343 for (expr, equivs) in std::iter::zip(
1344 subtree_refs.iter(),
1345 derived.results::<Equivalences>().into_iter(),
1346 ) {
1347 let analyses = annotations.entry(expr).or_default();
1348 analyses.equivalences = Some(match equivs.as_ref() {
1349 Some(equivs) => HumanizedEquivalenceClasses {
1350 equivalence_classes: equivs,
1351 cols: analyses.column_names.as_ref(),
1352 mode: HumanizedExplain::new(config.redacted),
1353 }
1354 .to_string(),
1355 None => "<empty collection>".to_string(),
1356 });
1357 }
1358 }
1359 }
1360
1361 Ok(AnnotatedPlan { plan, annotations })
1362 }
1363}
1364
1365/// Definition and helper structs for the [`Cardinality`] Analysis.
1366mod cardinality {
1367 use std::collections::{BTreeMap, BTreeSet};
1368
1369 use mz_expr::{
1370 BinaryFunc, Id, JoinImplementation, MirRelationExpr, MirScalarExpr, TableFunc, UnaryFunc,
1371 VariadicFunc,
1372 };
1373 use mz_ore::cast::{CastFrom, CastLossy, TryCastFrom};
1374 use mz_repr::GlobalId;
1375
1376 use ordered_float::OrderedFloat;
1377
1378 use super::{Analysis, Arity, SubtreeSize, UniqueKeys};
1379
1380 /// Compute the estimated cardinality of each subtree of a [MirRelationExpr] from the bottom up.
1381 #[allow(missing_debug_implementations)]
1382 pub struct Cardinality {
1383 /// Cardinalities for globally named entities
1384 pub stats: BTreeMap<GlobalId, usize>,
1385 }
1386
1387 impl Cardinality {
1388 /// A cardinality estimator with provided statistics for the given global identifiers
1389 pub fn with_stats(stats: BTreeMap<GlobalId, usize>) -> Self {
1390 Cardinality { stats }
1391 }
1392 }
1393
1394 impl Default for Cardinality {
1395 fn default() -> Self {
1396 Cardinality {
1397 stats: BTreeMap::new(),
1398 }
1399 }
1400 }
1401
1402 /// Cardinality estimates
1403 #[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
1404 pub enum CardinalityEstimate {
1405 Unknown,
1406 Estimate(OrderedFloat<f64>),
1407 }
1408
1409 impl CardinalityEstimate {
1410 pub fn max(lhs: CardinalityEstimate, rhs: CardinalityEstimate) -> CardinalityEstimate {
1411 use CardinalityEstimate::*;
1412 match (lhs, rhs) {
1413 (Estimate(lhs), Estimate(rhs)) => Estimate(std::cmp::max(lhs, rhs)),
1414 _ => Unknown,
1415 }
1416 }
1417
1418 pub fn rounded(&self) -> Option<usize> {
1419 match self {
1420 CardinalityEstimate::Estimate(OrderedFloat(f)) => {
1421 let rounded = f.ceil();
1422 let flattened = usize::cast_from(
1423 u64::try_cast_from(rounded)
1424 .expect("positive and representable cardinality estimate"),
1425 );
1426
1427 Some(flattened)
1428 }
1429 CardinalityEstimate::Unknown => None,
1430 }
1431 }
1432 }
1433
1434 impl std::ops::Add for CardinalityEstimate {
1435 type Output = CardinalityEstimate;
1436
1437 fn add(self, rhs: Self) -> Self::Output {
1438 use CardinalityEstimate::*;
1439 match (self, rhs) {
1440 (Estimate(lhs), Estimate(rhs)) => Estimate(lhs + rhs),
1441 _ => Unknown,
1442 }
1443 }
1444 }
1445
1446 impl std::ops::Sub for CardinalityEstimate {
1447 type Output = CardinalityEstimate;
1448
1449 fn sub(self, rhs: Self) -> Self::Output {
1450 use CardinalityEstimate::*;
1451 match (self, rhs) {
1452 (Estimate(lhs), Estimate(rhs)) => Estimate(lhs - rhs),
