Hi, Vladimir.
As another topic, it is highly recommended that you split the aggregation
in logical stages, not only for traits related matters. It is true that you
need to annotate the node with different flags or subclasses and it's a
large refactor. But after that, you may find much much bigger benefits.
The most important benefit is aggregation pushing down. For example, the
query:
select t1.value, agg(t2.value) from t1 join t2 on t1.key = t2.key;
You may be able to generate such plan:
PhysicalAggregatePhase(group=t1.value, method=agg_final(t2.value))
Exchange(dist = t1.value)
Join (t1.key = t2.key)
Exchange(dist = t1.key)
scan(t1)
Exchange(dist = t2.key)
PhysicalAggregationPhase(group = t2.key, f_partial(a))
scan(t2)
The pushed "PhysicalAggregationPhase(group = t2.key, f_partial(a))" may be
able to reduce the input data size of the exchange operation dramatically.
There has been lots of research on aggregation push down. But partial
aggregate pushing down could achieve much more benefits:
1. Unlike pushing down a full aggregation, the partial aggregate requires
no extra exchanges. So it could be a pure gain.
2. The pushing down can apply to any aggregation functions, including
user-defined aggregation functions.
3. By introducing the middle phase (the 3-pass aggregation implementation).
Aggregation can be splitted into any number of phases and partial
aggregation can be pushed down through any number of joins, somewhat like:
AggregatePhase(final)
Exchange
AggregatePhase(middle)
JOIN
Exchange
AggregatePhase(middle)
JOIN
Exchange
AggregatePhase(middle)
...
JOIN
Exchange
AggregatePhase(partial)
TableScan
...
Note that AggregatePhase(middle) could work in an adaptive manner: after
processing some data, if it discovers no data reduction, it could
just degenerate to a NOP operation and can be very light weight.
Thanks,
Jinpeng Wu
On Sat, May 29, 2021 at 8:32 AM Haisheng Yuan <hy...@apache.org> wrote:
> > 2) Optimization requests are basically sent to RelSet-s, not RelSubset-s,
> > as we make pairwise comparisons between the requested RelSubset and other
> > subsets in the set [5][6].
>
> I agree with you. There could be some waste when the new delivered /
> required traitset is generated by "passThrough"/ "derive", in which case,
> we only need enforcer between the pair of subsets, instead of pairing with
> all other required / delivered subsets in the RelSet. i.e.
> In the MEMO group, we have 2 required traitsets:
> 1) Hash[a] Sort[b]
> 2) Hash[b] Sort[c]
>
> When we try to pass Hash[a] Sort[b] to one of physical operators say
> Project, we found that we can pass down Hash[a] down to its child, then we
> get a new physical Project with traitset Hash[a], we only need enforcer
> between Hash[a] and Hash[a]Sort[b], but currently in method
> "addConverters", we also generate enforcer between Hash[a] and
> Hash[b]Sort[c], which is not actually what we want.
>
> I think it is definitely worth trying to optimize.
>
> Regards,
> Haisheng Yuan
> On 2021/05/28 19:15:03, Haisheng Yuan <hy...@apache.org> wrote:
> > Hi Vladimir,
> >
> > The top-down optimizer does NOT require implementation rule to generate
> 1 to 1 physical operator for a logical operator, as you can see, if you
> generate a 2 phase physical aggregates for the logical aggregate in the
> implementation rule, it still works. Window is special because we can
> reshuffle the execution order of window functions, and that order makes a
> difference according to different parent physical property request. A
> single converged physical Window operator catered for this speciality.
> However as I said I don't think it is a common scenario.
> >
> > > the whole decision of whether to go with 1-phase or 2-phase
> > > aggregate is a physical decision that should be made based on
> available (or
> > > assumed) input traits.
> > What is the problem of generating both 1-phase and 2-phase aggregates
> and choose the best one based on the cost?
> >
> > Let's see the following query:
> > select a, min(b) from (select * from foo, bar where foo.a=bar.a) t group
> by a;
> > suppose foo is randomly distributed fact table, and bar is randomly
> distributed dimension table.
> > Consider the 2 following plans:
> > 1)
> > PhysicalAggregate
> > +-- HashJoin
> > +-- HashDistribute by a
> > +-- TableScan on foo
> > +-- HashDistribute by a
> > +-- TableScan on bar
> >
> > 2)
> > PhysicalAggregate(global)
> > +-- HashDistribute by a
> > +---- PhysicalAggregate(local)
> > +---- HashJoin
> > +-- TableScan on foo
> > +-- Broadcast
> > +-- TableScan on bar
> >
> > Can you tell that the single phase aggregate plan is always better than
> the 2 phase aggregate plan?
> >
> > > Therefore, the typical way to optimize
> > > LogicalAggregate is to split in the physical phase (implementation
> rule,
> > > pass-through, derive). Practical systems like Dremio [1] and Flink [2]
> > > work this way.
> > Dremio and Flink work this way doesn't mean it is a good way. Greenplum
> Orca and Alibaba MaxCompute optimizer work in another way. In Flink and
> Dremio, they have HashAggPRule to generate 1 phase HashAgg and 2 phase
> HashAgg, SortAggPRule to generate 1 phase SortAgg and 2 phase SortAgg.
