query/dist_plan/
commutativity.rs

1// Copyright 2023 Greptime Team
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15use std::collections::HashSet;
16use std::sync::Arc;
17
18use common_function::aggrs::aggr_wrapper::{StateMergeHelper, is_all_aggr_exprs_steppable};
19use common_telemetry::debug;
20use datafusion::error::Result as DfResult;
21use datafusion_common::tree_node::{TreeNode, TreeNodeRecursion};
22use datafusion_expr::{Expr, LogicalPlan, UserDefinedLogicalNode};
23use promql::extension_plan::{
24    EmptyMetric, InstantManipulate, RangeManipulate, SeriesDivide, SeriesNormalize,
25};
26use store_api::metric_engine_consts::DATA_SCHEMA_TSID_COLUMN_NAME;
27
28use crate::dist_plan::MergeScanLogicalPlan;
29use crate::dist_plan::analyzer::AliasMapping;
30use crate::dist_plan::merge_sort::{MergeSortLogicalPlan, merge_sort_transformer};
31
32pub struct StepTransformAction {
33    extra_parent_plans: Vec<LogicalPlan>,
34    new_child_plan: Option<LogicalPlan>,
35}
36
37/// generate the upper aggregation plan that will execute on the frontend.
38/// Basically a logical plan resembling the following:
39/// Projection:
40///     Aggregate:
41///
42/// from Aggregate
43///
44/// The upper Projection exists sole to make sure parent plan can recognize the output
45/// of the upper aggregation plan.
46pub fn step_aggr_to_upper_aggr(
47    aggr_plan: &LogicalPlan,
48) -> datafusion_common::Result<StepTransformAction> {
49    let LogicalPlan::Aggregate(input_aggr) = aggr_plan else {
50        return Err(datafusion_common::DataFusionError::Plan(
51            "step_aggr_to_upper_aggr only accepts Aggregate plan".to_string(),
52        ));
53    };
54    if !is_all_aggr_exprs_steppable(&input_aggr.aggr_expr) {
55        return Err(datafusion_common::DataFusionError::NotImplemented(format!(
56            "Some aggregate expressions are not steppable in [{}]",
57            input_aggr
58                .aggr_expr
59                .iter()
60                .map(|e| e.to_string())
61                .collect::<Vec<_>>()
62                .join(", ")
63        )));
64    }
65
66    let step_aggr_plan = StateMergeHelper::split_aggr_node(input_aggr.clone())?;
67
68    // TODO(discord9): remove duplication
69    let ret = StepTransformAction {
70        extra_parent_plans: vec![step_aggr_plan.upper_merge.clone()],
71        new_child_plan: Some(step_aggr_plan.lower_state.clone()),
72    };
73    Ok(ret)
74}
75
76#[allow(dead_code)]
77pub enum Commutativity {
78    Commutative,
79    PartialCommutative,
80    ConditionalCommutative(Option<Transformer>),
81    TransformedCommutative {
82        /// Return plans from parent to child order
83        transformer: Option<StageTransformer>,
84    },
85    NonCommutative,
86    Unimplemented,
87    /// For unrelated plans like DDL
88    Unsupported,
89}
90
91pub struct Categorizer {}
92
93impl Categorizer {
94    pub fn check_plan(
95        plan: &LogicalPlan,
96        partition_cols: Option<AliasMapping>,
97    ) -> DfResult<Commutativity> {
98        // Subquery is treated separately in `inspect_plan_with_subquery`. To avoid rewrite the
99        // "maybe rewritten" plan, stop the check here.
100        if has_subquery(plan)? {
101            return Ok(Commutativity::Unimplemented);
102        }
103
104        let partition_cols = partition_cols.unwrap_or_default();
105
106        let comm = match plan {
107            LogicalPlan::Projection(proj) => {
108                for expr in &proj.expr {
109                    let commutativity = Self::check_expr(expr);
110                    if !matches!(commutativity, Commutativity::Commutative) {
111                        return Ok(commutativity);
112                    }
113                }
114                Commutativity::Commutative
115            }
116            LogicalPlan::Filter(filter) => Self::check_expr(&filter.predicate),
117            LogicalPlan::Window(_) => Commutativity::Unimplemented,
118            LogicalPlan::Aggregate(aggr) => {
119                // The state/merge split maps each group expression to one output column,
120                // which doesn't hold for grouping sets.
