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
// 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.
//! Whole-dataflow optimization
//!
//! A dataflow may contain multiple views, each of which may only be
//! optimized locally. However, information like demand and predicate
//! pushdown can be applied across views once we understand the context
//! in which the views will be executed.
use dataflow_types::{DataflowDesc, LinearOperator};
use expr::{GlobalId, Id, LocalId, MirRelationExpr, MirScalarExpr};
use ore::id_gen::IdGen;
use std::collections::{BTreeSet, HashMap, HashSet};
use crate::{monotonic::MonotonicFlag, Optimizer, TransformError};
/// Optimizes the implementation of each dataflow.
///
/// Inlines views, performs a full optimization pass including physical
/// planning using the supplied indexes, propagates filtering and projection
/// information to dataflow sources and lifts monotonicity information.
pub fn optimize_dataflow(
dataflow: &mut DataflowDesc,
indexes: &HashMap<GlobalId, Vec<(GlobalId, Vec<MirScalarExpr>)>>,
) -> Result<(), TransformError> {
// Inline views that are used in only one other view.
inline_views(dataflow)?;
// Logical optimization pass after view inlining
optimize_dataflow_relations(dataflow, indexes, &Optimizer::logical_optimizer())?;
optimize_dataflow_filters(dataflow)?;
// TODO: when the linear operator contract ensures that propagated
// predicates are always applied, projections and filters can be removed
// from where they come from. Once projections and filters can be removed,
// TODO: it would be useful for demand to be optimized after filters
// that way demand only includes the columns that are still necessary after
// the filters are applied.
optimize_dataflow_demand(dataflow)?;
// A smaller logical optimization pass after projections and filters are
// pushed down across views.
optimize_dataflow_relations(dataflow, indexes, &Optimizer::logical_cleanup_pass())?;
// Physical optimization pass
optimize_dataflow_relations(dataflow, indexes, &Optimizer::physical_optimizer())?;
optimize_dataflow_monotonic(dataflow)?;
Ok(())
}
/// Inline views used in one other view, and in no exported objects.
fn inline_views(dataflow: &mut DataflowDesc) -> Result<(), TransformError> {
// We cannot inline anything whose `BuildDesc::id` appears in either the
// `index_exports` or `sink_exports` of `dataflow`, because we lose our
// ability to name it.
// A view can / should be in-lined in another view if it is only used by
// one subsequent view. If there are two distinct views that have not
// themselves been merged, then too bad and it doesn't get inlined.
// Starting from the *last* object to build, walk backwards and inline
// any view that is neither referenced by a `index_exports` nor
// `sink_exports` nor more than two remaining objects to build.
for index in (0..dataflow.objects_to_build.len()).rev() {
// Capture the name used by others to reference this view.
let global_id = dataflow.objects_to_build[index].id;
// Determine if any exports directly reference this view.
let mut occurs_in_export = false;
for (_gid, sink_desc) in dataflow.sink_exports.iter() {
if sink_desc.from == global_id {
occurs_in_export = true;
}
}
for (_, index_desc, _) in dataflow.index_exports.iter() {
if index_desc.on_id == global_id {
occurs_in_export = true;
}
}
// Count the number of subsequent views that reference this view.
let mut occurrences_in_later_views = Vec::new();
for other in (index + 1)..dataflow.objects_to_build.len() {
if dataflow.objects_to_build[other]
.view
.global_uses()
.contains(&global_id)
{
occurrences_in_later_views.push(other);
}
}
// Inline if the view is referenced in one view and no exports.
if !occurs_in_export && occurrences_in_later_views.len() == 1 {
let other = occurrences_in_later_views[0];
// We can remove this view and insert it in the later view,
// but are not able to relocate the later view `other`.