> However do you think there is possibility that the global SortAgg combined
> with local HashAgg, or the global HashAgg combined with local SortAgg may
> perform better in difference cases? Are you going to generate all the 4
> combinations in the implementation rule? There are some cases we found we'd
> better to split the aggregate into 3 phase aggregate [1], in which case,
> will the implementation rule generate 3 HashAggs or 3 SortAggs, or all the
> 6 combinations?
> >
> > In our system, we have 1 phase, 2 phase, 3 phase logical aggregate rules
> to transform the LogicalAggregate to another kind of logical aggregate(s)
> with phase info, say LogicalXXXAggregate, then our physical aggregate rules
> match this kind of node to generate HashAgg or StreamAgg. Of course, in the
> logical rules, we can add business logic to guess the possible traits
> delivered by child nodes to determine whether the rule definitely won't
> generate a better alternative and may decide to abort this transformation
> early. But I would rather let the cost model decide.
> >
> > Admittedly, the current top-down optimization is not pure on-demand
> request oriented, because it will always generate a physical request
> regardless the parent nodes' trait request. For example the following query
> in a non-distributed environment:
> > select a, b, c, max(d) from foo group by a, b, c order by a desc;
> >
> > It will first generate a StreamAgg[a ASC, b ASC, c ASC] no matter what
> the parent node requires, then the "passThrough" tells StreamAgg that
> parent requires [a DESC], we get a StreamAgg[a DESC, b ASC, c ASC]. It
> would be ideal if we only generate StreamAgg[a DESC, b ASC, c ASC] by
> request, but I don't think that will make much difference, the bottleneck
> relies on the join order enumeration and the Project related operation.
> >
> > Regards,
> > Haisheng Yuan
> >
> > [1]
> https://github.com/greenplum-db/gporca/blob/master/libgpopt/src/xforms/CXformSplitDQA.cpp
> >
> > On 2021/05/28 09:17:45, Vladimir Ozerov <ppoze...@gmail.com> wrote:
> > > Hi Jinpeng, Haisheng,
> > >
> > > Thank you for your inputs. I really appreciate that. Let me try to
> address
> > > some of your comments and share some experience with the
> implementation of
> > > optimizers for a distributed engine I am currently working with.
> > >
> > > First of all, I would argue that multiple logical operators do not
> have a
> > > 1-1 mapping to physical operators, and Window is not special here. For
> > > instance, LogicalAggregate doesn't have 1-1 mapping to physical
> aggregates
> > > because the physical implementation can be either 1-phase or 2-phase.
> It
> > > doesn't matter that the 2-phase aggregate is a composition of two
> 1-phase
> > > aggregates: the whole decision of whether to go with 1-phase or 2-phase
> > > aggregate is a physical decision that should be made based on
> available (or
> > > assumed) input traits.
> > >
> > > Consider the following logical tree:
> > > LogicalAggregate[group=$0, agg=SUM($1)]
> > > Input
> > >
> > > If I do the split on the logical phase with a separate transformation
> rule,
> > > I will get the following tree:
> > > LogicalAggregate[group=$0, agg=SUM($1)]
> > > LogicalAggregate[group=$0, agg=SUM($1)]
> > > Input
> > >
> > > Now we have an infinite loop because the rule takes one aggregate and
> > > produces two aggregates. To fix that, we may extend the
> LogicalAggregate
> > > with some flag or so. But this (1) potentially breaks other
> LogicalAggregate
> > > optimizations (e.g., transpose with other operators), and (2) breaks
> the
> > > whole idea of the logical operators because the execution phase
> > > (pre-aggregate of final aggregate) is a property of concrete backend,
> not a
> > > property of relational algebra. Therefore, the typical way to optimize
> > > LogicalAggregate is to split in the physical phase (implementation
> rule,
> > > pass-through, derive). Practical systems like Dremio [1] and Flink [2]
> > > work this way.
> > >
> > > That said, as an optimizer developer, I need the flexibility to emit
> any
> > > physical trees for the given logical operator, and 1-1 mapping cannot
> be
> > > assumed. Calcite's API allows for that, and I am not aware of formal
> > > documentation or guidelines that discourage that.
> > >
> > > Now the question when exactly to emit the operators. Normally, we
> produce
> > > operators from rules. As discussed above, if the logical operator may
> > > produce different physical trees depending on input traits, the
> > > recommendation is to emit all combinations, even though we do not know
> > > whether there would be good inputs for that alternatives. This
> contradicts
> > > the idea of the guided top-down search, where we explore the search
> space
> > > in response to a concrete optimization request, rather than with a
> > > pessimistic assumption that a certain plan might be required in the
> future.
> > >
> > > I found a way to mitigate this problem partially. Funny, my solution is
> > > almost similar to what Haisheng proposed for the Window operator.
> > > 1. For every logical operator, I emit a single physical operator from
> the
> > > implementation rule, maintaining the exact 1-1 mapping. The emitted
> > > operators (1) have a special flag "template" which makes their const
> > > infinite, (2) never exposes or demands non-default traits except for
> > > convention, (3) have OMAKASE derivation mode.
> > > 2. When the input is optimized, the "derive" is called on the template,
> > > which produces the concrete physical tree, that is not necessarily 1-1
> to
> > > the original logical node.
