"I prefer the KISS principle" is a bit unfair. What I'm advocating is
also simple. But maybe from a different perspective.
I am looking at it from the perspective of people who use rules, and
compose them into rule sets to perform particular optimizations. I
believe this is the most important perspective to focus on. These
people are the key audience to serve. The corpus of re-usable,
composable rules may be Calcite's most important contribution.
To these people, a rule is not a class but an object. It has a
particular signature (in terms of the pattern of RelNodes that it
matches), maybe one or two configuration parameters (e.g.
FilterJoinRule.smart controls whether to automatically strengthen a
LEFT join to INNER), and it has code to generate a new RelNode when
matched.
What do I mean by 'object'? Think of how you use regular expressions
in Java. You call java.util.regex.Pattern.compile("a[bc]*") and it
gives you a Pattern object. You do not know what fields are inside
that Pattern, you do not care whether the object you receive is a
Pattern or some sub-class of Pattern, but you know there are one or
two methods such as 'matcher(CharSequence)' that you can call.
Our current API treats planner rules as classes. If you want to
customize a property of a rule (say make a FilterJoinRule with
smart=false, or change its RelBuilder, or make it match a
CassandraFilter rather than a LogicalFilter) then you have to make a
new rule instance by calling the constructor.
When you call the constructor, you have to supply ALL of the
properties, including operands, and including properties that you
don't know or care about. If someone in three months adds a new
property, the signature of the constructor will change, and you will
have to go back and change your code. This is broken.
Haisheng asked for evidence that the current system is broken.
https://issues.apache.org/jira/browse/CALCITE-3825 ("Split
AbstractMaterializedViewRule into multiple classes") is one example.
In my company, that change literally broke our code because we had
changed properties of some rules.
Another example is https://issues.apache.org/jira/browse/CALCITE-3975
("Add options to ProjectFilterTransposeRule to push down project and
filter expressions whole, not just field references"). With that
change, I broke my own code. I added two new arguments, 'boolean
wholeProject, boolean wholeFilter' to ProjectFilterTransposeRule's
constructor, deprecated the previous constructor, and created quite a
few conflicts in another dev branch that I was working on.
These kinds of problems are not uncommon. Because it is difficult to
reconfigure rules, or evolve them by adding extra arguments, people
are more likely to copy-paste rule code into their own projects, and
less like to contribute back.
In most rules, operands are created evaluating a complex expression -
made up of calls to static methods such as RelOptRule.operand - in a
constructor as an argument to the base class constructor. This is a
code smell. Such code is hard to move elsewhere.
We treat operands as immutable, but actually they are mutable. Every
operand contains mutable fields 'solveOrder', 'ordinalInParent' and
'ordinalInRule' that are assigned in the RelOptRule constructor. We
don't prevent you from passing the same operand to two different rule
instances, but if you did, bad things would happen. Actually I think
people would love to be able to say 'create a rule instance the same
as this, but replacing LogicalFilter with CassandraFilter' but we
don't make it easy, and copying operands from one rule to another is
certainly the wrong way to achieve it.
Given all of these problems, the solution was to convert rules into
data objects. Those data objects are the new Config interface and its
sub-interfaces. (I used an interface because the ImmutableBeans
framework makes it easy to create immutable data objects without
generating code or writing a lot of boilerplate code by hand.)
These Config interfaces mean one extra class per rule, and about 5
extra lines of source code. If you measure simplicity in lines of
code, then I suppose they make things a little less simple.
The new way of building operands, using a builder that makes
call-backs for each nested operand, and has a method for each
attribute of an operand (e.g. predicate()) is also 'less simple' if
you count lines of code. But it is more elegant in that it can
(potentially) produce genuinely immutable operands, is strongly typed,
and offers helpful code-completion when you use it in an IDE.
I hope you now understand the problems I am trying to solve, and why
this new approach is better. That said, this change isn't perfect. We
can evolve this change to make the code more concise, or easier to
understand.
Let's discuss some possible improvements.
