-
Notifications
You must be signed in to change notification settings - Fork 1.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
optimize queries where lhs and rhs of predicate are equal #10444
Changes from 5 commits
32da3e2
dc932e4
7ca6972
7d3c85c
c7c578f
c432e13
b303175
bacf64e
ffd0f02
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
|
@@ -26,6 +26,7 @@ | |
import org.apache.pinot.common.request.PinotQuery; | ||
import org.apache.pinot.core.query.optimizer.filter.FilterOptimizer; | ||
import org.apache.pinot.core.query.optimizer.filter.FlattenAndOrFilterOptimizer; | ||
import org.apache.pinot.core.query.optimizer.filter.IdenticalPredicateFilterOptimizer; | ||
import org.apache.pinot.core.query.optimizer.filter.MergeEqInFilterOptimizer; | ||
import org.apache.pinot.core.query.optimizer.filter.MergeRangeFilterOptimizer; | ||
import org.apache.pinot.core.query.optimizer.filter.NumericalFilterOptimizer; | ||
|
@@ -44,7 +45,7 @@ public class QueryOptimizer { | |
// values to the proper format so that they can be properly parsed | ||
private static final List<FilterOptimizer> FILTER_OPTIMIZERS = | ||
Arrays.asList(new FlattenAndOrFilterOptimizer(), new MergeEqInFilterOptimizer(), new NumericalFilterOptimizer(), | ||
new TimePredicateFilterOptimizer(), new MergeRangeFilterOptimizer()); | ||
new TimePredicateFilterOptimizer(), new MergeRangeFilterOptimizer(), new IdenticalPredicateFilterOptimizer()); | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Should we apply this optimizer in the end? If it doesn't rely on other optimizers, we can put it next to the flatten optimizer to avoid other optimizer to optimize on identical predicate There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. good call. it actually caught another null point exception moving it earlier |
||
|
||
private static final List<StatementOptimizer> STATEMENT_OPTIMIZERS = | ||
Collections.singletonList(new StringPredicateFilterOptimizer()); | ||
|
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,98 @@ | ||
/** | ||
* Licensed to the Apache Software Foundation (ASF) under one | ||
* or more contributor license agreements. See the NOTICE file | ||
* distributed with this work for additional information | ||
* regarding copyright ownership. The ASF licenses this file | ||
* to you under the Apache License, Version 2.0 (the | ||
* "License"); you may not use this file except in compliance | ||
* with the License. You may obtain a copy of the License at | ||
* | ||
* http://www.apache.org/licenses/LICENSE-2.0 | ||
* | ||
* Unless required by applicable law or agreed to in writing, | ||
* software distributed under the License is distributed on an | ||
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY | ||
* KIND, either express or implied. See the License for the | ||
* specific language governing permissions and limitations | ||
* under the License. | ||
*/ | ||
package org.apache.pinot.core.query.optimizer.filter; | ||
|
||
import java.util.List; | ||
import javax.annotation.Nullable; | ||
import org.apache.pinot.common.request.Expression; | ||
import org.apache.pinot.common.request.Function; | ||
import org.apache.pinot.common.utils.request.RequestUtils; | ||
import org.apache.pinot.spi.data.Schema; | ||
import org.apache.pinot.sql.FilterKind; | ||
|
||
/** | ||
* This base class acts as a helper for any optimizer that is effectively removing filter conditions. | ||
* It provides TRUE/FALSE literal classes that can be used to replace filter expressions that are always true/false. | ||
* It provides an optimization implementation for AND/OR/NOT expressions. | ||
*/ | ||
public abstract class BaseAndOrBooleanFilterOptimizer implements FilterOptimizer { | ||
|
||
protected static final Expression TRUE = RequestUtils.getLiteralExpression(true); | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This file doesn't follow the Pinot Style There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. good eye. I was working on a new laptop and hadn't set that up. |
||
protected static final Expression FALSE = RequestUtils.getLiteralExpression(false); | ||
|
||
@Override | ||
public abstract Expression optimize(Expression filterExpression, @Nullable Schema schema); | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. (minor) No need to override this API to an abstract method |
||
|
||
/** | ||
* If any of the operands of AND function is "false", then the AND function itself is false and can be replaced with | ||
