Hi, Haisheng and Igor. I think we do need the ability for logical space pruning. But we can achieve it step by step. In the first trial, we implement and optimize a rel only after it is fully explored. And then, after solving problems like group sharing stats and logical cost accuracy, we move it forward. Haisheng's current design does not prevent future refactoring.
I believe there are ways to calculate the logical cost, but I think it's not that simple as "cardinality * unit_copy_cost.". The "current best cost" is computed by the actual cost model from physical plan, and it is not comparable to such a simple cost model. To achieve this goal, it may require some modification to the RelNode interface. For the discussion mentioned in last part of Haisheng's proposal, I am more into the second way: Add a new XXXXPlanner. Here is my opinions: 1. This proposal is a long term work item. It may take months or even years before it is ready for production. The product I am working on is serving millions of queries every day. Sometimes a simple change may cause big problems. I cannot afford the risk and testing efforts if there are always so many big changes every release. 2. This work is more an exploration than a step by step work. Sometimes you may prefer an implementation and after moving forward to the next goal you find another design may be better. It may contain refactoring on refactoring. Modifying the VolcanoPlanner directly not only introduces more risk but also makes it more difficult to design as you need to keep backward compatible always in mind. With a new planner, we can try things wilder. We can even commit some useless staged code to help cooperations. 3. Though it is a new planner, it still satisfy the Planner interface. And I see not difficulties switching between VolcanoPlanner and the new one. Actually a simple factory that creates planner instances by config is the only effort it needs. This applies to test cases, too. So all the tests are sharable. And this feature, switching between old and new implementation with one config, is very helpful for online service publishing. 4. The new planner can derive from the VolcanoPlanner and share the same data structure. So it does not introduce more coding efforts than modifying the VolcanoPlanner directly. And, with such derivation, after some achievements are considered stable and helpful, it is also possible to apply the changes to the VolcanoPlanner directly to help improving the old planner. Xiening Dai <xndai....@gmail.com> 于2020年4月20日周一 下午12:02写道： > Hi Igor, > > Your comment - "because actual cost may be calculated correctly using > physical operators only. So won't be able to implement Branch and Bound > Space Pruning.“ is actually not true. In Cascade’s lower bound / upper > bound pruning algorithm, you can get cost lower bound of input RelNode > using cardinality * unit_copy_cost. The unit_copy_cost is a constant, which > stands for the minimal cost for processing a tuple from the input. So this > will be the minimal cost the input RelNode can achieve, and if this is > indeed larger than the current best cost, this input node can be pruned. > > In this proposal, we are not trying to split logical and physical planning > entirely. But for any given equivalent set, we would need to finish > exploration before implementation. But for the entire memo tree, each set > could be in different planning stage, and an equivalent set can be pruned > even before it’s implemented. A text book example of aforementioned “lower > bound / upper bound pruning” would be the join ordering case. > > Regarding #3, I think we can still achieve that (partially) through this > proposal. Remember every time when we start the optimization, we pass down > an upper bound limit. Initially this upper bound for root is infinite - > meaning that no feasible plan is available. Every time when we find a > physical plan we update this upper bound, then start the search again. We > could stop the search when the cost is less than a pre-defined threshold - > which gives you a “good enough” plan with early termination. Still this > wouldn't avoid the logical exploration. For that, you would probably > archive through rule configurations, and avoid some expansive > transformation to keep the cost down. > > > > On Apr 19, 2020, at 7:30 PM, Haisheng Yuan <hy...