1453 _ => Unknown,
1454 }
1455 }
1456 }
1457
1458 impl std::ops::Sub<CardinalityEstimate> for f64 {
1459 type Output = CardinalityEstimate;
1460
1461 fn sub(self, rhs: CardinalityEstimate) -> Self::Output {
1462 use CardinalityEstimate::*;
1463 if let Estimate(OrderedFloat(rhs)) = rhs {
1464 Estimate(OrderedFloat(self - rhs))
1465 } else {
1466 Unknown
1467 }
1468 }
1469 }
1470
1471 impl std::ops::Mul for CardinalityEstimate {
1472 type Output = CardinalityEstimate;
1473
1474 fn mul(self, rhs: Self) -> Self::Output {
1475 use CardinalityEstimate::*;
1476 match (self, rhs) {
1477 (Estimate(lhs), Estimate(rhs)) => Estimate(lhs * rhs),
1478 _ => Unknown,
1479 }
1480 }
1481 }
1482
1483 impl std::ops::Mul<f64> for CardinalityEstimate {
1484 type Output = CardinalityEstimate;
1485
1486 fn mul(self, rhs: f64) -> Self::Output {
1487 if let CardinalityEstimate::Estimate(OrderedFloat(lhs)) = self {
1488 CardinalityEstimate::Estimate(OrderedFloat(lhs * rhs))
1489 } else {
1490 CardinalityEstimate::Unknown
1491 }
1492 }
1493 }
1494
1495 impl std::ops::Div for CardinalityEstimate {
1496 type Output = CardinalityEstimate;
1497
1498 fn div(self, rhs: Self) -> Self::Output {
1499 use CardinalityEstimate::*;
1500 match (self, rhs) {
1501 (Estimate(lhs), Estimate(rhs)) => Estimate(lhs / rhs),
1502 _ => Unknown,
1503 }
1504 }
1505 }
1506
1507 impl std::ops::Div<f64> for CardinalityEstimate {
1508 type Output = CardinalityEstimate;
1509
1510 fn div(self, rhs: f64) -> Self::Output {
1511 use CardinalityEstimate::*;
1512 if let Estimate(lhs) = self {
1513 Estimate(lhs / OrderedFloat(rhs))
1514 } else {
1515 Unknown
1516 }
1517 }
1518 }
1519
1520 impl std::iter::Sum for CardinalityEstimate {
1521 fn sum<I: Iterator<Item = Self>>(iter: I) -> Self {
1522 iter.fold(CardinalityEstimate::from(0.0), |acc, elt| acc + elt)
1523 }
1524 }
1525
1526 impl std::iter::Product for CardinalityEstimate {
1527 fn product<I: Iterator<Item = Self>>(iter: I) -> Self {
1528 iter.fold(CardinalityEstimate::from(1.0), |acc, elt| acc * elt)
1529 }
1530 }
1531
1532 impl From<usize> for CardinalityEstimate {
1533 fn from(value: usize) -> Self {
1534 Self::Estimate(OrderedFloat(f64::cast_lossy(value)))
1535 }
1536 }
1537
1538 impl From<f64> for CardinalityEstimate {
1539 fn from(value: f64) -> Self {
1540 Self::Estimate(OrderedFloat(value))
1541 }
1542 }
1543
1544 /// The default selectivity for predicates we know nothing about.
1545 ///
1546 /// But see also expr/src/scalar.rs for `FilterCharacteristics::worst_case_scaling_factor()` for a more nuanced take.
1547 pub const WORST_CASE_SELECTIVITY: OrderedFloat<f64> = OrderedFloat(0.1);
1548
1549 // This section defines how we estimate cardinality for each syntactic construct.
1550 //
1551 // We split it up into functions to make it all a bit more tractable to work with.
1552 impl Cardinality {
1553 fn flat_map(&self, tf: &TableFunc, input: CardinalityEstimate) -> CardinalityEstimate {
1554 match tf {
1555 TableFunc::Wrap { types, width } => {
1556 input * (f64::cast_lossy(types.len()) / f64::cast_lossy(*width))
1557 }
1558 _ => {
1559 // TODO(mgree) what explosion factor should we make up?