> > >
> > > Before rule:
> > > LogicalAggregate[group=$0, agg=SUM($1)]
> > > LogicalInput
> > >
> > > After rule:
> > > PhysicalAggregate[group=$0, agg=SUM($1), template=true, cost=infinite]
> > > LogicalInput
> > >
> > > After "derive" if the input is not shared on $0:
> > > PhysicalAggregate[group=$0, agg=SUM($1)]
> > > Exchange
> > > PhysicalAggregate[group=$0, agg=SUM($1)]
> > > PhysicalInputNotSharded
> > >
> > > After "derive" if the input is shared on $0:
> > > PhysicalAggregate[group=$0, agg=SUM($1)]
> > > PhysicalInputNotSharded
> > >
> > > This approach allows me to avoid the generation of unnecessary
> alternatives
> > > by delaying the optimization to derive phase. The aggregate split is
> > > implemented in rules in Dremio/Flink, but in my case, this logic
> migrates
> > > to "derive".
> > >
> > > This solution worked well for the whole TPC-DS suite until we wanted to
> > > optimize combinations of operators rather than individual operators. A
> good
> > > example is TPC-DS query 1 [3]. During the logical optimization, we get
> the
> > > following logical tree, which is exactly the case that I demonstrated
> at
> > > the beginning of this mail thread:
> > > G1: Aggregate(groupBy=[ctr_store_sk])
> > > G2: Aggregate(groupBy=[ctr_customer_sk, ctr_store_sk]
> > >
> > > And this is where I got stuck. I need to do a simple thing - propagate
> an
> > > optimization request from G1 to G2, informing G2 that it should
> consider
> > > the distribution [ctr_store_sk]. I can deliver that request to my
> physical
> > > template in G2 through "convert". But the problem is that the current
> > > Calcite implementation doesn't allow me to satisfy this request later
> on in
> > > the derivation stage. Instead, I am forced to emit the final execution
> tree
> > > from the "passThrough" method, which will not be notified at the
> derivation
> > > stage. I prepared a scheme [4] that demonstrates the problem.
> > >
> > > It feels that I almost achieved what I need. The last step is to ensure
> > > that "derive" is called on the newly created template. And this is
> where I
> > > think I reach the inflexibility of the current top-down optimizer
> > > implementation. The current design forces us to define all possible
> > > structures of physical operators in advance, but I want to delay the
> > > decision to the derive stage when input traits are known because these
> > > traits are essential to make the proper physical decisions.
> > >
> > > There are some similarities with Haisheng's proposal about the Window
> > > operator. We also maintain the 1-1 correspondence between the logical
> > > operator and a physical template. However, Haisheng's proposal is
> basically
> > > heuristic, as we split optimization into two phases (implementation,
> > > post-processing). It is impossible to properly calculate the cost of
> the
> > > Window operator because we do not know which exchanges would be needed
> > > before the post-processing. In my case, we do the proper cost
> estimation
> > > within a single expanded MEMO.
> > >
> > > Now switching to theoretical considerations. We may make several
> > > observations from the previous discussion:
> > > 1) Our ideas converge to the solution where every logical operator has
> a
> > > single corresponding physical operator, which is later expanded into
> more
> > > alternatives.
> > > 2) Optimization requests are basically sent to RelSet-s, not
> RelSubset-s,
> > > as we make pairwise comparisons between the requested RelSubset and
> other
> > > subsets in the set [5][6].
> > > 3) Irrespective of the design, the complete exploration requires
> multiple
> > > invocations of some implementation logic for different combinations of
> > > required traits and available input traits.
> > >
> > > These observations led me to think that maybe trait propagation through
> > > some dedicated nodes (templates in my case and Haisheng's Window
> proposal,
> > > or pessimistically emitted physical nodes in the previous
> Jinpeng/Haisheng
> > > proposal) is not the ideal design, at least for some cases.
> > >
> > > From the design standpoint, we propagate traits top-down and bottom-up
> > > across equivalence groups, not individual RelSubset-s or RelNode-s.
> > > Currently, we ignore the optimization context when optimizing the group
> > > (except for the cost pruning). Rules emit partially constructed nodes
> since
> > > neither parent requirements nor child traits are available to the rule.
> > >
> > > Instead, there could exist a true guided top-down optimization flow
> when
> > > the "guided" term applies to rules as well:
> > > 1. Pass-through: RelSet receives an optimization request and chooses
> > > appropriate implementation rules to fire. A rule receives optimization
> > > requests, constructs optimization requests for children (adjusting
> traits,
> > > optimization budget, etc.), then sends these requests down. The process
> > > repeated recursively until we either reach the bottom node or some set
> that
> > > is already optimized for this request.
> > > 3. Derive: given the now known input traits, emit appropriate physical
> > > nodes from the rule. Then notify the parent. Repeat the process
> recursively.
> > >
> > > For common use cases, this design would require the same logic, which
> is
> > > currently split between rules, "derive" and "passThrough", just the
> code
> > > location will be different, as everything now converges to the rule.
> But
> > > for the advanced use cases, that approach may allow for more flexible
> > > optimization patterns, like for these two chained aggregates.
> > >
> > > I'll try to implement both solutions - (1) emit multiple nodes from a
> > > physical rule, and (2) enable derivation for some nodes emitted from
> > > "passThrough", and share the results here.
> > >
> > > Regards,
> > > Vladimir.
