Package org.apache.calcite.plan.volcano

package org.apache.calcite.plan.volcano
Optimizes relational expressions.

Overview

A planner (also known as an optimizer) finds the most efficient implementation of a relational expression.

Interface RelOptPlanner defines a planner, and class VolcanoPlanner is an implementation which uses a dynamic programming technique. It is based upon the Volcano optimizer [1].

Interface RelOptCost defines a cost model; class VolcanoCost is the implementation for a VolcanoPlanner.

A RelSet is a set of equivalent relational expressions. They are equivalent because they will produce the same result for any set of input data. It is an equivalence class: two expressions are in the same set if and only if they are in the same RelSet.

One of the unique features of the optimizer is that expressions can take on a variety of physical traits. Each relational expression has a set of traits. Each trait is described by an implementation of RelTraitDef. Manifestations of the trait implement RelTrait. The most common example of a trait is calling convention: the protocol used to receive and transmit data. ConventionTraitDef defines the trait and Convention enumerates the protocols. Every relational expression has a single calling convention by which it returns its results. Some examples:

  • JdbcConvention is a fairly conventional convention; the results are rows from a JDBC result set.
  • Convention.NONE means that a relational expression cannot be implemented; typically there are rules which can transform it to equivalent, implementable expressions.
  • EnumerableConvention implements the expression by generating Java code. The code places the current row in a Java variable, then calls the piece of code which implements the consuming relational expression. For example, a Java array reader of the names array would generate the following code: 
    String[] names;
 for (int i = 0; i < names.length; i++) {
     String name = names[i];
     // ... code for consuming relational expression ...
 }

New traits are added to the planner in one of two ways:

  1. If the new trait is integral to Calcite, then each and every implementation of RelNode should include its manifestation of the trait as part of the RelTraitSet passed to AbstractRelNode's constructor. It may be useful to provide alternate AbstractRelNode constructors if most relational expressions use a single manifestation of the trait.
  2. If the new trait describes some aspect of a Farrago extension, then the RelNodes passed to VolcanoPlanner.setRoot(org.apache.calcite.rel.RelNode) should have their trait sets expanded before the setRoot(RelNode) call.

The second trait extension mechanism requires that implementations of Object.clone() must not assume the type and quantity of traits in their trait set. In either case, the new RelTraitDef implementation must be VolcanoPlanner.addRelTraitDef(org.apache.calcite.plan.RelTraitDef) registered with the planner.

A RelSubset is a subset of a RelSet containing expressions which are equivalent and which have the same Convention. Like RelSet, it is an equivalence class.

Related packages

Details

Sets merge when the result of a rule already exists in another set. This implies that all of the expressions are equivalent. The RelSets are merged, and so are the contained RelSubsets.

Expression registration.

  • Expression is added to a set. We may find that an equivalent expression already exists. Otherwise, this is the moment when an expression becomes public, and fixed. Its digest is assigned, which allows us to quickly find identical expressions.
  • We match operands, figure out which rules are applicable, and generate rule calls. The rule calls are placed on a queue, and the important ones are called later.
  • RuleCalls allow us to defer the invocation of rules. When an expression is registered

Algorithm

To optimize a relational expression R:

1. Register R.

2. Create rule-calls for all applicable rules.

3. Rank the rule calls by importance.

4. Call the most important rule

5. Repeat.

Importance. A rule-call is important if it is likely to produce better implementation of a relexp on the plan's critical path. Hence (a) it produces a member of an important RelSubset, (b) its children are cheap.

Conversion. Conversions are difficult because we have to work backwards from the goal.

Rule triggering

The rules are:

  1. PushFilterThroughProjectRule. Operands: 
    Filter
   Project
  2. CombineProjectsRule. Operands: 
    Project
   Project

A rule can be triggered by a change to any of its operands. Consider the rule to combine two filters into one. It would have operands [Filter [Filter]]. If I register a new Filter, it will trigger the rule in 2 places. Consider: 

Project (deptno)                              [exp 1, subset A]
   Filter (gender='F')                         [exp 2, subset B]
     Project (deptno, gender, empno)           [exp 3, subset C]
       Project (deptno, gender, empno, salary) [exp 4, subset D]
         TableScan (emp)                       [exp 0, subset X]

Apply PushFilterThroughProjectRule to [exp 2, exp 3]: 

Project (deptno)                              [exp 1, subset A]
   Project (deptno, gender, empno)             [exp 5, subset B]
     Filter (gender='F')                       [exp 6, subset E]
       Project (deptno, gender, empno, salary) [exp 4, subset D]
         TableScan (emp)                       [exp 0, subset X]

Two new expressions are created. Expression 5 is in subset B (because it is equivalent to expression 2), and expression 6 is in a new equivalence class, subset E.

The products of a applying a rule can trigger a cascade of rules. Even in this simple system (2 rules and 4 initial expressions), two more rules are triggered:

  • Registering exp 5 triggers CombineProjectsRule(exp 1, exp 5), which creates 
    Project (deptno)                              [exp 7, subset A]
   Filter (gender='F')                         [exp 6, subset E]
     Project (deptno, gender, empno, salary)   [exp 4, subset D]
       TableScan (emp)                         [exp 0, subset X]
  • Registering exp 6 triggers PushFilterThroughProjectRule(exp 6, exp 4), which creates 
    Project (deptno)                              [exp 1, subset A]
   Project (deptno, gender, empno)             [exp 5, subset B]
     Project (deptno, gender, empno, salary)   [exp 8, subset E]
       Filter (gender='F')                     [exp 9, subset F]
         TableScan (emp)                       [exp 0, subset X]

Each rule application adds additional members to existing subsets. The non-singleton subsets are now A {1, 7}, B {2, 5} and E {6, 8}, and new combinations are possible. For example, CombineProjectsRule(exp 7, exp 8) further reduces the depth of the tree to: 

Project (deptno)                          [exp 10, subset A]
   Filter (gender='F')                     [exp 9, subset F]
     TableScan (emp)                       [exp 0, subset X]

Todo: show how rules can cause subsets to merge.

Conclusion:

  1. A rule can be triggered by any of its operands.
  2. If a subset is a child of more than one parent, it can trigger rule matches for any of its parents.
  3. Registering one relexp can trigger several rules (and even the same rule several times).
  4. Firing rules can cause subsets to merge.

References

1. The Volcano Optimizer Generator: Extensibility and Efficient Search - Goetz Graefe, William J. McKenna (1993).