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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938
// 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.
//! Lowering [`DataflowDescription`]s from MIR ([`MirRelationExpr`]) to LIR ([`Plan`]).
use std::collections::BTreeMap;
use mz_expr::JoinImplementation::{DeltaQuery, Differential, IndexedFilter, Unimplemented};
use mz_expr::{
permutation_for_arrangement, Id, JoinInputMapper, MapFilterProject, MirRelationExpr,
MirScalarExpr, OptimizedMirRelationExpr,
};
use mz_ore::{assert_none, soft_assert_eq_or_log, soft_panic_or_log};
use mz_repr::optimize::OptimizerFeatures;
use mz_repr::GlobalId;
use timely::progress::Timestamp;
use crate::dataflows::{BuildDesc, DataflowDescription, IndexImport};
use crate::plan::join::{DeltaJoinPlan, JoinPlan, LinearJoinPlan};
use crate::plan::reduce::{KeyValPlan, ReducePlan};
use crate::plan::threshold::ThresholdPlan;
use crate::plan::top_k::TopKPlan;
use crate::plan::{AvailableCollections, GetPlan, LirId, Plan};
pub(super) struct Context {
/// Known bindings to (possibly arranged) collections.
arrangements: BTreeMap<Id, AvailableCollections>,
/// Tracks the next available `LirId`.
next_lir_id: LirId,
/// Information to print along with error messages.
debug_info: LirDebugInfo,
/// Whether to enable fusion of MFPs in reductions.
enable_reduce_mfp_fusion: bool,
}
impl Context {
pub fn new(debug_name: String, features: &OptimizerFeatures) -> Self {
Self {
arrangements: Default::default(),
next_lir_id: 0,
debug_info: LirDebugInfo {
debug_name,
id: GlobalId::Transient(0),
},
enable_reduce_mfp_fusion: features.enable_reduce_mfp_fusion,
}
}
fn allocate_lir_id(&mut self) -> LirId {
let id = self.next_lir_id;
self.next_lir_id += 1;
id
}
pub fn lower<T: Timestamp>(
mut self,
desc: DataflowDescription<OptimizedMirRelationExpr>,
) -> Result<DataflowDescription<Plan<T>>, String> {
// Sources might provide arranged forms of their data, in the future.
// Indexes provide arranged forms of their data.
for IndexImport {
desc: index_desc,
typ,
..
} in desc.index_imports.values()
{
let key = index_desc.key.clone();
// TODO[btv] - We should be told the permutation by
// `index_desc`, and it should have been generated
// at the same point the thinning logic was.
//
// We should for sure do that soon, but it requires
// a bit of a refactor, so for now we just
// _assume_ that they were both generated by `permutation_for_arrangement`,
// and recover it here.
let (permutation, thinning) = permutation_for_arrangement(&key, typ.arity());
let index_keys = self
.arrangements
.entry(Id::Global(index_desc.on_id))
.or_insert_with(AvailableCollections::default);
index_keys.arranged.push((key, permutation, thinning));
index_keys.types = Some(typ.column_types.clone());
}
for id in desc.source_imports.keys() {
self.arrangements
.entry(Id::Global(*id))
.or_insert_with(AvailableCollections::new_raw);
}
// Build each object in order, registering the arrangements it forms.
let mut objects_to_build = Vec::with_capacity(desc.objects_to_build.len());
for build in desc.objects_to_build {
self.debug_info.id = build.id;
let (plan, keys) = self.lower_mir_expr(&build.plan)?;
self.arrangements.insert(Id::Global(build.id), keys);
objects_to_build.push(BuildDesc { id: build.id, plan });
}
Ok(DataflowDescription {
source_imports: desc.source_imports,
index_imports: desc.index_imports,
objects_to_build,
index_exports: desc.index_exports,
sink_exports: desc.sink_exports,
as_of: desc.as_of,
until: desc.until,
initial_storage_as_of: desc.initial_storage_as_of,
refresh_schedule: desc.refresh_schedule,
debug_name: desc.debug_name,
})
}
/// This method converts a MirRelationExpr into a plan that can be directly rendered.
///
/// The rough structure is that we repeatedly extract map/filter/project operators
/// from each expression we see, bundle them up as a `MapFilterProject` object, and
/// then produce a plan for the combination of that with the next operator.