1560 input * CardinalityEstimate::from(4.0)
1561 }
1562 }
1563 }
1564
1565 fn predicate(
1566 &self,
1567 predicate_expr: &MirScalarExpr,
1568 unique_columns: &BTreeSet<usize>,
1569 ) -> OrderedFloat<f64> {
1570 let index_selectivity = |expr: &MirScalarExpr| -> Option<OrderedFloat<f64>> {
1571 match expr {
1572 MirScalarExpr::Column(col) => {
1573 if unique_columns.contains(col) {
1574 // TODO(mgree): when we have index cardinality statistics, they should go here when `expr` is a `MirScalarExpr::Column` that's in `unique_columns`
1575 None
1576 } else {
1577 None
1578 }
1579 }
1580 _ => None,
1581 }
1582 };
1583
1584 match predicate_expr {
1585 MirScalarExpr::Column(_)
1586 | MirScalarExpr::Literal(_, _)
1587 | MirScalarExpr::CallUnmaterializable(_) => OrderedFloat(1.0),
1588 MirScalarExpr::CallUnary { func, expr } => match func {
1589 UnaryFunc::Not(_) => OrderedFloat(1.0) - self.predicate(expr, unique_columns),
1590 UnaryFunc::IsTrue(_) | UnaryFunc::IsFalse(_) => OrderedFloat(0.5),
1591 UnaryFunc::IsNull(_) => {
1592 if let Some(icard) = index_selectivity(expr) {
1593 icard
1594 } else {
1595 WORST_CASE_SELECTIVITY
1596 }
1597 }
1598 _ => WORST_CASE_SELECTIVITY,
1599 },
1600 MirScalarExpr::CallBinary { func, expr1, expr2 } => {
1601 match func {
1602 BinaryFunc::Eq => {
1603 match (index_selectivity(expr1), index_selectivity(expr2)) {
1604 (Some(isel1), Some(isel2)) => std::cmp::max(isel1, isel2),
1605 (Some(isel), None) | (None, Some(isel)) => isel,
1606 (None, None) => WORST_CASE_SELECTIVITY,
1607 }
1608 }
1609 // 1.0 - the Eq case
1610 BinaryFunc::NotEq => {
1611 match (index_selectivity(expr1), index_selectivity(expr2)) {
1612 (Some(isel1), Some(isel2)) => {
1613 OrderedFloat(1.0) - std::cmp::max(isel1, isel2)
1614 }
1615 (Some(isel), None) | (None, Some(isel)) => OrderedFloat(1.0) - isel,
1616 (None, None) => OrderedFloat(1.0) - WORST_CASE_SELECTIVITY,
1617 }
1618 }
1619 BinaryFunc::Lt | BinaryFunc::Lte | BinaryFunc::Gt | BinaryFunc::Gte => {
1620 // TODO(mgree) if we have high/low key values and one of the columns is an index, we can do better
1621 OrderedFloat(0.33)
1622 }
1623 _ => OrderedFloat(1.0), // TOOD(mgree): are there other interesting cases?
1624 }
1625 }
1626 MirScalarExpr::CallVariadic { func, exprs } => match func {
1627 VariadicFunc::And => exprs
1628 .iter()
1629 .map(|expr| self.predicate(expr, unique_columns))
1630 .product(),
1631 VariadicFunc::Or => {
1632 // TODO(mgree): BETWEEN will get compiled down to an AND of appropriate bounds---we could try to detect it and be clever
1633
1634 // F(expr1 OR expr2) = F(expr1) + F(expr2) - F(expr1) * F(expr2), but generalized
1635 let mut exprs = exprs.into_iter();
1636
1637 let mut expr1;
1638
1639 if let Some(first) = exprs.next() {
1640 expr1 = self.predicate(first, unique_columns);
1641 } else {
1642 return OrderedFloat(1.0);
1643 }
1644
1645 for expr2 in exprs {
1646 let expr2 = self.predicate(expr2, unique_columns);
1647 expr1 = expr1 + expr2 - expr1 * expr2;
1648 }
1649 expr1
1650 }
1651 _ => OrderedFloat(1.0),
1652 },
1653 MirScalarExpr::If { cond: _, then, els } => std::cmp::max(
1654 self.predicate(then, unique_columns),
1655 self.predicate(els, unique_columns),
1656 ),
1657 }
1658 }
1659
1660 fn filter(
1661 &self,
1662 predicates: &Vec<MirScalarExpr>,
1663 keys: &Vec<Vec<usize>>,
1664 input: CardinalityEstimate,
1665 ) -> CardinalityEstimate {
1666 // TODO(mgree): should we try to do something for indices built on multiple columns?