121                let has_grouping_set = aggr
122                    .group_expr
123                    .iter()
124                    .any(|expr| matches!(expr, Expr::GroupingSet(_)));
125                let is_all_steppable =
126                    !has_grouping_set && is_all_aggr_exprs_steppable(&aggr.aggr_expr);
127                let matches_partition = Self::check_partition(&aggr.group_expr, &partition_cols);
128                if !matches_partition && is_all_steppable {
129                    debug!("Plan is steppable: {plan}");
130                    return Ok(Commutativity::TransformedCommutative {
131                        transformer: Some(Arc::new(|plan: &LogicalPlan| {
132                            debug!("Before Step optimize: {plan}");
133                            let ret = step_aggr_to_upper_aggr(plan);
134                            ret.inspect_err(|err| {
135                                common_telemetry::error!("Failed to step aggregate plan: {err:?}");
136                            })
137                            .map(|s| TransformerAction {
138                                extra_parent_plans: s.extra_parent_plans,
139                                new_child_plan: s.new_child_plan,
140                            })
141                        })),
142                    });
143                }
144                if !matches_partition {
145                    return Ok(Commutativity::NonCommutative);
146                }
147                for expr in &aggr.aggr_expr {
148                    let commutativity = Self::check_expr(expr);
149                    if !matches!(commutativity, Commutativity::Commutative) {
150                        return Ok(commutativity);
151                    }
152                }
153                // all group by expressions are partition columns can push down, unless
154                // another push down(including `Limit` or `Sort`) is already in progress(which will then prevent next cond commutative node from being push down).
155                // TODO(discord9): This is a temporary solution(that works), a better description of
156                // commutativity is needed under this situation.
157                Commutativity::ConditionalCommutative(None)
158            }
159            LogicalPlan::Sort(_sort) => {
160                if partition_cols.is_empty() {
161                    return Ok(Commutativity::Commutative);
162                }
163
164                // sort plan needs to consider column priority
165                // Change Sort to MergeSort which assumes the input streams are already sorted hence can be more efficient.
166                Commutativity::ConditionalCommutative(Some(Arc::new(merge_sort_transformer)))
167            }
168            LogicalPlan::Join(_) => Commutativity::NonCommutative,
169            LogicalPlan::Repartition(_) => {
170                // unsupported? or non-commutative
171                Commutativity::Unimplemented
172            }
173            LogicalPlan::Union(_) => Commutativity::Unimplemented,
174            LogicalPlan::TableScan(_) => Commutativity::Commutative,
175            LogicalPlan::EmptyRelation(_) => Commutativity::NonCommutative,
176            LogicalPlan::Subquery(_) => Commutativity::Unimplemented,
177            LogicalPlan::SubqueryAlias(_) => Commutativity::Commutative,
178            LogicalPlan::Limit(limit) => {
179                // Only execute `fetch` on remote nodes.
180                // wait for https://github.com/apache/arrow-datafusion/pull/7669
181                if partition_cols.is_empty() && limit.fetch.is_some() {
182                    Commutativity::Commutative
183                } else if limit.skip.is_none() && limit.fetch.is_some() {
184                    Commutativity::PartialCommutative
185                } else {
186                    Commutativity::Unimplemented
187                }
188            }
189            LogicalPlan::Extension(extension) => {
190                Self::check_extension_plan(extension.node.as_ref() as _, &partition_cols)
191            }
192            LogicalPlan::Distinct(_) => {
193                if partition_cols.is_empty() {
194                    Commutativity::Commutative
195                } else {
196                    Commutativity::PartialCommutative
197                }
198            }
199            LogicalPlan::Unnest(_) => Commutativity::Commutative,
200            LogicalPlan::Statement(_) => Commutativity::Unsupported,
201            LogicalPlan::Values(_) => Commutativity::Unsupported,
202            LogicalPlan::Explain(_) => Commutativity::Unsupported,
203            LogicalPlan::Analyze(_) => Commutativity::Unsupported,
204            LogicalPlan::DescribeTable(_) => Commutativity::Unsupported,
205            LogicalPlan::Dml(_) => Commutativity::Unsupported,
206            LogicalPlan::Ddl(_) => Commutativity::Unsupported,
207            LogicalPlan::Copy(_) => Commutativity::Unsupported,
208            LogicalPlan::RecursiveQuery(_) => Commutativity::Unsupported,
209        };
210
211        Ok(comm)
212    }
213
214    pub fn check_extension_plan(
215        plan: &dyn UserDefinedLogicalNode,
216        partition_cols: &AliasMapping,
217    ) -> Commutativity {
218        match plan.name() {
219            name if name == SeriesDivide::name() => {
220                let series_divide = plan.as_any().downcast_ref::<SeriesDivide>().unwrap();
221                // Metric engine `__tsid` uniquely identifies a time-series. Treat a series divide
222                // that keys by `__tsid` as commutative across regions so it can be pushed down.