///
/// The method accesses `self.arrangements`, which it will locally add to and remove from for
/// `Let` bindings (by the end of the call it should contain the same bindings as when it
/// started).
///
/// The result of the method is both a `Plan`, but also a list of arrangements that
/// are certain to be produced, which can be relied on by the next steps in the plan.
/// Each of the arrangement keys is associated with an MFP that must be applied if that
/// arrangement is used, to back out the permutation associated with that arrangement.
///
/// An empty list of arrangement keys indicates that only a `Collection` stream can
/// be assumed to exist.
fn lower_mir_expr<T: Timestamp>(
&mut self,
expr: &MirRelationExpr,
) -> Result<(Plan<T>, AvailableCollections), String> {
// This function is recursive and can overflow its stack, so grow it if
// needed. The growth here is unbounded. Our general solution for this problem
// is to use [`ore::stack::RecursionGuard`] to additionally limit the stack
// depth. That however requires upstream error handling. This function is
// currently called by the Coordinator after calls to `catalog_transact`,
// and thus are not allowed to fail. Until that allows errors, we choose
// to allow the unbounded growth here. We are though somewhat protected by
// higher levels enforcing their own limits on stack depth (in the parser,
// transformer/desugarer, and planner).
mz_ore::stack::maybe_grow(|| self.lower_mir_expr_stack_safe(expr))
}
fn lower_mir_expr_stack_safe<T>(
&mut self,
expr: &MirRelationExpr,
) -> Result<(Plan<T>, AvailableCollections), String>
where
T: Timestamp,
{
// Extract a maximally large MapFilterProject from `expr`.
// We will then try and push this in to the resulting expression.
//
// Importantly, `mfp` may contain temporal operators and not be a "safe" MFP.
// While we would eventually like all plan stages to be able to absorb such
// general operators, not all of them can.
let (mut mfp, expr) = MapFilterProject::extract_from_expression(expr);
// We attempt to plan what we have remaining, in the context of `mfp`.
// We may not be able to do this, and must wrap some operators with a `Mfp` stage.
let (mut plan, mut keys) = match expr {
// These operators should have been extracted from the expression.
MirRelationExpr::Map { .. } => {
panic!("This operator should have been extracted");
}
MirRelationExpr::Filter { .. } => {
panic!("This operator should have been extracted");
}
MirRelationExpr::Project { .. } => {
panic!("This operator should have been extracted");
}
// These operators may not have been extracted, and need to result in a `Plan`.
MirRelationExpr::Constant { rows, typ: _ } => {
let plan = Plan::Constant {
rows: rows.clone().map(|rows| {
rows.into_iter()
.map(|(row, diff)| (row, T::minimum(), diff))
.collect()
}),
lir_id: self.allocate_lir_id(),
};
// The plan, not arranged in any way.
(plan, AvailableCollections::new_raw())
}
MirRelationExpr::Get { id, typ: _, .. } => {
// This stage can absorb arbitrary MFP operators.
let mut mfp = mfp.take();
// If `mfp` is the identity, we can surface all imported arrangements.
// Otherwise, we apply `mfp` and promise no arrangements.
let mut in_keys = self
.arrangements
.get(id)
.cloned()
.unwrap_or_else(AvailableCollections::new_raw);
// Seek out an arrangement key that might be constrained to a literal.
// Note: this code has very little use nowadays, as its job was mostly taken over
// by `LiteralConstraints` (see in the below longer comment).
let key_val = in_keys
.arranged
.iter()
.filter_map(|key| {
mfp.literal_constraints(&key.0)
.map(|val| (key.clone(), val))
})
.max_by_key(|(key, _val)| key.0.len());
// Determine the plan of action for the `Get` stage.
let plan = if let Some(((key, permutation, thinning), val)) = &key_val {
// This code path used to handle looking up literals from indexes, but it's
// mostly deprecated, as this is nowadays performed by the `LiteralConstraints`
// MIR transform instead. However, it's still called in a couple of tricky
// special cases:
// - `LiteralConstraints` handles only Gets of global ids, so this code still
// gets to handle Filters on top of Gets of local ids.
// - Lowering does a `MapFilterProject::extract_from_expression`, while
// `LiteralConstraints` does
// `MapFilterProject::extract_non_errors_from_expr_mut`.
// - It might happen that new literal constraint optimization opportunities
// appear somewhere near the end of the MIR optimizer after
// `LiteralConstraints` has already run.