1667 let mut unique_columns = BTreeSet::new();
1668 for key in keys {
1669 if key.len() == 1 {
1670 unique_columns.insert(key[0]);
1671 }
1672 }
1673
1674 let mut estimate = input;
1675 for expr in predicates {
1676 let selectivity = self.predicate(expr, &unique_columns);
1677 debug_assert!(
1678 OrderedFloat(0.0) <= selectivity && selectivity <= OrderedFloat(1.0),
1679 "predicate selectivity {selectivity} should be in the range [0,1]"
1680 );
1681 estimate = estimate * selectivity.0;
1682 }
1683
1684 estimate
1685 }
1686
1687 fn join(
1688 &self,
1689 equivalences: &Vec<Vec<MirScalarExpr>>,
1690 _implementation: &JoinImplementation,
1691 unique_columns: BTreeMap<usize, usize>,
1692 mut inputs: Vec<CardinalityEstimate>,
1693 ) -> CardinalityEstimate {
1694 if inputs.is_empty() {
1695 return CardinalityEstimate::from(0.0);
1696 }
1697
1698 for equiv in equivalences {
1699 // those sources which have a unique key
1700 let mut unique_sources = BTreeSet::new();
1701 let mut all_unique = true;
1702
1703 for expr in equiv {
1704 if let MirScalarExpr::Column(col) = expr {
1705 if let Some(idx) = unique_columns.get(col) {
1706 unique_sources.insert(*idx);
1707 } else {
1708 all_unique = false;
1709 }
1710 } else {
1711 all_unique = false;
1712 }
1713 }
1714
1715 // no unique columns in this equivalence
1716 if unique_sources.is_empty() {
1717 continue;
1718 }
1719
1720 // ALL unique columns in this equivalence
1721 if all_unique {
1722 // these inputs have unique keys for _all_ of the equivalence, so they're a bound on how many rows we'll get from those sources
1723 // we'll find the leftmost such input and use it to hold the minimum; the other sources we set to 1.0 (so they have no effect)
1724 let mut sources = unique_sources.iter();
1725
1726 let lhs_idx = *sources.next().unwrap();
1727 let mut lhs =
1728 std::mem::replace(&mut inputs[lhs_idx], CardinalityEstimate::from(1.0));
1729 for &rhs_idx in sources {
1730 let rhs =
1731 std::mem::replace(&mut inputs[rhs_idx], CardinalityEstimate::from(1.0));
1732 lhs = CardinalityEstimate::min(lhs, rhs);
1733 }
1734
1735 inputs[lhs_idx] = lhs;
1736
1737 // best option! go look at the next equivalence
1738 continue;
1739 }
1740
1741 // some unique columns in this equivalence
1742 for idx in unique_sources {
1743 // when joining R and S on R.x = S.x, if R.x is unique and S.x is not, we're bounded above by the cardinality of S
1744 inputs[idx] = CardinalityEstimate::from(1.0);
1745 }
1746 }
1747
1748 let mut product = CardinalityEstimate::from(1.0);
1749 for input in inputs {
1750 product = product * input;
1751 }
1752 product
1753 }
1754
1755 fn reduce(
1756 &self,
1757 group_key: &Vec<MirScalarExpr>,
1758 expected_group_size: &Option<u64>,
1759 input: CardinalityEstimate,
1760 ) -> CardinalityEstimate {
1761 // TODO(mgree): if no `group_key` is present, we can do way better
1762
1763 if let Some(group_size) = expected_group_size {
1764 input / f64::cast_lossy(*group_size)
1765 } else if group_key.is_empty() {
1766 CardinalityEstimate::from(1.0)
1767 } else {
1768 // in the worst case, every row is its own group
1769 input
1770 }
1771 }
1772
1773 fn topk(
1774 &self,
1775 group_key: &Vec<usize>,
1776 limit: &Option<MirScalarExpr>,
1777 expected_group_size: &Option<u64>,
1778 input: CardinalityEstimate,
1779 ) -> CardinalityEstimate {
1780 // TODO: support simple arithmetic expressions
1781 let k = limit
1782 .as_ref()
1783 .and_then(|l| l.as_literal_int64())
1784 .map_or(1, |l| std::cmp::max(0, l));
1785
1786 if let Some(group_size) = expected_group_size {
1787 input * (f64::cast_lossy(k) / f64::cast_lossy(*group_size))
1788 } else if group_key.is_empty() {
1789 CardinalityEstimate::from(f64::cast_lossy(k))
1790 } else {
1791 // in the worst case, every row is its own group
1792 input.clone()
1793 }
1794 }
1795
1796 fn threshold(&self, input: CardinalityEstimate) -> CardinalityEstimate {
1797 // worst case scaling factor is 1
1798 input.clone()
1799 }
1800 }
1801