> > >
> > > [1]
> > >
> https://github.com/dremio/dremio-oss/blob/8e85901e7222c81ccad3436ba9b63485503472ac/sabot/kernel/src/main/java/com/dremio/exec/planner/physical/HashAggPrule.java#L123-L146
> > > [2]
> > >
> https://github.com/apache/flink/blob/release-1.13.0/flink-table/flink-table-planner-blink/src/main/scala/org/apache/flink/table/planner/plan/rules/physical/batch/BatchPhysicalHashAggRule.scala#L101-L147
> > > [3] https://github.com/Agirish/tpcds/blob/master/query1.sql
> > > [4]
> > >
> https://docs.google.com/document/d/1u7V4bNp51OjytKnS2kzD_ikHLiNUPbhjZfRjkTfdsDA/edit
> > > [5]
> > >
> https://github.com/apache/calcite/blob/calcite-1.26.0/core/src/main/java/org/apache/calcite/plan/volcano/RelSet.java#L196
> > > [6]
> > >
> https://github.com/apache/calcite/blob/calcite-1.26.0/core/src/main/java/org/apache/calcite/plan/volcano/TopDownRuleDriver.java#L203
> > >
> > > пт, 28 мая 2021 г. в 00:10, Haisheng Yuan <hy...@apache.org>:
> > >
> > > > Getting back to your window query example:
> > > >
> > > > > Consider the Window function:
> > > > > SELECT
> > > > > AGG1 over (partition by a),
> > > > > AGG2 over (partition by b),
> > > > > AGG3 over (partition by c),
> > > > > ...
> > > > > FROM input
> > > >
> > > > Window is quite special because the logical vs physical operator
> count is
> > > > not 1 to 1, generally we generate a physical window operator for each
> > > > window function with different partition column. That determines
> that once
> > > > the physical operators are created, their order can't be changed.
> Hence
> > > > your proposal of passing required traits to physical rule can
> mitigate the
> > > > problem.
> > > >
> > > > But things would be much easier if we define a different physical
> window
> > > > operator.
> > > > For the above query, we can generate the *Single* physical window
> operator
> > > > like this:
> > > > PhysicalWindow[AGG1 over (partition by a), AGG2 over (partition by
> b),
> > > > AGG3 over (partition by c)]
> > > > or PhysicalWindow(a, b, c) for brevity.
> > > > How do we define the physical properties for it?
> > > > The operator delivers hash distribution on first window partition
> column
> > > > a, but requires its child input to be hash distributed by its last
> window
> > > > partition column c.
> > > >
> > > > If the parent operator request hash distribution on b, or c, the
> window
> > > > operator will be called on "passthrough" method and generate
> > > > PhysicalWindow(b, a, c), or PhysicalWindow(c, a, b). After final
> plan is
> > > > generated, during post processing, we can replace the window
> operator with
> > > > multiple layer nested window operators, and insert Exchange
> operators if
> > > > necessary. But frankly speaking, I haven't seen any use cases of
> this kind
> > > > in production.
> > > >
> > > > Regarding the rule alternative you proposed;
> > > > > class PhysicalAggregateRule extends PhysicalRule {
> > > > > void onMatch(RelOptRuleCall call, *RelTraitSet requiredTraits*)
> {...
> > > >
> > > > Consider the following plan:
> > > > InnerJoin (on a)
> > > > +-- Agg (on b)
> > > > +-- Scan
> > > >
> > > > For the inner join, we can generate sort merge join and hash join.
> > > > The sort merge join can request the following traits to Agg:
> > > > 1) Singleton
> > > > 2) hash distribution on a, sorted by a
> > > > The hash join can request the following traits to Agg:
> > > > 1) Singleton
> > > > 2) hash distribution on a
> > > > 3) any distribution
> > > > 4) broadcast distribution
> > > >
> > > > The PhysicalAggregateRule will be called and executed 5 times, while
> > > > generating the same physical aggregate candidates, unless we pass a
> whole
> > > > list of required traits to the physical rule, which I have
> prototyped some
> > > > time ago with the exact idea.
> > > >
> > > > Regards,
> > > > Haisheng Yuan
> > > >
> > > > On 2021/05/27 17:44:23, Haisheng Yuan <hy...@apache.org> wrote:
> > > > > > In distributed systems, an implementation rule may produce
> different
> > > > > > physical operators depending on the input traits. Examples are
> > > > Aggregate,
> > > > > > Sort, Window.
> > > > >
> > > > > No, in most cases, physical operators are generated regardless the
> > > > input, because the input traits are not know yet. Window might be an
> > > > exception.
> > > > >
> > > > > > Since input traits are not known when the rule is fired, we
> must
> > > > > > generate *all possible combinations* of physical operators
> that we
> > > > may
> > > > > > need. For LogicalAggregate, we must generate 1-phase and
> 2-phase
> > > > > > alternatives. For LogicalSort, we also have 1-phase and
> 2-phase
> > > > > > alternatives. Etc.
> > > > >
> > > > > IMHO, 1 phase and 2 phase are just different logical alternatives,
> that
> > > > is also why I call it a logical rule to split the aggregate into a 2
> phase
> > > > aggregate. But HashAggregate and StreamAggregate are indeed the
> different
> > > > physical alternatives for a LogicalAggregate.
> > > > >
> > > > >
> > > > > > Unlike Aggregate or Sort, which may have only 1 or 2 phases,
> certain
> > > > > > logical operators may have many physical alternatives.
> Consider the
> > > > Window
> > > > > > function:......