One possible improvement would be to stop using rule instances
(RelOptRule) and instead use rule config instances
(RelOptRule.Config). The planner could easily call Config.toRule()
when building a program. The sub-classes of RelOptRule would be
largely invisible (private sub-classes of the Config classes), and
that is appropriate, because all they provide is the implementation of
the onMatch() method. There would be a one-time impact because a lot
of types would change from RelOptRule to RelOptRule.Config.
Another improvement would be to move around the rule constants.
FilterCorrelateRule.INSTANCE and AggregateFilterTransposeRule.INSTANCE
would become the fields FILTER_CORRELATE and
AGGREGATE_FILTER_TRANSPOSE in a new class, LogicalRules. People would
no longer need to reference the names of rule classes in their code,
just the instances. And of course they can now easily customize those
instances.
I hope I have convinced you that rules need to become 'data objects'
with a small amount of behavior, and that people need to be able to
customize them without calling their constructor or even knowing their
precise class.
I would love to hear suggestions for how we can make the transition to
this new model as smooth as possible.
Julian
On Sun, Jun 14, 2020 at 9:15 PM Haisheng Yuan <hy...@apache.org> wrote:
>
> Druid adapter rules are not used by Druid. As far as I know, only Hive uses
> these rules, I even think they should be moved to Hive project. :) If any
> other projects other than Hive are using them, please let us know. Basically,
> these Druid rules are not generally applicable. The PR changed many built-in
> rules to specifically accommodate Druid adapter rules, especially with third
> operand that can match any operator, even EnumerableTableScan is legit.
>
> Most converter rules are not designed to be flexible, no one would extend
> them. RelOptRule#operand() methods are deprecated, which implies someday,
> they will be removed and migrated to the new Config. But most down stream
> projects don't need to worry about the their rules' extensibility and
> compatibility, because they can modify their own rules freely, anytime.
>
> At the end of the day, we may find that only a small fraction of
> transformation rules need refactoring, e.g. ProjectFilterTransposeRule. Note
> that even FilterProjectTransposeRule doesn't need refactoring, we can safely
> remove copyFilter and copyProject, it is safe to always copy them, do we
> really want to match physical operators and generate logical alternatives?
>
> Last but the most important, rule operands, IMO, should be out of Config's
> reach. How frequent do we need to change rule operands or its matching class?
> IMO, operands of rule constructor should remain unchanged, RelFactory and
> Rule description can also remain the same, or adapted by Config silently
> without changing the rule itself, if there are no other additional
> parameters. Other than that, everything can be put into Config. Therefore I
> don't think RelOptNewRule is needed, because RelOptRule should be able to
> integrate with Config seamlessly, without breaking anything.
>
> All in all, I prefer the KISS principle, keep it simple, stupid.
>
> So my opinion is +0.5.
>
> Thanks,
> Haisheng
>
> On 2020/06/14 23:36:21, Stamatis Zampetakis <zabe...@gmail.com> wrote:
> > Hello,
> >
> > Thanks for putting this together Julian.
> >
> > I went quickly over the PR just to grasp better the idea and here are some
> > first thoughts.
> >
> > I really like the fact that rules creation is now much more uniform than
> > before.
> > I also like the fact that in some cases subclassing can be avoided thanks
> > to the flexible configuration. As a fan of the Liskov
> > substitution principle I find it positive to be able to avoid classes such
> > as DruidProjectFilterTransposeRule although to be honest I don't know why
> > these Druid rules impose those additional restrictions.
> >
> > On the other hand, I also feel that the new approach is a bit harder to
> > follow than the previous one. The fact that we have more extension points
> > gives more flexibility but at the same time complicates the implementation
> > a bit. I guess regular committers will not have much trouble to adapt to
> > the changes but for newcomers there may be more questions. For instance:
> > What do we do when we want to customize the behavior of an existing rule?
> > * Use an existing config with different parameters.
> > * Extend the config.
> > * Extend the rule.
> > * Extend the config and the rule.
> > * Create a new rule (if yes under which interface RelOptRule or
> > RelOptNewRule).