1802 impl Analysis for Cardinality {
1803 type Value = CardinalityEstimate;
1804
1805 fn announce_dependencies(builder: &mut crate::analysis::DerivedBuilder) {
1806 builder.require(crate::analysis::Arity);
1807 builder.require(crate::analysis::UniqueKeys);
1808 }
1809
1810 fn derive(
1811 &self,
1812 expr: &MirRelationExpr,
1813 index: usize,
1814 results: &[Self::Value],
1815 depends: &crate::analysis::Derived,
1816 ) -> Self::Value {
1817 use MirRelationExpr::*;
1818
1819 let sizes = depends.as_view().results::<SubtreeSize>();
1820 let arity = depends.as_view().results::<Arity>();
1821 let keys = depends.as_view().results::<UniqueKeys>();
1822
1823 match expr {
1824 Constant { rows, .. } => {
1825 CardinalityEstimate::from(rows.as_ref().map_or_else(|_| 0, |v| v.len()))
1826 }
1827 Get { id, .. } => match id {
1828 Id::Local(id) => depends
1829 .bindings()
1830 .get(id)
1831 .and_then(|id| results.get(*id))
1832 .copied()
1833 .unwrap_or(CardinalityEstimate::Unknown),
1834 Id::Global(id) => self
1835 .stats
1836 .get(id)
1837 .copied()
1838 .map(CardinalityEstimate::from)
1839 .unwrap_or(CardinalityEstimate::Unknown),
1840 },
1841 Let { .. } | Project { .. } | Map { .. } | ArrangeBy { .. } | Negate { .. } => {
1842 results[index - 1].clone()
1843 }
1844 LetRec { .. } =>
1845 // TODO(mgree): implement a recurrence-based approach (or at least identify common idioms, e.g. transitive closure)
1846 {
1847 CardinalityEstimate::Unknown
1848 }
1849 Union { base: _, inputs: _ } => depends
1850 .children_of_rev(index, expr.children().count())
1851 .map(|off| results[off].clone())
1852 .sum(),
1853 FlatMap { func, .. } => {
1854 let input = results[index - 1];
1855 self.flat_map(func, input)
1856 }
1857 Filter { predicates, .. } => {
1858 let input = results[index - 1];
1859 let keys = depends.results::<UniqueKeys>();
1860 let keys = &keys[index - 1];
1861 self.filter(predicates, keys, input)
1862 }
1863 Join {
1864 equivalences,
1865 implementation,
1866 inputs,
1867 ..
1868 } => {
1869 let mut input_results = Vec::with_capacity(inputs.len());
1870
1871 // maps a column to the index in `inputs` that it belongs to
1872 let mut unique_columns = BTreeMap::new();
1873 let mut key_offset = 0;
1874
1875 let mut offset = 1;
1876 for idx in 0..inputs.len() {
1877 let input = results[index - offset];
1878 input_results.push(input);
1879
1880 let arity = arity[index - offset];
1881 let keys = &keys[index - offset];
1882 for key in keys {
1883 if key.len() == 1 {
1884 unique_columns.insert(key_offset + key[0], idx);
1885 }
1886 }
1887 key_offset += arity;
1888
1889 offset += &sizes[index - offset];
1890 }
1891
1892 self.join(equivalences, implementation, unique_columns, input_results)
1893 }
1894 Reduce {
1895 group_key,
1896 expected_group_size,
1897 ..
1898 } => {
1899 let input = results[index - 1];
1900 self.reduce(group_key, expected_group_size, input)
1901 }
1902 TopK {
1903 group_key,
1904 limit,
1905 expected_group_size,
1906 ..
1907 } => {
1908 let input = results[index - 1];
1909 self.topk(group_key, limit, expected_group_size, input)
1910 }
1911 Threshold { .. } => {
1912 let input = results[index - 1];
1913 self.threshold(input)
1914 }
1915 }
1916 }
1917 }
1918
1919 impl std::fmt::Display for CardinalityEstimate {
1920 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1921 match self {
1922 CardinalityEstimate::Estimate(OrderedFloat(estimate)) => write!(f, "{estimate}"),
1923 CardinalityEstimate::Unknown => write!(f, "<UNKNOWN>"),
1924 }
1925 }
1926 }
1927}
1928
1929mod common_lattice {
1930 use crate::analysis::Lattice;
1931
1932 pub struct BoolLattice;
1933
1934 impl Lattice<bool> for BoolLattice {
1935 // `true` > `false`.
1936 fn top(&self) -> bool {
1937 true
1938 }
1939 // `false` is the greatest lower bound. `into` changes if it's true and `item` is false.
1940 fn meet_assign(&self, into: &mut bool, item: bool) -> bool {
1941 let changed = *into && !item;
1942 *into = *into && item;
1943 changed
1944 }
1945 }
1946}