// (Also note that a similar literal constraint handling machinery is also
// present when handling the leftover MFP after this big match.)
mfp.permute(permutation.clone(), thinning.len() + key.len());
in_keys.arranged = vec![(key.clone(), permutation.clone(), thinning.clone())];
GetPlan::Arrangement(key.clone(), Some(val.clone()), mfp)
} else if !mfp.is_identity() {
// We need to ensure a collection exists, which means we must form it.
if let Some((key, permutation, thinning)) =
in_keys.arbitrary_arrangement().cloned()
{
mfp.permute(permutation.clone(), thinning.len() + key.len());
in_keys.arranged = vec![(key.clone(), permutation, thinning)];
GetPlan::Arrangement(key, None, mfp)
} else {
GetPlan::Collection(mfp)
}
} else {
// By default, just pass input arrangements through.
GetPlan::PassArrangements
};
let out_keys = if let GetPlan::PassArrangements = plan {
in_keys.clone()
} else {
AvailableCollections::new_raw()
};
// Return the plan, and any keys if an identity `mfp`.
(
Plan::Get {
id: id.clone(),
keys: in_keys,
plan,
lir_id: self.allocate_lir_id(),
},
out_keys,
)
}
MirRelationExpr::Let { id, value, body } => {
// It would be unfortunate to have a non-trivial `mfp` here, as we hope
// that they would be pushed down. I am not sure if we should take the
// initiative to push down the `mfp` ourselves.
// Plan the value using only the initial arrangements, but
// introduce any resulting arrangements bound to `id`.
let (value, v_keys) = self.lower_mir_expr(value)?;
let pre_existing = self.arrangements.insert(Id::Local(*id), v_keys);
assert_none!(pre_existing);
// Plan the body using initial and `value` arrangements,
// and then remove reference to the value arrangements.
let (body, b_keys) = self.lower_mir_expr(body)?;
self.arrangements.remove(&Id::Local(*id));
// Return the plan, and any `body` arrangements.
(
Plan::Let {
id: id.clone(),
value: Box::new(value),
body: Box::new(body),
lir_id: self.allocate_lir_id(),
},
b_keys,
)
}
MirRelationExpr::LetRec {
ids,
values,
limits,
body,
} => {
assert_eq!(ids.len(), values.len());
assert_eq!(ids.len(), limits.len());
// Plan the values using only the available arrangements, but
// introduce any resulting arrangements bound to each `id`.
// Arrangements made available cannot be used by prior bindings,
// as we cannot circulate an arrangement through a `Variable` yet.
let mut lir_values = Vec::with_capacity(values.len());
for (id, value) in ids.iter().zip(values) {
let (mut lir_value, mut v_keys) = self.lower_mir_expr(value)?;
// If `v_keys` does not contain an unarranged collection, we must form it.
if !v_keys.raw {
// Choose an "arbitrary" arrangement; TODO: prefer a specific one.
let (input_key, permutation, thinning) =
v_keys.arbitrary_arrangement().unwrap();
let mut input_mfp = MapFilterProject::new(value.arity());
input_mfp.permute(permutation.clone(), thinning.len() + input_key.len());
let input_key = Some(input_key.clone());
let forms = AvailableCollections::new_raw();
// We just want to insert an `ArrangeBy` to form an unarranged collection,
// but there is a complication: We shouldn't break the invariant (created by
// `NormalizeLets`, and relied upon by the rendering) that there isn't
// anything between two `LetRec`s. So if `lir_value` is itself a `LetRec`,
// then we insert the `ArrangeBy` on the `body` of the inner `LetRec`,
// instead of on top of the inner `LetRec`.
lir_value = match lir_value {
Plan::LetRec {
ids,
values,
limits,
body,
lir_id,
} => Plan::LetRec {
ids,
values,
limits,
body: Box::new(Plan::ArrangeBy {
input: body,
forms,
input_key,
input_mfp,
lir_id: self.allocate_lir_id(),
}),
lir_id,
},
lir_value => Plan::ArrangeBy {
input: Box::new(lir_value),
forms,
input_key,
input_mfp,
lir_id: self.allocate_lir_id(),
},
};
v_keys.raw = true;
}
let pre_existing = self.arrangements.insert(Id::Local(*id), v_keys);
assert_none!(pre_existing);
lir_values.push(lir_value);
}
// As we exit the iterative scope, we must leave all arrangements behind,
// as they reference a timestamp coordinate that must be stripped off.
for id in ids.iter() {
self.arrangements
.insert(Id::Local(*id), AvailableCollections::new_raw());
}
// Plan the body using initial and `value` arrangements,
// and then remove reference to the value arrangements.
let (body, b_keys) = self.lower_mir_expr(body)?;
for id in ids.iter() {
self.arrangements.remove(&Id::Local(*id));
}
// Return the plan, and any `body` arrangements.