> >
> > As Haisheng mentioned, the fact that every rule can be configured with any
> > RelOptRuleOperand is also a point possibly deserving more discussion.
> > Ideally, the API should be able to guide developers to pass the correct
> > configuration parameters and not fail at runtime.
> >
> > Overall, I have mixed feelings about the proposed refactoring (definitely
> > not something blocking), I guess cause I haven't seen as many use-cases as
> > Julian.
> > I'm curious to see what others think about the changes. It's a pity to take
> > a decision based on the feedback of only two/three people.
> >
> > Best,
> > Stamatis
> >
> >
> >
> >
> > On Sun, Jun 14, 2020 at 9:36 PM Haisheng Yuan <hy...@apache.org> wrote:
> >
> > > Hi Julian,
> > >
> > > Thanks for working on this.
> > >
> > > We haven't reached a consensus yet.
> > >
> > > Frankly speaking, I agree with what Stamatis said earlier. Flexibility
> > > doesn't come with no cost. Admittedly, with this patch, any rule
> > > constructor refactoring won't need to deprecate old constructors and break
> > > backward compatibility, however, it also makes the rule definition much
> > > more verbose, less readable and understandable. IMHO, it does more harm
> > > than good.
> > >
> > > Let's see how CassandraFilterRule becomes before and after the change.
> > >
> > > Before this change:
> > >
> > > private static class CassandraFilterRule extends RelOptRule {
> > > private static final CassandraFilterRule INSTANCE = new
> > > CassandraFilterRule();
> > >
> > > private CassandraFilterRule() {
> > > super(operand(LogicalFilter.class, operand(CassandraTableScan.class,
> > > none())),
> > > "CassandraFilterRule");
> > > }
> > > }
> > >
> > > After this change:
> > >
> > > private static class CassandraFilterRule
> > > extends RelOptNewRule<CassandraFilterRule.Config> {
> > > private static final CassandraFilterRule INSTANCE =
> > > Config.EMPTY
> > > .withOperandSupplier(b0 ->
> > > b0.operand(LogicalFilter.class)
> > > .oneInput(b1 -> b1.operand(CassandraTableScan.class)
> > > .noInputs()))
> > > .as(Config.class)
> > > .toRule();
> > >
> > > /** Creates a CassandraFilterRule. */
> > > protected CassandraFilterRule(Config config) {
> > > super(config);
> > > }
> > >
> > > /** Rule configuration. */
> > > public interface Config extends RelOptNewRule.Config {
> > > @Override default CassandraFilterRule toRule() {
> > > return new CassandraFilterRule(this);
> > > }
> > > }
> > > }
> > >
> > >
> > > Intuitively, if we want to check what does the rule generally match or how
> > > it is defined, we just check the constructor. Now we checkout the
> > > constructor, only config is there, go to Config, there is still nothing
> > > interesting, we have to go to the INSTANCE. What is the difference with
> > > just using operand and optionConfig as the rule constructor?
> > >
> > > protected CassandraFilterRule(RelOptRuleOperand operand, Config config)
> > > {
> > > super(operand, config);
> > > }
> > >
> > > Or even simply replace Config with int, with each bit represent an option,
> > > if all of them are boolean options.
> > >
> > > Nothing is more flexible than just using RelOptRuleOperand as the
> > > parameter, just like the base class RelOptRule does. But do we want it?
> > >
> > > At the same time, with the new approach, it is now legit to create the
> > > following instance:
> > >
> > > private static final CassandraFilterRule INSTANCE2 =
> > > Config.EMPTY
> > > .withOperandSupplier(b0 ->
> > > b0.operand(LogicalProject.class) // Even the is intended
> > > to match Filter, but it compiles
> > > .oneInput(b1 -> b1.operand(CassandraTableScan.class)
> > > .noInputs()))
> > > .as(Config.class)
> > > .toRule();
> > >
> > > Before we continue to the discussion and code review, we need to go back
> > > to the original problem, is it a real problem that is facing us? Is there
> > > any real demand or just artificial demand? How about we conduct a Calcite
> > > user survey to see how Calcite devs and users think? Then let's see how to
> > > move forward.