* "false" literal. Otherwise, remove all the "true" operands of the AND function. Similarly, if any of the operands | ||
* of OR function is "true", then the OR function itself is true and can be replaced with "true" literal. Otherwise, | ||
* remove all the "false" operands of the OR function. | ||
*/ | ||
protected Expression optimizeCurrent(Expression expression) { | ||
Function function = expression.getFunctionCall(); | ||
String operator = function.getOperator(); | ||
List<Expression> operands = function.getOperands(); | ||
if (operator.equals(FilterKind.AND.name())) { | ||
// If any of the literal operands are always false, then replace AND function with FALSE. | ||
for (Expression operand : operands) { | ||
if (isAlwaysFalse(operand)) { | ||
return FALSE; | ||
} | ||
} | ||
|
||
// Remove all Literal operands that are always true. | ||
operands.removeIf(this::isAlwaysTrue); | ||
if (operands.isEmpty()) { | ||
return TRUE; | ||
} | ||
} else if (operator.equals(FilterKind.OR.name())) { | ||
// If any of the literal operands are always true, then replace OR function with TRUE | ||
for (Expression operand : operands) { | ||
if (isAlwaysTrue(operand)) { | ||
return TRUE; | ||
} | ||
} | ||
|
||
// Remove all Literal operands that are always false. | ||
operands.removeIf(this::isAlwaysFalse); | ||
if (operands.isEmpty()) { | ||
return FALSE; | ||
} | ||
} else if (operator.equals(FilterKind.NOT.name())) { | ||
assert operands.size() == 1; | ||
Expression operand = operands.get(0); | ||
if (isAlwaysTrue(operand)) { | ||
return FALSE; | ||
} | ||
if (isAlwaysFalse(operand)) { | ||
return TRUE; | ||
} | ||
} | ||
return expression; | ||
} | ||
|
||
protected boolean isAlwaysFalse(Expression operand) { | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. If we don't need to override this method (see comment below), we can change There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I've reworked the interface to be a little clearer. The base class handles the DFS, and users just implement the base case. Let me know if this looks better. |
||
return operand.equals(FALSE); | ||
} | ||
|
||
protected boolean isAlwaysTrue(Expression operand) { | ||
return operand.equals(TRUE); | ||
} | ||
} |
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,130 @@ | ||
/** | ||
* Licensed to the Apache Software Foundation (ASF) under one | ||
* or more contributor license agreements. See the NOTICE file | ||
* distributed with this work for additional information | ||
* regarding copyright ownership. The ASF licenses this file | ||
* to you under the Apache License, Version 2.0 (the | ||
* "License"); you may not use this file except in compliance | ||
* with the License. You may obtain a copy of the License at | ||
* | ||
* http://www.apache.org/licenses/LICENSE-2.0 | ||
* | ||
* Unless required by applicable law or agreed to in writing, | ||
* software distributed under the License is distributed on an | ||
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY | ||
* KIND, either express or implied. See the License for the | ||
* specific language governing permissions and limitations | ||
* under the License. | ||
*/ | ||
package org.apache.pinot.core.query.optimizer.filter; | ||
|
||
import java.util.List; | ||
import javax.annotation.Nullable; | ||
import org.apache.pinot.common.request.Expression; | ||
import org.apache.pinot.common.request.Function; | ||
import org.apache.pinot.spi.data.Schema; | ||
import org.apache.pinot.sql.FilterKind; | ||
|
||
|
||
/** | ||
* This optimizer converts all predicates where the left hand side == right hand side to | ||
* a simple TRUE/FALSE literal value. While filters like, WHERE 1=1 OR "col1"="col1" are not | ||
* typical, they end up expensive in Pinot because they are rewritten as A-A==0. | ||
*/ | ||
public class IdenticalPredicateFilterOptimizer extends BaseAndOrBooleanFilterOptimizer { | ||
|
||
@Override | ||
public Expression optimize(Expression filterExpression, @Nullable Schema schema) { | ||
Function function = filterExpression.getFunctionCall(); | ||
if (function == null) { | ||
return filterExpression; | ||
} | ||
|
||
List<Expression> operands = function.getOperands(); | ||
FilterKind kind = FilterKind.valueOf(function.getOperator()); | ||
switch (kind) { | ||
case AND: | ||
case OR: | ||
case NOT: | ||
// Recursively traverse the expression tree to find an operator node that can be rewritten. | ||
operands.forEach(operand -> optimize(operand, schema)); | ||