223                if series_divide
224                    .tags()
225                    .iter()
226                    .any(|tag| tag == DATA_SCHEMA_TSID_COLUMN_NAME)
227                {
228                    return Commutativity::Commutative;
229                }
230
231                let tags = series_divide.tags().iter().collect::<HashSet<_>>();
232
233                for all_alias in partition_cols.values() {
234                    let all_alias = all_alias.iter().map(|c| &c.name).collect::<HashSet<_>>();
235                    if tags.intersection(&all_alias).count() == 0 {
236                        return Commutativity::NonCommutative;
237                    }
238                }
239
240                Commutativity::Commutative
241            }
242            name if name == SeriesNormalize::name()
243                || name == InstantManipulate::name()
244                || name == RangeManipulate::name() =>
245            {
246                // They should always follows Series Divide.
247                // Either all commutative or all non-commutative (which will be blocked by SeriesDivide).
248                Commutativity::Commutative
249            }
250            name if name == EmptyMetric::name()
251                || name == MergeScanLogicalPlan::name()
252                || name == MergeSortLogicalPlan::name() =>
253            {
254                Commutativity::Unimplemented
255            }
256            _ => Commutativity::Unsupported,
257        }
258    }
259
260    pub fn check_expr(expr: &Expr) -> Commutativity {
261        #[allow(deprecated)]
262        match expr {
263            Expr::Column(_)
264            | Expr::ScalarVariable(_, _)
265            | Expr::Literal(_, _)
266            | Expr::BinaryExpr(_)
267            | Expr::Not(_)
268            | Expr::IsNotNull(_)
269            | Expr::IsNull(_)
270            | Expr::IsTrue(_)
271            | Expr::IsFalse(_)
272            | Expr::IsNotTrue(_)
273            | Expr::IsNotFalse(_)
274            | Expr::Negative(_)
275            | Expr::Between(_)
276            | Expr::Exists(_)
277            | Expr::InList(_)
278            | Expr::Like(_)
279            | Expr::Case(_) => Commutativity::Commutative,
280            // Annotation collection must not affect distribution. Keep scalar UDFs pushdownable;
281            // TODO: Preserve datanode PromQL annotations through plan decoding, return them in
282            // Flight stream metadata, and merge them on the frontend.
283            Expr::ScalarFunction(_) => Commutativity::Commutative,
284            Expr::AggregateFunction(_udaf) => Commutativity::Commutative,
285
286            Expr::SimilarTo(_)
287            | Expr::IsUnknown(_)
288            | Expr::IsNotUnknown(_)
289            | Expr::WindowFunction(_)
290            | Expr::InSubquery(_)
291            | Expr::ScalarSubquery(_)
292            | Expr::HigherOrderFunction(_)
293            | Expr::Lambda(_)
294            | Expr::LambdaVariable(_)
295            | Expr::Wildcard { .. } => Commutativity::Unimplemented,
296
297            Expr::Alias(alias) => Self::check_expr(&alias.expr),
298            Expr::Cast(cast) => Self::check_expr(&cast.expr),
299            Expr::TryCast(try_cast) => Self::check_expr(&try_cast.expr),
300
301            Expr::Unnest(_)
302            | Expr::GroupingSet(_)
303            | Expr::Placeholder(_)
304            | Expr::OuterReferenceColumn(_, _)
305            | Expr::SetComparison(_) => Commutativity::Unimplemented,
306        }
307    }
308
309    /// Return true if the given expr and partition cols satisfied the rule.
310    /// In this case the plan can be treated as fully commutative.
311    ///
312    /// So only if every partition column is itself one of `exprs`, return true.
313    /// Otherwise return false.
314    ///
315    /// An expression that only references a partition column, like `substr(host, 3, 1)`
316    /// or `k % 2`, doesn't count: it can put rows from different partitions into the same
317    /// group.