> > >
> > > [+1] Hell yeah, the new approach is awesome, let's go with it.
> > > [+0] Meh, I am OK with current approach, I don't see any burden or problem
> > > with it.
> > > [-1] Hell no, current approach is bad, the new one is even worse.
> > >
> > >
> > > Thanks,
> > > Haisheng
> > >
> > >
> > > On 2020/06/12 23:51:56, Julian Hyde <jh...@apache.org> wrote:
> > > > There is now a PR: https://github.com/apache/calcite/pull/2024. Can
> > > > people please review?
> > > >
> > > > Here's the TL;DR:
> > > >
> > > > Previously, it was not easy to customize, re-use or extend planner
> > > > rules. If you wished to customize a rule (i.e. create a new instance
> > > > of a rule with different properties) you would have to call the rule's
> > > > constructor. Frequently the required constructor did not exist, so we
> > > > would have to add a new constructor and deprecate the old one.
> > > >
> > > > After this change, you start off from an instance of the rule, modify
> > > > its configuration, and call toRule() on the configuration. (Rule
> > > > constructors are now private, because only the configuration ever
> > > > calls them.)
> > > >
> > > > A good illustration of this is DruidRules, which used to contain many
> > > > sub-classes. Those sub-classes are no longer needed. Old code:
> > > >
> > > > public static final DruidSortProjectTransposeRule
> > > SORT_PROJECT_TRANSPOSE =
> > > > new DruidSortProjectTransposeRule(RelFactories.LOGICAL_BUILDER);
> > > >
> > > > public static class DruidSortProjectTransposeRule
> > > > extends SortProjectTransposeRule {
> > > > public DruidSortProjectTransposeRule(RelBuilderFactory
> > > > relBuilderFactory) {
> > > > super(
> > > > operand(Sort.class,
> > > > operand(Project.class, operand(DruidQuery.class,
> > > none()))),
> > > > relBuilderFactory, null);
> > > > }
> > > > }
> > > >
> > > > New code:
> > > >
> > > > public static final SortProjectTransposeRule SORT_PROJECT_TRANSPOSE =
> > > > SortProjectTransposeRule.INSTANCE.config
> > > > .withOperandFor(Sort.class, Project.class, DruidQuery.class)
> > > > .toRule();
> > > >
> > > > The change maintains backwards compatibility to a large degree. In a
> > > > few places, I had to change rule instances from type RelOptRule to
> > > > Supplier<RelOptRule>, to avoid deadlocks during class loading. For
> > > > instance, instead of writing FilterJoinRule.FILTER_ON_JOIN you must
> > > > now write FilterJoinRule.FILTER_ON_JOIN.get().
> > > >
> > > > Julian
> > > >
> > > >
> > > > On Thu, Apr 16, 2020 at 12:08 PM Julian Hyde <jh...@apache.org> wrote:
> > > > >
> > > > > I have now pushed a dev branch with a prototype. Please see
> > > > > https://issues.apache.org/jira/browse/CALCITE-3923 for details.
> > > > >
> > > > > Having built the prototype, I believe that this change is beneficial
> > > > > and we should do it. But I would like to get to consensus on the
> > > > > design before we pull the trigger.
> > > > >
> > > > > Julian
> > > > >
> > > > > On Tue, Apr 14, 2020 at 2:06 PM Julian Hyde <jh...@apache.org> wrote:
> > > > > >
> > > > > > Haisheng,
> > > > > >
> > > > > > I hear you. I agree that major changes to rules will require new
> > > > > > rule
> > > > > > classes (not merely sub-classes). People should copy-paste,
> > > > > > refactor,
> > > > > > and all that good stuff. But I think there are a lot of cases where
> > > we
> > > > > > need to make minor changes to rules (there are many of these in the
> > > > > > code base already), and this change will help.