> > > > >
> > > > > In window implementation rule, when building physical operator for
> > > > Window that has multiple window functions but with different
> partition
> > > > columns, we can infer the possible traits that can be delivered by
> input
> > > > operators by creating your own RelMetaData, hence multiple window
> > > > combination with certain order, but not exhausted enumeration. In
> fact, the
> > > > window ordering problem exists in every different kind of optimizer.
> > > > >
> > > > > > As input traits are not known when the rule is fired, the nodes
> emitted
> > > > > > from the implementation rules most likely would not be used in
> the
> > > > final
> > > > > > plan.
> > > > >
> > > > > That is quite normal, any operator generated by implementation rule
> > > > might not be used in the final plan, because there may be tens of
> thousands
> > > > of alternatives, we only choose the one with lowest cost.
> > > > >
> > > > > > For example, I can create a physical aggregate that demands
> > > > > > non-strict distribution {a,b} from its input, meaning that both
> [a,b]
> > > > and
> > > > > > [b,a] is ok. However, in the final plan, we are obligated to
> have a
> > > > strict
> > > > > > distribution - either [a, b] in that order, or [b, a] in that
> order -
> > > > > > otherwise, physical operators like Join and Union will not work.
> > > > >
> > > > > It depends on your own satisfaction model and how do you coordinate
> > > > property requirement among child operators. Unlike Orca optimizer,
> where
> > > > there is exact match, partial satisfying, orderless match etc,
> Calcite's
> > > > default implementation always require exact satisfying. But we can
> still
> > > > make use of "passThrough" and "derive" to achieve our goal. i.e. the
> > > > aggregate generated by implementation rule requires itself and its
> child to
> > > > delivered distribution on [a,b], but the "derive" method tells
> Aggregate
> > > > that [b,a] is available, it can generate another option to require
> [b,a]
> > > > instead.
> > > > >
> > > > > > In distributed engines, the nodes emitted from rules are
> basically
> > > > "templates"
> > > > > > that must be replaced with normal nodes.
> > > > >
> > > > > There is no difference between distributed and non-distributed
> engines
> > > > when dealing with this. In Orca and CockroachDB optimizer, the nodes
> > > > emitted from rules are operators without physical properties, the
> optimizer
> > > > then request physical properties in top-down manner, either
> recursively or
> > > > stack, or state machine. Calcite is quite different. when the
> physical
> > > > operator is generated by implementation rule, the physical operator
> must
> > > > has its own traits, at the same time, the traits that it expects its
> child
> > > > operators to deliver. So in Calcite, they are not "templates". The
> > > > difference is there since Calcite's inception.
> > > > >
> > > > > Regards,
> > > > > Haisheng Yuan
> > > > >
> > > > > On 2021/05/27 08:59:33, Vladimir Ozerov <ppoze...@gmail.com>
> wrote:
> > > > > > Hi Haisheng,
> > > > > >
> > > > > > Thank you for your inputs. They are really helpful. Let me
> summarize
> > > > your
> > > > > > feedback in my own words to verify that I understand it
> correctly.
> > > > > >
> > > > > > 1. In distributed systems, an implementation rule may produce
> > > > different
> > > > > > physical operators depending on the input traits. Examples are
> > > > Aggregate,
> > > > > > Sort, Window.
> > > > > > 2. Since input traits are not known when the rule is fired,
> we must
> > > > > > generate *all possible combinations* of physical operators
> that we
> > > > may
> > > > > > need. For LogicalAggregate, we must generate 1-phase and
> 2-phase
> > > > > > alternatives. For LogicalSort, we also have 1-phase and
> 2-phase
> > > > > > alternatives. Etc.
> > > > > > 3. If all combinations are generated, it is expected that
> > > > "passThrough"
> > > > > > and "derive" would be just trivial replacements of traits for
> most
> > > > cases.
> > > > > > This is why "passThroughTraits" and "deriveTraits" are
> recommended.
> > > > A
> > > > > > notable exception is TableScan that may emit alternative
> indexes in
> > > > > > response to the pass-through requests.
> > > > > >
> > > > > > If my understanding is correct, then there are several issues
> with this
> > > > > > approach still.
> > > > > >
> > > > > > 1. Unlike Aggregate or Sort, which may have only 1 or 2 phases,
> certain
> > > > > > logical operators may have many physical alternatives. Consider
> the
> > > > Window
> > > > > > function:
> > > > > > SELECT
> > > > > > AGG1 over (partition by a),
> > > > > > AGG2 over (partition by b),
> > > > > > AGG3 over (partition by c),
> > > > > > ...
> > > > > > FROM input
> > > > > >
> > > > > > To calculate each aggregate, we need to re-shuffle the input
> based on
> > > > the
> > > > > > partition key. The key question is the order of reshuffling. If
> the
> > > > input
> > > > > > is shared by [a], I want to calculate AGG1 locally and then
> re-shuffle
> > > > the
> > > > > > input to calculate other aggregates. For the remaining AGG2 and
> AGG3,
> > > > the
> > > > > > order is also important. If the parent demands sharding by [b],
> then
> > > > the
> > > > > > proper sequence is b-c-a:
> > > > > > 1: Window[AGG2 over (partition by b)] // SHARDED[b]
> > > > > > 2: Window[AGG3 over (partition by c)] // SHARDED[c]
> > > > > > 3: Window[AGG1 over (partition by a)] // SHARDED[a]
> > > > > > 4: Input // SHARDED[a]
> > > > > >
> > > > > > But if the parent demands [c], the proper sequence is c-b-a.