318    fn check_partition(exprs: &[Expr], partition_cols: &AliasMapping) -> bool {
319        let group_cols = exprs
320            .iter()
321            .filter_map(|expr| {
322                let mut expr = expr;
323                while let Expr::Alias(alias) = expr {
324                    expr = &alias.expr;
325                }
326                match expr {
327                    Expr::Column(column) => Some(column.name.clone()),
328                    _ => None,
329                }
330            })
331            .collect::<HashSet<_>>();
332        for all_alias in partition_cols.values() {
333            let all_alias = all_alias
334                .iter()
335                .map(|c| c.name.clone())
336                .collect::<HashSet<_>>();
337            // check if group columns intersect with all alias of partition columns
338            // is empty, if it's empty, not all partition columns show up in `exprs`
339            if group_cols.intersection(&all_alias).count() == 0 {
340                return false;
341            }
342        }
343
344        true
345    }
346}
347
348#[cfg(test)]
349mod tests {
350    use std::collections::{BTreeMap, BTreeSet};
351
352    use datafusion_common::Column;
353    use datafusion_expr::LogicalPlanBuilder;
354    use datafusion_expr::expr::ScalarFunction;
355    use datafusion_functions::core::coalesce;
356    use promql::functions::{
357        NativeHistogramDrop, NativeHistogramFraction, NativeHistogramQuantile,
358    };
359
360    use super::*;
361
362    #[test]
363    fn series_divide_by_tsid_is_commutative() {
364        let input = LogicalPlanBuilder::empty(false).build().unwrap();
365        let series_divide = SeriesDivide::new(
366            vec![DATA_SCHEMA_TSID_COLUMN_NAME.to_string()],
367            "ts".to_string(),
368            input,
369        );
370
371        let partition_cols: AliasMapping = BTreeMap::from([(
372            "some_partition_col".to_string(),
373            BTreeSet::from([Column::from_name("some_partition_col")]),
374        )]);
375
376        let commutativity = Categorizer::check_extension_plan(&series_divide, &partition_cols);
377        assert!(matches!(commutativity, Commutativity::Commutative));
378    }
379
380    #[test]
381    fn annotated_histogram_helpers_do_not_block_pushdown() {
382        for udf in [
383            NativeHistogramQuantile::scalar_udf(),
384            NativeHistogramFraction::scalar_udf(),
385            NativeHistogramDrop::warning_bool_false_udf(String::new(), None),
386        ] {
387            let helper = Expr::ScalarFunction(ScalarFunction::new_udf(Arc::new(udf), vec![]));
388            let expr = Expr::ScalarFunction(ScalarFunction::new_udf(coalesce(), vec![helper]));
389            assert!(matches!(
390                Categorizer::check_expr(&expr),
391                Commutativity::Commutative
392            ));
393        }
394    }
395}
396
397pub type Transformer = Arc<dyn Fn(&LogicalPlan) -> Option<LogicalPlan>>;
398
399/// Returns transformer action that need to be applied
400pub type StageTransformer = Arc<dyn Fn(&LogicalPlan) -> DfResult<TransformerAction>>;
401
402/// The Action that a transformer should take on the plan.
403pub struct TransformerAction {
404    /// list of plans that need to be applied to parent plans, in the order of parent to child.
405    /// i.e. if this returns `[Projection, Aggregate]`, then the parent plan should be transformed to
406    /// ```ignore
407    /// Original Parent Plan:
408    ///     Projection:
409    ///         Aggregate:
410    ///             MergeScan: ...
411    /// ```
412    pub extra_parent_plans: Vec<LogicalPlan>,
413    /// new child plan, if None, use the original plan.
414    pub new_child_plan: Option<LogicalPlan>,
415}
416
417pub fn partial_commutative_transformer(plan: &LogicalPlan) -> Option<LogicalPlan> {
418    Some(plan.clone())
419}
420
421fn has_subquery(plan: &LogicalPlan) -> DfResult<bool> {
422    let mut found = false;
423    plan.apply_expressions(|e| {
424        e.apply(|x| {
425            if matches!(
426                x,
427                Expr::Exists(_) | Expr::InSubquery(_) | Expr::ScalarSubquery(_)
428            ) {
429                found = true;
430                Ok(TreeNodeRecursion::Stop)
431            } else {
432                Ok(TreeNodeRecursion::Continue)
433            }
434        })
435    })?;
436    Ok(found)
437}
438
439#[cfg(test)]
440mod test {
441    use datafusion_expr::{LogicalPlanBuilder, Sort};
442
443    use super::*;
444
445    #[test]
446    fn sort_on_empty_partition() {
447        let plan = LogicalPlan::Sort(Sort {
448            expr: vec![],
449            input: Arc::new(LogicalPlanBuilder::empty(false).build().unwrap()),
450            fetch: None,
451        });
452        assert!(matches!(
453            Categorizer::check_plan(&plan, Some(Default::default())).unwrap(),
454            Commutativity::Commutative
455        ));
456    }
457}
query/dist_plan/commutativity.rs

query/dist_plan/
commutativity.rs