// When splicing in the `index` view, we need to create disjoint
// identifiers for the Let's `body` and `value`, as well as a new
// identifier for the binding itself. Following `UpdateLet`, we
// go with the binding first, then the value, then the body.
let update_let = crate::update_let::UpdateLet::default();
let mut id_gen = crate::IdGen::default();
let new_local = LocalId::new(id_gen.allocate_id());
// Use the same `id_gen` to assign new identifiers to `index`.
update_let.action(
dataflow.objects_to_build[index].view.as_inner_mut(),
&mut HashMap::new(),
&mut id_gen,
)?;
// Assign new identifiers to the other relation.
update_let.action(
dataflow.objects_to_build[other].view.as_inner_mut(),
&mut HashMap::new(),
&mut id_gen,
)?;
// Install the `new_local` name wherever `global_id` was used.
dataflow.objects_to_build[other]
.view
.as_inner_mut()
.visit_mut_post(&mut |expr| {
if let MirRelationExpr::Get { id, .. } = expr {
if id == &Id::Global(global_id) {
*id = Id::Local(new_local);
}
}
});
// With identifiers rewritten, we can replace `other` with
// a `MirRelationExpr::Let` binding, whose value is `index` and
// whose body is `other`.
let body = dataflow.objects_to_build[other]
.view
.as_inner_mut()
.take_dangerous();
let value = dataflow.objects_to_build[index]
.view
.as_inner_mut()
.take_dangerous();
*dataflow.objects_to_build[other].view.as_inner_mut() = MirRelationExpr::Let {
id: new_local,
value: Box::new(value),
body: Box::new(body),
};
dataflow.objects_to_build.remove(index);
}
}
Ok(())
}
/// Performs either the logical or the physical optimization pass on the
/// dataflow using the supplied set of indexes.
fn optimize_dataflow_relations(
dataflow: &mut DataflowDesc,
indexes: &HashMap<GlobalId, Vec<(GlobalId, Vec<MirScalarExpr>)>>,
optimizer: &Optimizer,
) -> Result<(), TransformError> {
// Re-optimize each dataflow
// TODO(mcsherry): we should determine indexes from the optimized representation
// just before we plan to install the dataflow. This would also allow us to not
// add indexes imperatively to `DataflowDesc`.
for object in dataflow.objects_to_build.iter_mut() {
// Re-name bindings to accommodate other analyses, specifically
// `InlineLet` which probably wants a reworking in any case.
// Re-run all optimizations on the composite views.
optimizer.transform(object.view.as_inner_mut(), &indexes)?;
}
Ok(())
}
/// Pushes demand information from published outputs to dataflow inputs,
/// projecting away unnecessary columns.
///
/// Dataflows that exist for the sake of generating plan explanations do not
/// have published outputs. In this case, we push demand information from views
/// not depended on by other views to dataflow inputs.
fn optimize_dataflow_demand(dataflow: &mut DataflowDesc) -> Result<(), TransformError> {
// Maps id -> union of known columns demanded from the source/view with the
// corresponding id.
let mut demand = HashMap::new();
// Demand all columns of inputs to sinks.
for (_id, sink) in dataflow.sink_exports.iter() {
let input_id = sink.from;
demand
.entry(Id::Global(input_id))
.or_insert_with(BTreeSet::new)
.extend(0..dataflow.arity_of(&input_id));
}
// Demand all columns of inputs to exported indexes.
for (_id, desc, _typ) in dataflow.index_exports.iter() {
let input_id = desc.on_id;
demand
.entry(Id::Global(input_id))
.or_insert_with(BTreeSet::new)
.extend(0..dataflow.arity_of(&input_id));
}
optimize_dataflow_demand_inner(
dataflow
.objects_to_build
.iter_mut()
.rev()
.map(|build_desc| (Id::Global(build_desc.id), build_desc.view.as_inner_mut())),
&mut demand,
)?;
// Push demand information into the SourceDesc.
for (source_id, source) in dataflow.source_imports.iter_mut() {
if let Some(columns) = demand.get(&Id::Global(*source_id)).clone() {
// Install no-op demand information if none exists.
if source.operators.is_none() {
source.operators = Some(LinearOperator {
predicates: Vec::new(),
projection: (0..source.description.desc.arity()).collect(),
})
}
// Restrict required columns by those identified as demanded.
if let Some(operator) = &mut source.operators {
operator.projection.retain(|col| columns.contains(col));
}
}
}
Ok(())
}
/// Pushes demand through views in `view_sequence` in order, removing
/// columns not demanded.
///
/// This method is made public for the sake of testing.