@apache.org> wrote: > > > > Igor, > > > > a) Given current Calcite's stats derivation strategy, mixing logical and > physical planning won't make it better. I hope you went through my email to > the end, currently operators inside a memo group don't share stats info, > each operator's stats may differ with the other one, and the RelSubset's > best node and stats may vary during optimization. So logical and physical > rule pruning are not safe at the moment, otherwise it almost implies > changing downstream project's cost model silently. > > > > On the other hand, ensuring child group exploration task finish first > will make rule mutual exclusivity check possible, like the result of > ReduceExpressionRule won't need trigger the same rule again, The join > result of JoinCommuteRule won't need trigger JoinCommuteRule and > ReduceExpressionRule again. > > > > More importantly, most if not all the the long planning queries in > Calcite are not caused by too many alternatives, but mutual triggering, or > cyclic triggering, which can be avoided by customizing the rules instead of > using the default one. Unless you use dynamic programming (Calcite use > heuristic) on join reordering (left-deep, right-deep, bushy), space pruning > won't help make long / infinite running query faster. > > > > b) No evidence shows current version of Calcite will return the most > promising plan in first planning iteration. Instead of praying for getting > good enough plan in the first iteration, why not focus on fixing rules that > causes the issue? > > > > c) That is not the goal. > > > > On 2020/04/19 15:14:57, Seliverstov Igor <gvvinbl...@gmail.com> wrote: > >> Haisheng, > >> > >> From my point of view splitting logical and physical planning parts > isn’t a good idea. > >> > >> I think so because actual cost may be calculated correctly using > physical operators only. So that we: > >> a) won't be able to implement Branch and Bound Space Pruning (as far as > I understand, at exploring time there are no physical operators, no correct > costs, but assumptions only, I don’t think we should rely on them) > >> b) won’t be able to get good enough plan (without Branch and Bound > Space Pruning it’s non-trivial task to get right search direction and most > promising plans in first planning iterations) > >> c) won’t be able to tune the good enough plan during several similar > queries are executed > >> > >> Regards, > >> Igor > >> > >> > >>> 19 апр. 2020 г., в 17:37, Haisheng Yuan <hy...@apache.org> написал(а): > >>> > >>> Hi Igor, > >>> > >>> You can't have your cake and eat it. > >>> But one feasible work item definitely we can do is that once timeout, > stop exploring, use the first available physical operator in each group and > optimize it. > >>> > >>> Because most, if not all, of the long / infinite running optimizations > are caused by project related transpose, join reordering (join commute + > associative), constant reduction for large expression (see CALCITE-3460), > all of which are logical transformations rules and many of which have > corresponding JIRAs. So another alternative is, let's fix these bugs to > improve Calcite to make timeout option less usable. > >>> > >>> Another thing worth noting is that sql server optimizer timeout > mechanism is not based on clock time, but the number of optimization tasks > it has done [1]. > >>> > >>> [1] > https://techcommunity.microsoft.com/t5/sql-server-support/understanding-optimizer-timeout-and-how-complex-queries-can-be/ba-p/319188 > >>> > >>> Regards, > >>> Haisheng > >>> > >>> On 2020/04/19 11:31:27, Seliverstov Igor <gvvinbl...@gmail.com> > wrote: > >>>> Haisheng, > >>>> > >>>> Ok, then such notification isn't needed > >>>> > >>>> But in this case we don't have any control over how long planning > takes > >>>> > >>>> For some systems it's necessary to get good enough plan right now > instead > >>>> of best one after while > >>>> > >>>> For example we've been considering a case when a query is optimised > several > >>>> times in short iterations in case it's impossible to get the best > plan in > >>>> reasonable period of time (for example there is an SLA for response > time) > >>>> > >>>> This mean we need all needed physical implementations after each > logical > >>>> transformation is applied. > >>>> > >>>> Regards, > >>>> Igor > >>>> > >>>> > >>>> вс, 19 апр. 2020 г., 13:55 Haisheng Yuan <hy...