> Since we
> > > > do
> > > > > > not know real distributions when the rule is fired, we must emit
> all
> > > > the
> > > > > > permutations to ensure that no optimization opportunity is
> missed. But
> > > > with
> > > > > > complex window aggregate, this might be impractical because we
> will
> > > > emit
> > > > > > lots of unnecessary nodes.
> > > > > >
> > > > > > 2. As input traits are not known when the rule is fired, the
> nodes
> > > > emitted
> > > > > > from the implementation rules most likely would not be used in
> the
> > > > final
> > > > > > plan. For example, I can create a physical aggregate that demands
> > > > > > non-strict distribution {a,b} from its input, meaning that both
> [a,b]
> > > > and
> > > > > > [b,a] is ok. However, in the final plan, we are obligated to
> have a
> > > > strict
> > > > > > distribution - either [a, b] in that order, or [b, a] in that
> order -
> > > > > > otherwise, physical operators like Join and Union will not work.
> In
> > > > > > distributed engines, the nodes emitted from rules are basically
> > > > "templates"
> > > > > > that must be replaced with normal nodes.
> > > > > >
> > > > > > Does this reasoning make any sense? If yes, it means that the
> current
> > > > > > approach forces us to produce many unnecessary nodes to explore
> the
> > > > full
> > > > > > search space. The question is whether alternative approaches
> could
> > > > better
> > > > > > fit the requirements of the distributed engine? This is a purely
> > > > > > theoretical question. I am currently looking deeper at
> CockroachDB.
> > > > They
> > > > > > have very different architecture: no separation between logical
> and
> > > > > > physical nodes, physical properties are completely decoupled from
> > > > nodes,
> > > > > > usage of recursion instead of the stack, etc.
> > > > > >
> > > > > > Regards,
> > > > > > Vladimir.
> > > > > >
> > > > > > чт, 27 мая 2021 г. в 03:19, Haisheng Yuan <hy...@apache.org>:
> > > > > >
> > > > > > > Another point I would like to mention is that it is not
> recommended
> > > > to
> > > > > > > override method "passThrough" and "derive" directly, override
> > > > > > > "passThroughTraits" and "deriveTraits" instead, so that we can
> make
> > > > sure
> > > > > > > only the same type of physical node is created and no nested
> > > > relnodes or
> > > > > > > additional RelSets are created, unless you know you have to
> create
> > > > > > > different type of nodes. For example, if the table foo has an
> btree
> > > > index
> > > > > > > on column a, and the parent relnode is requesting ordering on
> column
> > > > a,
> > > > > > > then we may consider to override "passThrough" of TableScan to
> > > > return an
> > > > > > > IndexScan instead of a TableScan.
> > > > > > >
> > > > > > > Regards,
> > > > > > > Haisheng Yuan
> > > > > > > On 2021/05/26 22:45:20, Haisheng Yuan <hy...@apache.org>
> wrote:
> > > > > > > > Hi Vladimir,
> > > > > > > >
> > > > > > > > 1. You need a logical rule to split the aggregate into a
> local
> > > > aggregate
> > > > > > > and global aggregate, for example:
> > > > > > > >
> > > > > > >
> > > >
> https://github.com/greenplum-db/gporca/blob/master/libgpopt/src/xforms/CXformSplitGbAgg.cpp
> > > > > > > > Only implementation rules can convert a logical node to a
> physical
> > > > node
> > > > > > > or multiple physical nodes.
> > > > > > > > After physical implementation, you have 2 physical
> alternatives:
> > > > > > > > 1) single phase global physical aggregate,
> > > > > > > > 2) 2 phase physical aggregate with local and global
> aggregate.
> > > > > > > > It should be up to the cost to decide which one to choose.
> > > > > > > >
> > > > > > > > 2. Given a desired traitset from parent node, the current
> relnode
> > > > only
> > > > > > > needs to generate a single relnode after passing down the
> traitset.
> > > > Given a
> > > > > > > traitset delivered by child node, the current relnode only
> derive a
> > > > single
> > > > > > > relnode. Quite unlike other optimizer, in Calcite's top-down
> > > > optimizer, you
> > > > > > > don't need to worry about issuing multiple optimization
> requests to
> > > > inputs,
> > > > > > > which is handled by Calcite framework secretly. i.e.
> > > > > > > > SELECT a, b, min(c) from foo group by a, b;
> > > > > > > > In many other optimizer, we probably need ask the aggregate
> to
> > > > issue 3
> > > > > > > distribution requests for tablescan on foo, which are
> > > > > > > > 1) hash distributed by a,
> > > > > > > > 2) hash distributed by b,
> > > > > > > > 3) hash distributed by a, b
> > > > > > > > However in Calcite top-down optimizer, your physical
> > > > implementation rule
> > > > > > > for global aggregate only need generate a single physical node
> with
> > > > hash
> > > > > > > distribution by a, b. In case the table foo happens to be
> > > > distributed by a,
> > > > > > > or b, the derive() method will tell you there is an
> opportunity.
> > > > This is
> > > > > > > the feature that Calcite's top-down optimizer excels over other
> > > > optimizers,
> > > > > > > because this can dramatically reduce the search space while
> keeping
> > > > the
> > > > > > > optimal optimization opportunity.