(
Plan::LetRec {
ids: ids.clone(),
values: lir_values,
limits: limits.clone(),
body: Box::new(body),
lir_id: self.allocate_lir_id(),
},
b_keys,
)
}
MirRelationExpr::FlatMap { input, func, exprs } => {
let (input, keys) = self.lower_mir_expr(input)?;
// This stage can absorb arbitrary MFP instances.
let mfp = mfp.take();
let mut exprs = exprs.clone();
let input_key = if let Some((k, permutation, _)) = keys.arbitrary_arrangement() {
// We don't permute the MFP here, because it runs _after_ the table function,
// whose output is in a fixed order.
//
// We _do_, however, need to permute the `expr`s that provide input to the
// `func`.
for expr in &mut exprs {
expr.permute_map(permutation);
}
Some(k.clone())
} else {
None
};
// Return the plan, and no arrangements.
(
Plan::FlatMap {
input: Box::new(input),
func: func.clone(),
exprs: exprs.clone(),
mfp_after: mfp,
input_key,
lir_id: self.allocate_lir_id(),
},
AvailableCollections::new_raw(),
)
}
MirRelationExpr::Join {
inputs,
equivalences,
implementation,
} => {
let input_mapper = JoinInputMapper::new(inputs);
// Plan each of the join inputs independently.
// The `plans` get surfaced upwards, and the `input_keys` should
// be used as part of join planning / to validate the existing
// plans / to aid in indexed seeding of update streams.
let mut plans = Vec::new();
let mut input_keys = Vec::new();
let mut input_arities = Vec::new();
for input in inputs.iter() {
let (plan, keys) = self.lower_mir_expr(input)?;
input_arities.push(input.arity());
plans.push(plan);
input_keys.push(keys);
}
// Extract temporal predicates as joins cannot currently absorb them.
let (plan, missing) = match implementation {
IndexedFilter(_coll_id, _idx_id, key, _val) => {
// Start with the constant input. (This used to be important before #14059
// was fixed.)
let start: usize = 1;
let order = vec![(0usize, key.clone(), None)];
// All columns of the constant input will be part of the arrangement key.
let source_arrangement = (
(0..key.len())
.map(MirScalarExpr::Column)
.collect::<Vec<_>>(),
(0..key.len()).map(|i| (i, i)).collect::<BTreeMap<_, _>>(),
Vec::<usize>::new(),
);
let (ljp, missing) = LinearJoinPlan::create_from(
start,
Some(&source_arrangement),
equivalences,
&order,
input_mapper,
&mut mfp,
&input_keys,
);
(JoinPlan::Linear(ljp), missing)
}
Differential((start, start_arr, _start_characteristic), order) => {
let source_arrangement = start_arr.as_ref().and_then(|key| {
input_keys[*start]
.arranged
.iter()
.find(|(k, _, _)| k == key)
.clone()
});
let (ljp, missing) = LinearJoinPlan::create_from(
*start,
source_arrangement,
equivalences,
order,
input_mapper,
&mut mfp,
&input_keys,
);
(JoinPlan::Linear(ljp), missing)
}
DeltaQuery(orders) => {
let (djp, missing) = DeltaJoinPlan::create_from(
equivalences,
orders,
input_mapper,
&mut mfp,
&input_keys,
);
(JoinPlan::Delta(djp), missing)
}
// Other plans are errors, and should be reported as such.