@apache.org>: > >>>> > >>>>> Hi Igor, > >>>>> > >>>>> There will be no rule queue anymore. Y will be fully explored > (logical > >>>>> rules are matched and applied) before it can be implemented and > optimized. > >>>>> > >>>>> Thanks, > >>>>> Haisheng > >>>>> > >>>>> On 2020/04/19 10:11:45, Seliverstov Igor <gvvinbl...@gmail.com> > wrote: > >>>>>> Hi Haisheng, > >>>>>> > >>>>>> Great explanation, thanks. > >>>>>> > >>>>>> One thing I'd like to cover in advance is trait propagation process > (I > >>>>> need > >>>>>> it for Apache Ignite SQL engine implementation). > >>>>>> > >>>>>> For example: > >>>>>> > >>>>>> There are two nodes: Rel X and its child node Rel Y > >>>>>> > >>>>>> Both nodes are in Optimized state, and there is a Logical rule for > Rel Y > >>>>> in > >>>>>> a rules queue matched, > >>>>>> > >>>>>> After logical rule is applied, there is a logical rel Rel Y' and > >>>>> containing > >>>>>> it set moves into Exploring state (since there is a logical node > which > >>>>> has > >>>>>> to be implemented) > >>>>>> > >>>>>> After whole process Exploring -> ... -> Optimized we need to force > parent > >>>>>> Rel X set to switch its state from Optimized to Optimizing to peek > the > >>>>>> newly implemented child and (possible) produce a new Rel X' > physical rel > >>>>> on > >>>>>> the basis of previously requested from the Rel X set traits. > >>>>>> > >>>>>> If a new Rel X' is produced, a Rel X' parent should move its state > >>>>>> Optimized -> Optimizing and repeat described above operations. > >>>>>> > >>>>>> Does it look like true? > >>>>>> > >>>>>> Regards, > >>>>>> Igor > >>>>>> > >>>>>> > >>>>>> вс, 19 апр. 2020 г., 6:52 Haisheng Yuan <h.y...@alibaba-inc.com>: > >>>>>> > >>>>>>> Hi, > >>>>>>> > >>>>>>> In the past few months, we have discussed a lot about Cascades > style > >>>>>>> top-down optimization, including on-demand trait > derivation/request, > >>>>> rule > >>>>>>> apply, branch and bound space pruning. Now we think it is time to > move > >>>>>>> towards these targets. > >>>>>>> > >>>>>>> We will separate it into several small issues, and each one can be > >>>>>>> integrated as a standalone, independent feature, and most > importantly, > >>>>>>> meanwhile keep backward compatibility. > >>>>>>> > >>>>>>> 1. Top-down trait request > >>>>>>> In other words, pass traits requirements from parent nodes to child > >>>>> nodes. > >>>>>>> The trait requests happens after all the logical transformation > rules > >>>>> and > >>>>>>> physical implementation rules are done, in a top-down manner, > driven > >>>>> from > >>>>>>> root set. e.g.: > >>>>>>> SELECT a, sum(c) FROM > >>>>>>> (SELECT * FROM R JOIN S USING (a, b)) t > >>>>>>> GROUP BY a; > >>>>>>> > >>>>>>> Suppose we have the following plan tree in the MEMO, and let's only > >>>>>>> consider distribution for simplicity, each group represents a > RelSet > >>>>> in the > >>>>>>> MEMO. > >>>>>>> > >>>>>>> Group 1: Agg on [a] > >>>>>>> Group 2: +-- MergeJoin on [a, b] > >>>>>>> Group 3: |--- TableScan R > >>>>>>> Group 4: +--- TableScan S > >>>>>>> | Group No | Operator | Derived Traits | Required Traits | > >>>>>>> | ----------- | ------------- | --------------- | --------------- > | > >>>>>>> | Group 1 | Aggregate | Hash[a] | N/A | > >>>>>>> | Group 2 | MergeJoin | Hash[a,b] | Hash[a] | > >>>>>>> | Group 3 | TableScan R | None | Hash[a,b] | > >>>>>>> | Group 4 | TableScan S | None | Hash[a,b] | > >>>>>>> > >>>>>>> We will add new interface PhysicalNode (or extending RelNode) with > >>>>>>> methods: > >>>>>>> > >>>>>>> - Pair<RelTraitSet,List<RelTraitSet>> requireTraits(RelTraitSet > >>>>> required); > >>>>>>> pair.left is the current operator's new traitset, pair.right is the > >>>>> list > >>>>>>> of children's required traitset. > >>>>>>> > >>>>>>> - RelNode passThrough(RelTraitSet required); > >>>>>>> Default implementation will call above method requireTraits() and > >>>>>>> RelNode.copy() to create new RelNode. Available to be overriden for > >>>>>>> physical operators to customize the logic. > >>>>>>> > >>>>>>> The planner will call above method on MergeJoin operator to pass > the > >>>>>>> required traits (Hash[a]) to Mergejoin's child operators. > >>>>>>> > >>>>>>> We will get a new MergeJoin: > >>>>>>> MergeJoin hash[a] > >>>>>>> |---- TableScan R hash[a] (RelSubset) > >>>>>>> +---- TableScan S hash[a] (RelSubset) > >>>>>>> > >>>>>>> Now the MEMO group looks like: > >>>>>>> | Group No | Operator | Derived Traits | Required Traits > >>>>> | > >>>>>>> | ---------- | -------- ----- | -------------------- | > >>>>>>> --------------------- | > >>>>>>> | Group1 | Aggregate | Hash[a] | N/A > >>>>>>> | > >>>>>>> | Group2 | MergeJoin | Hash[a,b], Hash[a]| Hash[a] > >>>>>>> | > >>>>>>> | Group3 | TableScan R | None | Hash[a,b], > >>>>> Hash[a] > >>>>>>> | > >>>>>>> | Group4 | TableScan S | None | Hash[a,b], > >>>>> Hash[a] > >>>>>>> | > >>>>>>> > >>>>>>> Calcite user may choose to ignore / not implement the interface to > keep > >>>>>>> the original behavior. Each physical operator, according to its own > >>>>> logic, > >>>>>>> decides whether passThrough() should pass traits down or not by > >>>>> returning: > >>>>>>> - a non-null RelNode, which means it can pass down > >>>>>>> - null object, which means can't pass down > >>>>>>> > >>>>>>> 2. Provide option to disable AbstractConverter > >>>>>>> Once the plan can request traits in top-down way in the framework, > many > >>>>>>> system don't need AbstractConverter anymore, since it is just a > >>>>>>> intermediate operator to generate physical sort / exchange. For > those, > >>>>> we > >>>>>>> can provide option to disable AbstractConverter, generate physical > >>>>> enforcer > >>>>>>> directly by adding a method to interface Convention: > >>>>>>> - RelNode enforce(RelNode node, RelTraitSet traits); > >>>>>>> > >>>>>>> The default implementation may just calling > >>>>> changeTraitsUsingConverters(), > >>>>>>> but people may choose to override it if the system has special > needs, > >>>>> like > >>>>>>> several traits must implement together, or the position of > collation in > >>>>>>> RelTraitSet is before distribution. > >>>>>>> > >>>>>>> 3. Top-down driven, on-demand rule apply > >>>>>>> For every RelNode in a RelSet, rule is matched and applied > >>>>> sequentially, > >>>>>>> newly generated RelNodes are added to the end of RelNode list in > the > >>>>> RelSet > >>>>>>> waiting for rule apply. RuleQueue and DeferringRuleCall is not > needed > >>>>>>> anymore. This will make space pruning and rule mutual exclusivity > check > >>>>>>> possible. > >>>>>>> > >>>>>>> There are 3 stages for each RelSet: > >>>>>>> a). Exploration: logical transformation, yields logical nodes > >>>>>>> b). Implementation: physical transformation, yields physical nodes > >>>>>>> c). Optimization: trait request, enforcement > >>>>>>> > >>>>>>> The general process looks like: > >>>>>>> - optimize RelSet X: > >>>>>>> implement RelSet X > >>>>>>> for each physical relnode in RelSet X: > >>>>>>> // pass down trait requests to child RelSets > >>>>>>> for each child RelSet Y of current relnode: > >>>>>>> optimize RelSet Y > >>>>>>> > >>>>>>> - implement RelSet X: > >>>>>>> if X has been implemented: > >>>>>>> return > >>>>>>> explore RelSet X > >>>>>>> for each relnode in RelSet X: > >>>>>>> apply physical rules > >>>>>>> - explore RelSet X: > >>>>>>> if X has been explored > >>>>>>> return > >>>>>>> for each relnode in RelSet X: > >>>>>>> // ensure each child RelSet of current relnode is explored > >>>>>>> for each child RelSet Y of current relnode: > >>>>>>> explore RelSet Y > >>>>>>> apply logical rules on current relnode > >>>>>>> > >>>>>>> Basically it is a state machine with several states: Initialized, > >>>>>>> Explored, Exploring, Implemented, Implementing, Optimized, > Optimizing > >>>>> and > >>>>>>> several transition methods: exploreRelSet, exploreRelNode, > >>>>> implementRelSet, > >>>>>>> implementRelNode, optimizeRelSet, optimizeRelNode... > >>>>>>> > >>>>>>> To achieve this, we need to mark the rules either logical rule or > >>>>> physical > >>>>>>> rule. > >>>>>>> To keep backward compatibility, all the un-marked rules will be > >>>>> treated as > >>>>>>> logical rules, except rules that uses AbstractConverter as rule > >>>>> operand, > >>>>>>> these rules still need to applied top-down, or random order. > >>>>>>> > >>>>>>> > >>>>>>> 4. On-demand, bottom-up trait derivation > >>>>>>> It is called bottom-up, but actually driven by top-down, happens > same > >>>>> time > >>>>>>> as top-down trait request, in optimization stage mentioned above. > Many > >>>>>>> Calcite based bigdata system only propagate traits on Project and > >>>>> Filter by > >>>>>>> writing rules, which is very limited. In fact, we can extend trait > >>>>>>> propagation/derivation to all the operators, without rules, by > adding > >>>>>>> interface PhysicalNode (or extending RelNode) with method: > >>>>>>> - RelNode derive(RelTraitSet traits, int childId); > >>>>>>> > >>>>>>> Given the following plan (only consider distribution for > simplicity): > >>>>>>> Agg [a,b] > >>>>>>> +-- MergeJoin [a] > >>>>>>> |---- TableScan R > >>>>>>> +--- TableScan S > >>>>>>> > >>>>>>> Hash[a] won't satisfy Hash[a,b] without special treatment, because > >>>>> there > >>>>>>> isn't a mechanism to coordinate traits between children. > >>>>>>> > >>>>>>> Now we call derive method on Agg [a,b] node, derive(Hash[a], 0), > we get > >>>>>>> the new node: > >>>>>>> Agg [a] > >>>>>>> +-- MergeJoin [a] (RelSubset) > >>>>>>> > >>>>>>> We will provide different matching type, so each operator can > specify > >>>>> what > >>>>>>> kind of matching type it requires its children: > >>>>>>> - MatchType getMatchType(RelTrait trait, int childId); > >>>>>>> > >>>>>>> a) Exact: Hash[a,b] exact match Hash[a,b], aka, satisfy > >>>>>>> b) Partial: Hash[a] partial match Hash[a,b] > >>>>>>> c) Permuted: Sort[a,b,c] permuted match Sort[c,b,a] > >>>>>>> > >>>>>>> In addition, optimization order is provided for each opertor to > >>>>> specify: > >>>>>>> a) left to right > >>>>>>> b) right to left > >>>>>>> c) both > >>>>>>> > >>>>>>> For example, for query SELECT * FROM R join S on R.a=S.a and > R.b=S.b > >>>>> and > >>>>>>> R.c=S.c: > >>>>>>> Suppose R is distributed by a,b, and S is distributed by c. > >>>>>>> MergeJoin [a,b,c] > >>>>>>> |--- TableScan R [a,b] > >>>>>>> +-- TableScan S [c] > >>>>>>> a) left to right, call derive(Hash[a,b], 0), we get MergeJoin [a,b] > >>>>>>> b) right to left, call derive(Hash[c], 1), we get MergeJoin [c], > most > >>>>>>> likely a bad plan > >>>>>>> c) both, get above 2 plans. > >>>>>>> > >>>>>>> For system that doesn't have a fine-tuned stats and cost model, it > may > >>>>> not > >>>>>>> be able to make a right decision based purely on cost. Probably we > >>>>> need to > >>>>>>> provide the MergeJoin with both children's derived traitset list. > >>>>>>> - List<RelNode> derive(List<List<RelTraitSet>>); > >>>>>>> > >>>>>>> Of course, all above methods are optional to implement for those > who > >>>>>>> doesn't need this feature. > >>>>>>> > >>>>>>> 5. Branch and Bound Space Pruning > >>>>>>> After we implement on-demand, top-down trait enforcement and > >>>>> rule-apply, > >>>>>>> we can pass the cost limit at the time of passing down required > >>>>> traits, as > >>>>>>> described in the classical Cascades paper. Right now, Calcite > doesn't > >>>>>>> provide group level logical properties, including stats info, each > >>>>> operator > >>>>>>> in the same group has its own logical property and the stats may > vary, > >>>>> so > >>>>>>> we can only do limited space pruning for trait enforcement, still > >>>>> good. But > >>>>>>> if we agree to add option to share group level stats between > relnodes > >>>>> in a > >>>>>>> RelSet, we will be able to do more aggresive space pruning, which > will > >>>>> help > >>>>>>> boost the performance of join reorder planning. > >>>>>>> > >>>>>>> > >>>>>>> With all that being said, how do we move forward? > >>>>>>> > >>>>>>> There are 2 ways: > >>>>>>> a) Modify on current VolcanoPlanner. > >>>>>>> Pros: code reuse, existing numerous test cases and infrastructure, > >>>>> fast > >>>>>>> integration > >>>>>>> Cons: changing code always brings risk > >>>>>>> > >>>>>>> b) Add a new XXXXPlanner > >>>>>>> Pros: no risk, no tech debt, no need to worry about backward > >>>>>>> compatability > >>>>>>> Cons: separate test cases for new planner, one more planner to > >>>>> maintain > >>>>>>> > >>>>>>> We'd like to hear the community's thoughts and advices. > >>>>>>> > >>>>>>> Thanks. > >>>>>>> - Haisheng > >>>>>>> > >>>>>>> > >>>>>>> > >>>>>>> > >>>>>>> > >>>>>>> > >>>>>>> > >>>>>> > >>>>> > >>>> > >> > >> > >