> > > > > > > >
> > > > > > > > 3. This is by design. Nodes produced from "passThrough" and
> > > > "derive" and
> > > > > > > just sibling physical node with different traitset, we only
> need the
> > > > > > > initial physical nodes after implementation to avoid
> unnecessary
> > > > > > > operations. The fundamental reason is, unlike Orca optimizer
> where
> > > > physical
> > > > > > > node and physical property are separate things, Calcite's
> > > > logical/physical
> > > > > > > nodes contains traitset. With regard to the latter question,
> can you
> > > > give
> > > > > > > an example?
> > > > > > > >
> > > > > > > > Regards,
> > > > > > > > Haisheng Yuan
> > > > > > > >
> > > > > > > >
> > > > > > > > On 2021/05/26 20:11:57, Vladimir Ozerov <ppoze...@gmail.com>
> > > > wrote:
> > > > > > > > > Hi,
> > > > > > > > >
> > > > > > > > > I tried to optimize a certain combination of operators for
> the
> > > > > > > distributed
> > > > > > > > > engine and got stuck with the trait propagation in the
> top-down
> > > > > > > engine. I
> > > > > > > > > want to ask the community for advice on whether the
> problem is
> > > > solvable
> > > > > > > > > with the current Apache Calcite implementation or not.
> > > > > > > > >
> > > > > > > > > Consider the following logical tree:
> > > > > > > > > 3: LogicalAggregate[group=[a], F2(c)]
> > > > > > > > > 2: LogicalAggregate[group=[a,b], F1(c)]
> > > > > > > > > 1: LogicalScan[t]
> > > > > > > > >
> > > > > > > > > Consider that these two aggregates cannot be merged or
> > > > simplified for
> > > > > > > > > whatever reason. We have only a set of physical rules to
> > > > translate this
> > > > > > > > > logical tree to a physical tree. Also, there could be any
> number
> > > > of
> > > > > > > > > other operators between these two aggregates. We omit them
> for
> > > > clarity,
> > > > > > > > > assuming that the distribution is not destroyed.
> > > > > > > > >
> > > > > > > > > In the distributed environment, non-collocated aggregates
> are
> > > > often
> > > > > > > > > implemented in two phases: local pre-aggregation and final
> > > > aggregation,
> > > > > > > > > with an exchange in between. Consider that the Scan
> operator is
> > > > hash
> > > > > > > > > distributed by some key other than [a] or [b]. If we
> optimize
> > > > operators
> > > > > > > > > without considering the whole plan, we may optimize each
> operator
> > > > > > > > > independently, which would give us the following plan:
> > > > > > > > > 3: PhysicalAggregate[group=[a], F2_phase2(c)]
> //
> > > > > > > > > HASH_DISTRIBUTED [a]
> > > > > > > > > 3: Exchange[a]
> //
> > > > > > > > > HASH_DISTRIBUTED [a]
> > > > > > > > > 3: PhysicalAggregate[group=[a], F2_phase1(c)]
> //
> > > > > > > > > HASH_DISTRIBUTED [a,b]
> > > > > > > > > 2: PhysicalAggregate[group=[a,b], F1_phase2(c)]
> //
> > > > > > > > > HASH_DISTRIBUTED [a,b]
> > > > > > > > > 2: Exchange[a, b]
> //
> > > > > > > > > HASH_DISTRIBUTED [a,b]
> > > > > > > > > 2: PhysicalAggregate[group=[a,b], F1_phase1(c)]
> //
> > > > > > > > > HASH_DISTRIBUTED [d]
> > > > > > > > > 1: PhysicalScan[t]
> //
> > > > > > > > > HASH_DISTRIBUTED [d]
> > > > > > > > >
> > > > > > > > > This plan is not optimal, because we re-hash inputs twice.
> A
> > > > better
> > > > > > > plan
> > > > > > > > > that we want to get:
> > > > > > > > > 3: PhysicalAggregate[group=[a], F2(c)] //
> > > > > > > HASH_DISTRIBUTED
> > > > > > > > > [a]
> > > > > > > > > 2: PhysicalAggregate[group=[a,b], F1_phase2(c)] //
> > > > > > > HASH_DISTRIBUTED
> > > > > > > > > [a]
> > > > > > > > > 2: Exchange[a] //
> > > > > > > HASH_DISTRIBUTED
> > > > > > > > > [a]
> > > > > > > > > 2: PhysicalAggregate[group=[a,b], F1_phase1(c)] //
> > > > > > > HASH_DISTRIBUTED
> > > > > > > > > [d]
> > > > > > > > > 1: PhysicalScan[t] //
> > > > > > > HASH_DISTRIBUTED
> > > > > > > > > [d]
> > > > > > > > >
> > > > > > > > > In this case, we take advantage of the fact that the
> > > > distribution [a]
> > > > > > > is
> > > > > > > > > compatible with [a,b]. Therefore we may enforce only [a],
> > > > instead of
> > > > > > > doing
> > > > > > > > > [a,b] and then [a]. Since exchange operators are very
> expensive,
> > > > this
> > > > > > > > > optimization may bring a significant boost to the query
> engine.
> > > > Now the
> > > > > > > > > question - how do we reach that state? Intuitively, a
> > > > pass-through is
> > > > > > > > > exactly what we need. We may pass the optimization request
> from
> > > > top
> > > > > > > > > aggregate to bottom aggregate to find physical
> implementations
> > > > shared
> > > > > > > by
> > > > > > > > > [a]. But the devil is in the details - when and how
> exactly to
> > > > pass
> > > > > > > this
> > > > > > > > > request?