/// TODO: make this private once we allow multiple exports per dataflow.
pub fn optimize_dataflow_demand_inner<'a, I>(
view_sequence: I,
demand: &mut HashMap<Id, BTreeSet<usize>>,
) -> Result<(), TransformError>
where
I: Iterator<Item = (Id, &'a mut MirRelationExpr)>,
{
// Maps id -> The projection that was pushed down on the view with the
// corresponding id.
let mut applied_projection = HashMap::new();
// Collect the mutable references to views after pushing projection down
// in order to run cleanup actions on them in a second loop.
let mut view_refs = Vec::new();
let projection_pushdown = crate::projection_pushdown::ProjectionPushdown;
for (id, view) in view_sequence {
if let Some(columns) = demand.get(&id) {
let projection_pushed_down = columns.iter().map(|c| *c).collect();
// Push down the projection consisting of the entries of `columns`
// in increasing order.
projection_pushdown.action(view, &projection_pushed_down, demand);
applied_projection.insert(id, projection_pushed_down);
} else if id == Id::Global(GlobalId::Explain) {
// If we just want to explain the plan for a given view, then there
// will be no upstream demand. Just demand all columns from views
// that are not depended on by another view.
let arity = view.arity();
projection_pushdown.action(view, &(0..arity).collect(), demand);
}
view_refs.push(view);
}
let typ_update = crate::update_let::UpdateLet::default();
for view in view_refs {
// Update column references to views where projections were pushed down.
projection_pushdown.update_projection_around_get(view, &applied_projection);
// Types need to be updated after ProjectionPushdown
// because the width of each view may have changed.
typ_update.action(view, &mut HashMap::new(), &mut IdGen::default())?;
}
Ok(())
}
/// Pushes predicate to dataflow inputs.
fn optimize_dataflow_filters(dataflow: &mut DataflowDesc) -> Result<(), TransformError> {
// Contains id -> predicates map, describing those predicates that
// can (but need not) be applied to the collection named by `id`.
let mut predicates = HashMap::<Id, HashSet<expr::MirScalarExpr>>::new();
// Propagate predicate information from outputs to inputs.
optimize_dataflow_filters_inner(
dataflow
.objects_to_build
.iter_mut()
.rev()
.map(|build_desc| (Id::Global(build_desc.id), build_desc.view.as_inner_mut())),
&mut predicates,
)?;
// Push predicate information into the SourceDesc.
for (source_id, source) in dataflow.source_imports.iter_mut() {
if let Some(list) = predicates.get(&Id::Global(*source_id)).clone() {
// Install no-op predicate information if none exists.
if source.operators.is_none() {
source.operators = Some(LinearOperator {
predicates: Vec::new(),
projection: (0..source.description.desc.arity()).collect(),
})
}
// Add any predicates that can be pushed to the source.
if let Some(operator) = &mut source.operators {
operator.predicates.extend(list.iter().cloned());
operator.predicates.sort();
}
}
}
Ok(())
}
/// Pushes filters down through views in `view_sequence` in order.
///
/// This method is made public for the sake of testing.
/// TODO: make this private once we allow multiple exports per dataflow.
pub fn optimize_dataflow_filters_inner<'a, I>(
view_iter: I,
predicates: &mut HashMap<Id, HashSet<expr::MirScalarExpr>>,
) -> Result<(), TransformError>
where
I: Iterator<Item = (Id, &'a mut MirRelationExpr)>,
{
let transform = crate::predicate_pushdown::PredicatePushdown::default();
for (id, view) in view_iter {
if let Some(list) = predicates.get(&id).clone() {
if !list.is_empty() {
*view = view.take_dangerous().filter(list.iter().cloned());
}
}
transform.action(view, predicates)?;
}
Ok(())
}
/// Propagates information about monotonic inputs through views.
pub fn optimize_dataflow_monotonic(dataflow: &mut DataflowDesc) -> Result<(), TransformError> {
let mut monotonic = std::collections::HashSet::new();
for (source_id, source) in dataflow.source_imports.iter_mut() {
if let dataflow_types::sources::SourceConnector::External {
envelope: dataflow_types::sources::SourceEnvelope::None(_),
..
} = source.description.connector
{
monotonic.insert(source_id.clone());
}
}
let monotonic_flag = MonotonicFlag::default();
// Propagate predicate information from outputs to inputs.
for build_desc in dataflow.objects_to_build.iter_mut() {
monotonic_flag.apply(
build_desc.view.as_inner_mut(),
&monotonic,
&mut HashSet::new(),
)?;
}
Ok(())
}