Unimplemented => return Err("unimplemented join".to_string()),
};
// The renderer will expect certain arrangements to exist; if any of those are not available, the join planning functions above should have returned them in
// `missing`. We thus need to plan them here so they'll exist.
let is_delta = matches!(plan, JoinPlan::Delta(_));
for (((input_plan, input_keys), missing), arity) in plans
.iter_mut()
.zip(input_keys.iter())
.zip(missing.into_iter())
.zip(input_arities.iter().cloned())
{
if missing != Default::default() {
if is_delta {
// join_implementation.rs produced a sub-optimal plan here;
// we shouldn't plan delta joins at all if not all of the required
// arrangements are available. Soft panic in CI and log an error in
// production to increase the chances that we will catch all situations
// that violate this constraint.
soft_panic_or_log!("Arrangements depended on by delta join alarmingly absent: {:?}
Dataflow info: {}
This is not expected to cause incorrect results, but could indicate a performance issue in Materialize.", missing, self.debug_info);
} else {
soft_panic_or_log!("Arrangements depended on by a non-delta join are absent: {:?}
Dataflow info: {}
This is not expected to cause incorrect results, but could indicate a performance issue in Materialize.", missing, self.debug_info);
// Nowadays MIR transforms take care to insert MIR ArrangeBys for each
// Join input. (Earlier, they were missing in the following cases:
// - They were const-folded away for constant inputs. This is not
// happening since
// https://github.com/MaterializeInc/materialize/pull/16351
// - They were not being inserted for the constant input of
// `IndexedFilter`s. This was fixed in
// https://github.com/MaterializeInc/materialize/pull/20920
// - They were not being inserted for the first input of Differential
// joins. This was fixed in
// https://github.com/MaterializeInc/materialize/pull/16099)
}
let raw_plan = std::mem::replace(
input_plan,
Plan::Constant {
rows: Ok(Vec::new()),
lir_id: self.allocate_lir_id(),
},
);
*input_plan = self.arrange_by(raw_plan, missing, input_keys, arity);
}
}
// Return the plan, and no arrangements.
(
Plan::Join {
inputs: plans,
plan,
lir_id: self.allocate_lir_id(),
},
AvailableCollections::new_raw(),
)
}
MirRelationExpr::Reduce {
input,
group_key,
aggregates,
monotonic,
expected_group_size,
} => {
let input_arity = input.arity();
let (input, keys) = self.lower_mir_expr(input)?;
let (input_key, permutation_and_new_arity) = if let Some((
input_key,
permutation,
thinning,
)) = keys.arbitrary_arrangement()
{
(
Some(input_key.clone()),
Some((permutation.clone(), thinning.len() + input_key.len())),
)
} else {
(None, None)
};
let key_val_plan = KeyValPlan::new(
input_arity,
group_key,
aggregates,
permutation_and_new_arity,
);
let reduce_plan =
ReducePlan::create_from(aggregates.clone(), *monotonic, *expected_group_size);
// Return the plan, and the keys it produces.
let mfp_after;
let output_arity;
if self.enable_reduce_mfp_fusion {
(mfp_after, mfp, output_arity) =
reduce_plan.extract_mfp_after(mfp, group_key.len());
} else {
(mfp_after, output_arity) = (
MapFilterProject::new(mfp.input_arity),
group_key.len() + aggregates.len(),
);
soft_assert_eq_or_log!(
mfp.input_arity,
output_arity,
"Output arity of reduce must match input arity for MFP on top of it"
);
}
let output_keys = reduce_plan.keys(group_key.len(), output_arity);
(
Plan::Reduce {
input: Box::new(input),
key_val_plan,
plan: reduce_plan,
input_key,
mfp_after,
lir_id: self.allocate_lir_id(),
},
output_keys,
)
}
MirRelationExpr::TopK {
input,
group_key,
order_key,
limit,
offset,
monotonic,
expected_group_size,
} => {
let arity = input.arity();
let (input, keys) = self.lower_mir_expr(input)?;
let top_k_plan = TopKPlan::create_from(
group_key.clone(),
order_key.clone(),
*offset,
limit.clone(),
arity,
*monotonic,
*expected_group_size,
);
// We don't have an MFP here -- install an operator to permute the
// input, if necessary.
let input = if !keys.raw {
self.arrange_by(input, AvailableCollections::new_raw(), &keys, arity)
} else {
input
};
// Return the plan, and no arrangements.
(
Plan::TopK {
input: Box::new(input),
top_k_plan,
lir_id: self.allocate_lir_id(),
},
AvailableCollections::new_raw(),
)
}
MirRelationExpr::Negate { input } => {
let arity = input.arity();
let (input, keys) = self.lower_mir_expr(input)?;
// We don't have an MFP here -- install an operator to permute the
// input, if necessary.
let input = if !keys.raw {
self.arrange_by(input, AvailableCollections::new_raw(), &keys, arity)
} else {
input
};
// Return the plan, and no arrangements.