> > > > > > > > >
> > > > > > > > > Typically, we have a conversion rule that converts a
> logical
> > > > aggregate
> > > > > > > to a
> > > > > > > > > physical aggregate. We may invoke "convert" on the input to
> > > > initiate
> > > > > > > the
> > > > > > > > > pass-through:
> > > > > > > > >
> > > > > > > > > RelNode convert(...) {
> > > > > > > > > return new PhysicalAggregate(
> > > > > > > > > convert(input, HASH_DISTRIBUTED[a])
> > > > > > > > > )
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > > The first problem - we cannot create the normal physical
> > > > aggregate here
> > > > > > > > > because we do not know input traits yet. The final decision
> > > > whether to
> > > > > > > do a
> > > > > > > > > one-phase or two-phase aggregate can be made only in the
> > > > > > > > > "PhysicalNode.derive" method when concrete input traits are
> > > > resolved.
> > > > > > > > > Therefore the converter rule should create a kind of
> "template"
> > > > > > > physical
> > > > > > > > > operator, which would be used to construct the final
> operator(s)
> > > > when
> > > > > > > input
> > > > > > > > > traits are resolved. AFAIU Enumerable works similarly: we
> create
> > > > > > > operators
> > > > > > > > > with virtually arbitrary traits taken from logical nodes
> in the
> > > > > > > conversion
> > > > > > > > > rules. We only later do create normal nodes in the derive()
> > > > methods.
> > > > > > > > >
> > > > > > > > > The second problem - our top aggregate doesn't actually
> need the
> > > > > > > > > HASH_DISTRIBUTED[a] input. Instead, it may accept inputs
> with any
> > > > > > > > > distribution. What we really need is to inform the input
> (bottom
> > > > > > > aggregate)
> > > > > > > > > that it should look for additional implementations that
> satisfy
> > > > > > > > > HASH_DISTRIBUTED[a]. Therefore, enforcing a specific
> > > > distribution on
> > > > > > > the
> > > > > > > > > input using the "convert" method is not what we need
> because this
> > > > > > > > > conversion might enforce unnecessary exchanges.
> > > > > > > > >
> > > > > > > > > The third problem - derivation. Consider that we delivered
> the
> > > > > > > optimization
> > > > > > > > > request to the bottom aggregate. As an implementor, what
> am I
> > > > supposed
> > > > > > > to
> > > > > > > > > do in this method? I cannot return the final aggregate
> from here
> > > > > > > because
> > > > > > > > > the real input traits are not derived yet. Therefore, I
> can only
> > > > return
> > > > > > > > > another template, hoping that the "derive" method will be
> called
> > > > on it.
> > > > > > > > > However, this will not happen because trait derivation is
> > > > skipped on
> > > > > > > the
> > > > > > > > > nodes emitted from pass-through. See "DeriveTrait.perform"
> [1].
> > > > > > > > >
> > > > > > > > > BottomAggregate {
> > > > > > > > > RelNode passThrough(distribution=HASH_DISTRIBUTED[a]) {
> > > > > > > > > // ???
> > > > > > > > > }
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > > I feel that I am either going in the wrong direction, or
> some
> > > > gaps in
> > > > > > > the
> > > > > > > > > product disallow such optimization. So I would like to ask
> the
> > > > > > > community to
> > > > > > > > > assist with the following questions:
> > > > > > > > > 1. In the top-down optimizer, how should we convert a
> logical
> > > > node to a
> > > > > > > > > physical node, provided that "derive" is not called yet? I
> have
> > > > a gut
> > > > > > > > > feeling that the trait propagation is currently not
> implemented
> > > > to the
> > > > > > > full
> > > > > > > > > extent because based on Cascades paper I would expect that
> parent
> > > > > > > physical
> > > > > > > > > nodes are produced after the child physical nodes. But in
> our
> > > > rules,
> > > > > > > this
> > > > > > > > > is not the case - some physical nodes are produced before
> the
> > > > trait
> > > > > > > > > derivation.
> > > > > > > > > 2. How to propagate several optimization requests to
> inputs? We
> > > > need
> > > > > > > either
> > > > > > > > > inputs with a specific distribution or inputs with an
> arbitrary
> > > > > > > > > distribution in the example above. It seems that to achieve
> > > > that, I
> > > > > > > need to
> > > > > > > > > emit several alternative nodes with different requirements
> to
> > > > inputs.
> > > > > > > Does
> > > > > > > > > it make sense?
> > > > > > > > > 3. Why are nodes produced from the "passThrough" method
> excluded
> > > > from
> > > > > > > trait
> > > > > > > > > derivation? If this is by design, how can I preserve the
> > > > optimization
> > > > > > > > > request to satisfy it on the derivation stage when input
> traits
> > > > are
> > > > > > > > > available?
> > > > > > > > >
> > > > > > > > > Regards,
> > > > > > > > > Vladimir.
> > > > > > > > >
> > > > > > > > > [1]
> > > > > > > > >
> > > > > > >
> > > >
> https://github.com/apache/calcite/blob/calcite-1.26.0/core/src/main/java/org/apache/calcite/plan/volcano/TopDownRuleDriver.java#L828
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>