|
||
// We have rewritten the child operands, so rewrite the parent if needed. | ||
return optimizeCurrent(filterExpression); | ||
case EQUALS: | ||
if (hasIdenticalLhsAndRhs(filterExpression)) { | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. (minor) Directly pass the operands There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. ty, no need to recompute it all |
||
return TRUE; | ||
} | ||
return filterExpression; | ||
case NOT_EQUALS: | ||
if (hasIdenticalLhsAndRhs(filterExpression)) { | ||
return FALSE; | ||
} | ||
return filterExpression; | ||
default: | ||
return filterExpression; | ||
} | ||
} | ||
|
||
@Override | ||
protected boolean isAlwaysFalse(Expression operand) { | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Do we need to override this method? After the DFS, all the children should already be optimized |
||
if (super.isAlwaysFalse(operand)) { | ||
return true; | ||
} else if (hasIdenticalLhsAndRhs(operand)) { | ||
return operand.getFunctionCall().getOperator().equals(FilterKind.NOT_EQUALS.name()); | ||
} | ||
return false; | ||
} | ||
|
||
@Override | ||
protected boolean isAlwaysTrue(Expression operand) { | ||
if (super.isAlwaysTrue(operand)) { | ||
return true; | ||
} else if (hasIdenticalLhsAndRhs(operand)) { | ||
return operand.getFunctionCall().getOperator().equals(FilterKind.EQUALS.name()); | ||
} | ||
return false; | ||
} | ||
|
||
/** | ||
* Pinot queries of the WHERE 1 != 1 AND "col1" = "col2" variety are rewritten as | ||
* 1-1 != 0 AND "col1"-"col2" = 0. Therefore, we check specifically for the case where | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. The rewrite is already happening in Since we already get this implementation, we can add a TODO here and revisit later There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I didn't see that class until now. But even then, I think I slightly prefer this more as an optimization than a rewrite. But it's probably easier to do there before it gets converted just for us to convert it back. |
||
* the operand is set up in this fashion. | ||
* | ||
* We return false specifically after every check to ensure we're only continuing when | ||
* the input looks as expected. Otherwise, it's easy to for one of the operand functions | ||
* to return null and fail the query. | ||
*/ | ||
private boolean hasIdenticalLhsAndRhs(Expression operand) { | ||
List<Expression> children = operand.getFunctionCall().getOperands(); | ||
boolean hasTwoChildren = children.size() == 2; | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. (minor) Is there any case There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. a |
||
Expression firstChild = children.get(0); | ||
if (firstChild.getFunctionCall() == null || !hasTwoChildren) { | ||
return false; | ||
} | ||
boolean firstChildIsMinusOperator = firstChild.getFunctionCall().getOperator().equals("minus"); | ||
if (!firstChildIsMinusOperator) { | ||
return false; | ||
} | ||
boolean firstChildHasTwoOperands = firstChild.getFunctionCall().getOperandsSize() == 2; | ||
if (!firstChildHasTwoOperands) { | ||
return false; | ||
} | ||
Expression minusOperandFirstChild = firstChild.getFunctionCall().getOperands().get(0); | ||
Expression minusOperandSecondChild = firstChild.getFunctionCall().getOperands().get(1); | ||
if (minusOperandFirstChild == null || minusOperandSecondChild == null || !minusOperandFirstChild.equals( | ||
minusOperandSecondChild)) { | ||
return false; | ||
} | ||
Expression secondChild = children.get(1); | ||
return isLiteralZero(secondChild); | ||
} | ||
|
||
private boolean isLiteralZero(Expression expression) { | ||
if (!expression.isSetLiteral()) { | ||
return false; | ||
} | ||
Object literalValue = expression.getLiteral().getFieldValue(); | ||
return literalValue.equals(0) || literalValue.equals(0L) || literalValue.equals(0d); | ||
} | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good point. Ignoring the error is more robust, while failing the query can help catch the bug in the optimizer and prevent certain unexpected performance degradation. Currently optimize logic is applied in-place (there is no return value), so I personally prefer directly failing the query since the query might already be modified and messed up
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ya, I was thinking the same about the fact that it might be only semi optimized when this fails. I've updated the comment to reflect this state