(
Plan::Negate {
input: Box::new(input),
lir_id: self.allocate_lir_id(),
},
AvailableCollections::new_raw(),
)
}
MirRelationExpr::Threshold { input } => {
let arity = input.arity();
let (plan, keys) = self.lower_mir_expr(input)?;
// We don't have an MFP here -- install an operator to permute the
// input, if necessary.
let plan = if !keys.raw {
self.arrange_by(plan, AvailableCollections::new_raw(), &keys, arity)
} else {
plan
};
let (threshold_plan, required_arrangement) = ThresholdPlan::create_from(arity);
let mut types = keys.types.clone();
let plan = if !keys
.arranged
.iter()
.any(|(key, _, _)| key == &required_arrangement.0)
{
types = Some(input.typ().column_types);
self.arrange_by(
plan,
AvailableCollections::new_arranged(
vec![required_arrangement],
types.clone(),
),
&keys,
arity,
)
} else {
plan
};
let output_keys = threshold_plan.keys(types);
// Return the plan, and any produced keys.
(
Plan::Threshold {
input: Box::new(plan),
threshold_plan,
lir_id: self.allocate_lir_id(),
},
output_keys,
)
}
MirRelationExpr::Union { base, inputs } => {
let arity = base.arity();
let mut plans_keys = Vec::with_capacity(1 + inputs.len());
let (plan, keys) = self.lower_mir_expr(base)?;
plans_keys.push((plan, keys));
for input in inputs.iter() {
let (plan, keys) = self.lower_mir_expr(input)?;
plans_keys.push((plan, keys));
}
let plans = plans_keys
.into_iter()
.map(|(plan, keys)| {
// We don't have an MFP here -- install an operator to permute the
// input, if necessary.
if !keys.raw {
self.arrange_by(plan, AvailableCollections::new_raw(), &keys, arity)
} else {
plan
}
})
.collect();
// Return the plan and no arrangements.
let plan = Plan::Union {
inputs: plans,
consolidate_output: false,
lir_id: self.allocate_lir_id(),
};
(plan, AvailableCollections::new_raw())
}
MirRelationExpr::ArrangeBy { input, keys } => {
let arity = input.arity();
let types = Some(input.typ().column_types);
let (input, mut input_keys) = self.lower_mir_expr(input)?;
input_keys.types = types;
// Determine keys that are not present in `input_keys`.
let new_keys = keys
.iter()
.filter(|k1| !input_keys.arranged.iter().any(|(k2, _, _)| k1 == &k2))
.cloned()
.collect::<Vec<_>>();
if new_keys.is_empty() {
(input, input_keys)
} else {
let new_keys = new_keys.iter().cloned().map(|k| {
let (permutation, thinning) = permutation_for_arrangement(&k, arity);
(k, permutation, thinning)
});
let (input_key, input_mfp) = if let Some((input_key, permutation, thinning)) =
input_keys.arbitrary_arrangement()
{
let mut mfp = MapFilterProject::new(arity);
mfp.permute(permutation.clone(), thinning.len() + input_key.len());
(Some(input_key.clone()), mfp)
} else {
(None, MapFilterProject::new(arity))
};
input_keys.arranged.extend(new_keys);
input_keys.arranged.sort_by(|k1, k2| k1.0.cmp(&k2.0));
// Return the plan and extended keys.
(
Plan::ArrangeBy {
input: Box::new(input),
forms: input_keys.clone(),
input_key,
input_mfp,
lir_id: self.allocate_lir_id(),
},
input_keys,
)
}
}
};
// If the plan stage did not absorb all linear operators, introduce a new stage to implement them.
if !mfp.is_identity() {
// Seek out an arrangement key that might be constrained to a literal.
// TODO: Improve key selection heuristic.
let key_val = keys
.arranged
.iter()
.filter_map(|(key, permutation, thinning)| {
let mut mfp = mfp.clone();
mfp.permute(permutation.clone(), thinning.len() + key.len());
mfp.literal_constraints(key)
.map(|val| (key.clone(), permutation, thinning, val))
})
.max_by_key(|(key, _, _, _)| key.len());
// Input key selection strategy:
// (1) If we can read a key at a particular value, do so
// (2) Otherwise, if there is a key that causes the MFP to be the identity, and
// therefore allows us to avoid discarding the arrangement, use that.