> > > > > >
> > > > > > I have logged https://issues.apache.org/jira/browse/CALCITE-3923 and
> > > > > > am going to start working on a prototype. When we have a prototype
> > > > > > we
> > > > > > will be able to assess how big an impact the API change will have.
> > > > > > (E.g. whether it will be a breaking change.)
> > > > > >
> > > > > > Julian
> > > > > >
> > > > > > On Sat, Mar 14, 2020 at 8:22 PM Haisheng Yuan <
> > > h.y...@alibaba-inc.com> wrote:
> > > > > > >
> > > > > > > I don't think it is worth the refactoring. People who want to
> > > customize the rule, in most cases, won't be satisfied by a different
> > > parameter, they most likely still need to rewrite (copy & paste) the rule
> > > with some slightly their own logic. For many Calcite users, the rule is
> > > not
> > > reusable even with flexible configurations.
> > > > > > >
> > > > > > > - Haisheng
> > > > > > >
> > > > > > > ------------------------------------------------------------------
> > > > > > > 发件人：Stamatis Zampetakis<zabe...@gmail.com>
> > > > > > > 日 期：2020年03月14日 22:54:04
> > > > > > > 收件人：<dev@calcite.apache.org>
> > > > > > > 主 题：Re: [DISCUSS] Refactor how planner rules are parameterized
> > > > > > >
> > > > > > > Hello,
> > > > > > >
> > > > > > > Apologies for the late reply but I just realised that I had
> > > written the
> > > > > > > mail and never pressed the send button.
> > > > > > >
> > > > > > > I think it is a nice idea and certainly a problem worth
> > > addressing. If I
> > > > > > > understood well you're thinking something like the current
> > > constructor of
> > > > > > > the RelBuilder [1] that accepts a Context parameter. Indeed it
> > > seems that
> > > > > > > with this change even rules that are not designed to be configured
> > > can be
> > > > > > > changed much more gracefully (without adding new constructors and
> > > breaking
> > > > > > > changes).
> > > > > > >
> > > > > > > On the other hand, some of the advantages that you mention can
> > > also be
> > > > > > > turned into disadvantages. For instance, copying a rule without
> > > knowing the
> > > > > > > values of the other parameters is a bit risky and might be harder
> > > to reason
> > > > > > > about its correctness. Moreover, private constructors, final
> > > classes, etc.,
> > > > > > > are primarily used for encapsulation purposes so allowing the
> > > state of the
> > > > > > > rule escape somehow breaks the original design of the rule.
> > > > > > >
> > > > > > > Another problem with respect to rules is cross convention matching
> > > and
> > > > > > > transformations [2]. Many rules should not fire for operands that
> > > are in
> > > > > > > different conventions; a typical example that comes in my mind is
> > > > > > > FilterProjectTransposeRule [3]. In the same spirit most rules
> > > should not
> > > > > > > generate mixed convention transformations. Although a different
> > > problem, I
> > > > > > > am mentioning it here since it could affect the design of the new
> > > API.
> > > > > > >
> > > > > > > Best,
> > > > > > > Stamatis
> > > > > > >
> > > > > > > [1]
> > > > > > >
> > > https://github.com/apache/calcite/blob/f11115a2fe9e360f38910f112288581040e0ced5/core/src/main/java/org/apache/calcite/tools/RelBuilder.java#L155
> > > > > > >
> > > > > > > [2]
> > > > > > >
> > > https://lists.apache.org/thread.html/da1860f99f8bfd6ec7d26626c428ce1c55480e7c61ae7f83060a40c2%40%3Cdev.calcite.apache.org%3E
> > > > > > > [3]
> > > > > > >
> > > https://github.com/apache/calcite/blob/7c27b147414c64505fa33c947100ece094caa15c/core/src/main/java/org/apache/calcite/rel/rules/FilterProjectTransposeRule.java#L57
> > > > > > >
> > > > > > > On Thu, Feb 20, 2020 at 9:20 PM Michael Mior <mm...@apache.org>
> > > wrote:
> > > > > > >
> > > > > > > > This sounds reasonable to me. It also sounds like we could make
> > > this
> > > > > > > > backwards compatible by retaining (but deprecating) the existing
> > > > > > > > constructors and factory methods that will no longer be needed.
> > > > > > > > --
> > > > > > > > Michael Mior
> > > > > > > > mm...@apache.org
> > > > > > > >
> > > > > > > > Le jeu. 20 févr. 2020 à 13:11, Julian Hyde <jh...@apache.org> a
> > > écrit :
> > > > > > > > >
> > > > > > > > > I have an idea for a refactoring to RelOptRule. I haven’t
> > > fully thought
> > > > > > > > it through, but I’m going to sketch it out here to see whether
> > > folks agree
> > > > > > > > about the problems/solutions.
> > > > > > > > >
> > > > > > > > > It will be a breaking change (in the sense that people will
> > > have to
> > > > > > > > change their code in order to get it to compile) but relatively
> > > safe (in
> > > > > > > > that once the code compiles, it will have the same behavior as
> > > before).
> > > > > > > > Also it will give Calcite developers and users a lot more
> > > flexibility going
> > > > > > > > forward.
> > > > > > > > >
> > > > > > > > > The problem I see is that people often want different variants
> > > of
> > > > > > > > planner rules. An example is FilterJoinRule, which has a
> > > 'boolean smart’
> > > > > > > > parameter, a predicate (which returns whether to pull up filter
> > > > > > > > conditions), operands (which determine the precise sub-classes
> > > of RelNode
> > > > > > > > that the rule should match) and a relBuilderFactory (which
> > > controls the
> > > > > > > > type of RelNode created by this rule).
> > > > > > > > >
> > > > > > > > > Suppose you have an instance of FilterJoinRule and you want to
> > > change
> > > > > > > > ‘smart’ from true to false. The ‘smart’ parameter is immutable
> > > (good!) but
> > > > > > > > you can’t easily create a clone of the rule because you don’t
> > > know the
> > > > > > > > values of the other parameters. Your instance might even be
> > > (unbeknownst to
> > > > > > > > you) a sub-class with extra parameters and a private
> > > > > > > > constructor.
> > > > > > > > >
> > > > > > > > > So, my proposal is to put all of the config information of a
> > > RelOptRule
> > > > > > > > into a single ‘config’ parameter that contains all relevant
> > > properties.
> > > > > > > > Each sub-class of RelOptRule would have one constructor with
> > > just a
> > > > > > > > ‘config’ parameter. Each config knows which sub-class of
> > > RelOptRule to
> > > > > > > > create. Therefore it is easy to copy a config, change one or
> > > > > > > > more
> > > > > > > > properties, and create a new rule instance.
> > > > > > > > >
> > > > > > > > > Adding a property to a rule’s config does not require us to
> > > add or
> > > > > > > > deprecate any constructors.
> > > > > > > > >
> > > > > > > > > The operands are part of the config, so if you have a rule
> > > that matches
> > > > > > > > a EnumerableFilter on an EnumerableJoin and you want to make it
> > > match an
> > > > > > > > EnumerableFilter on an EnumerableNestedLoopJoin, you can easily
> > > create one
> > > > > > > > with one changed operand.
> > > > > > > > >
> > > > > > > > > The config is immutable and self-describing, so we can use it
> > > to
> > > > > > > > automatically generate a unique description for each rule
> > > instance.
> > > > > > > > >
> > > > > > > > > Julian
> > > > > > > > >
> > > > > > > > > [1]
> > > > > > > >
> > > https://github.com/apache/calcite/blob/5fa41609cb0fe310a0a11d86319d861423850a36/core/src/main/java/org/apache/calcite/rel/rules/FilterJoinRule.java#L93
> > > > > > > > <
> > > > > > > >
> > > https://github.com/apache/calcite/blob/5fa41609cb0fe310a0a11d86319d861423850a36/core/src/main/java/org/apache/calcite/rel/rules/FilterJoinRule.java#L93
> > > > > > > > >
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > >
> > >
> >