// (3) Otherwise, if there is _some_ key, use that,
// (4) Otherwise just read the raw collection.
let input_key_val = if let Some((key, permutation, thinning, val)) = key_val {
mfp.permute(permutation.clone(), thinning.len() + key.len());
Some((key, Some(val)))
} else if let Some((key, permutation, thinning)) =
keys.arranged.iter().find(|(key, permutation, thinning)| {
let mut mfp = mfp.clone();
mfp.permute(permutation.clone(), thinning.len() + key.len());
mfp.is_identity()
})
{
mfp.permute(permutation.clone(), thinning.len() + key.len());
Some((key.clone(), None))
} else if let Some((key, permutation, thinning)) = keys.arbitrary_arrangement() {
mfp.permute(permutation.clone(), thinning.len() + key.len());
Some((key.clone(), None))
} else {
None
};
if mfp.is_identity() {
// We have discovered a key
// whose permutation causes the MFP to actually
// be the identity! We can keep it around,
// but without its permutation this time,
// and with a trivial thinning of the right length.
let (key, val) = input_key_val.unwrap();
let (_old_key, old_permutation, old_thinning) = keys
.arranged
.iter_mut()
.find(|(key2, _, _)| key2 == &key)
.unwrap();
*old_permutation = (0..mfp.input_arity).map(|i| (i, i)).collect();
let old_thinned_arity = old_thinning.len();
*old_thinning = (0..old_thinned_arity).collect();
// Get rid of all other forms, as this is now the only one known to be valid.
// TODO[btv] we can probably save the other arrangements too, if we adjust their permutations.
// This is not hard to do, but leaving it for a quick follow-up to avoid making the present diff too unwieldy.
keys.arranged.retain(|(key2, _, _)| key2 == &key);
keys.raw = false;
// Creating a Plan::Mfp node is now logically unnecessary, but we
// should do so anyway when `val` is populated, so that
// the `key_val` optimization gets applied.
if val.is_some() {
plan = Plan::Mfp {
input: Box::new(plan),
mfp,
input_key_val: Some((key, val)),
lir_id: self.allocate_lir_id(),
}
}
} else {
plan = Plan::Mfp {
input: Box::new(plan),
mfp,
input_key_val,
lir_id: self.allocate_lir_id(),
};
keys = AvailableCollections::new_raw();
}
}
Ok((plan, keys))
}
/// Replace the plan with another one
/// that has the collection in some additional forms.
pub fn arrange_by<T>(
&mut self,
plan: Plan<T>,
collections: AvailableCollections,
old_collections: &AvailableCollections,
arity: usize,
) -> Plan<T> {
if let Plan::ArrangeBy {
input,
mut forms,
input_key,
input_mfp,
lir_id,
} = plan
{
forms.raw |= collections.raw;
forms.arranged.extend(collections.arranged);
forms.arranged.sort_by(|k1, k2| k1.0.cmp(&k2.0));
forms.arranged.dedup_by(|k1, k2| k1.0 == k2.0);
if forms.types.is_none() {
forms.types = collections.types;
} else {
assert!(collections.types.is_none() || collections.types == forms.types);
}
Plan::ArrangeBy {
input,
forms,
input_key,
input_mfp,
lir_id,
}
} else {
let (input_key, input_mfp) = if let Some((input_key, permutation, thinning)) =
old_collections.arbitrary_arrangement()
{
let mut mfp = MapFilterProject::new(arity);
mfp.permute(permutation.clone(), thinning.len() + input_key.len());
(Some(input_key.clone()), mfp)
} else {
(None, MapFilterProject::new(arity))
};
Plan::ArrangeBy {
input: Box::new(plan),
forms: collections,
input_key,
input_mfp,
lir_id: self.allocate_lir_id(),
}
}
}
}
/// Various bits of state to print along with error messages during LIR planning,
/// to aid debugging.
#[derive(Clone, Debug)]
pub struct LirDebugInfo {
debug_name: String,
id: GlobalId,
}
impl std::fmt::Display for LirDebugInfo {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
write!(f, "Debug name: {}; id: {}", self.debug_name, self.id)
}
}