Colleagues,
sorry for late reply.
@Igor, I wouldn't say that materialized views work like rules. There is
no pattern matching there (at least for trivial cases like sorted
indexes, for more complex views pattern matching will have place, see
VolcanoPlanner#registerMaterializations). Also it couldn't be
retriggered several times. The better analogy here is an extra call of
VolcanoPlanner#register for index scans and registering them in the same
sets as usual scans. See an example below.
@Vladimir, I'm not sure that phoenix approach will significantly reduce
optimization time, but it looks like materializations might save some
efforts. Lets consider an example with Merge join.
With drill-like approach initially we have:
LogicalJoin(emps.depId=deps.id)
LogicalScan(emps)
LogicalScan(deps)
then we apply converters to convert this tree into the physical
representation. Assume that both tables are sorted by 'id' column and
there is an index on 'emps.depId' column. Also they are collocated on
'emps.depId=deps.id' columns. You apply MergeJoinRule and demand it's
inputs are sorted on 'emps.depId' and 'deps.id' together. After applying
this rule, converting scans to the physical nodes and expanding
AbstractConverters you'll end up with
MergeJoin(emps.depId=deps.id)
Sort(emps.depId)
PhysicalScan(emps)
PhysicalScan(deps)
then you apply ScanToIndexScanRule
MergeJoin(emps.depId=deps.id)
Sort(emps.depId)
PhysicalIndexScan(emps.depId)
PhysicalScan(deps)
and finally after removing redundant sort by SortRemoveRule you'll get
MergeJoin(emps.depId=deps.id)
PhysicalIndexScan(emps.depId)
PhysicalScan(deps)
Now, lets take a look to the same optimization process with the
phoenix-like approach. Initially we have the same query tree:
LogicalJoin(emps.depId=deps.id)
LogicalScan(emps)
LogicalScan(deps)
But then, just before planning, we register materializations (see the
beginning of the VolcanoPlanner#findBestExp method). And query tree now
looks like
LogicalJoin(emps.depId=deps.id)
Set 0: [LogicalScan(emps), LogicalIndexScan(collation=emps.depId)]
LogicalScan(deps)
Note that we have two scans with different collations for 'emps' table
in the Set0. And this happened before the actual planning process. After
converting scans to the physical nodes we'll have:
LogicalJoin(emps.depId=deps.id)
Set 0: [PhysicalScan(emps), PhysicalIndexScan(collation=emps.depId)]
PhysicalScan(deps)
and after applying MergeJoinRule and demanding that 'deps' should be
sorted by 'id' column and 'emps' should be sorted by 'depId' column, we
will end up with a tree without Sort operator (unlike in drill case),
because we have already had a properly sorted subset for 'emps' scan.
The tree will look like this:
MergeJoin(emps.depId=deps.id)
PhysicalIndexScan(collation=emps.depId)
PhysicalScan(deps)
So, we get to the same point without creating and removing redundant
sort, because we have all possible index scans registered before the
planning is actually started and we can demand sortedness of table scans
directly without applying IndexRules and Abstract converters.
--
Kind Regards
Roman Kondakov
On 13.12.2019 12:38, Seliverstov Igor wrote:
> Colleagues,
>
> As far as I understand, materialization acts like a special rule, that
> matches some subtree pattern (a leaf part of a query plan) to a star table,
> which may have better cost than the subtree, it replaces. Saying that, in
> general, there is no difference between approaches - they do the same almost
> in the same way but using different API.
>
> My opinion is it’s better to do the deal using rules - it makes overall
> approach consistent.
>
> Regards,
> Igor
>
>> 12 дек. 2019 г., в 10:03, Vladimir Ozerov <ppoze...@gmail.com> написал(а):
>>
>> Roman,
>>
>> What I am trying to understand is what advantage of materialization API you
>> see over the normal optimization process? Does it save optimization time,
>> or reduce memory footprint, or maybe provide better plans? I am asking
>> because I do not see how expressing indexes as materializations fit
>> classical optimization process. We discussed Sort <- Scan optimization.
>> Let's consider another example:
>>
>> LogicalSort[a ASC]
>> LogicalJoin
>>
>> Initially, you do not know the implementation of the join, and hence do not
>> know it's collation. Then you may execute physical join rules, which
>> produce, say, PhysicalMergeJoin[a ASC]. If you execute sort implementation
>> rule afterwards, you may easily eliminate the sort, or make it simpler
>> (e.g. remove local sorting phase), depending on the distribution. In other
>> words, proper implementation of sorting optimization assumes that you have
>> a kind of SortRemoveRule anyway, irrespectively of whether you use
>> materializations or not, because sorting may be injected on top of any
>> operator. With this in mind, the use of materializations doesn't make the
>> planner simpler. Neither it improves the outcome of the whole optimization
>> process.
>>
>> What is left is either lower CPU or RAM usage? Is this the case?
>>
>> ср, 11 дек. 2019 г. в 18:37, Roman Kondakov <kondako...@mail.ru.invalid>:
>>
>>> Vladimir,
>>>
>>> the main advantage of the Phoenix approach I can see is the using of
>>> Calcite's native materializations API. Calcite has advanced support for
>>> materializations [1] and lattices [2]. Since secondary indexes can be
>>> considered as materialized views (it's just a sorted representation of
>>> the same table) we can seamlessly use views to simulate indexes behavior
>>> for Calcite planner.
>>>
>>>
>>> [1] https://calcite.apache.org/docs/materialized_views.html
>>> [2] https://calcite.apache.org/docs/lattice.html
>>>
>>> --
>>> Kind Regards
>>> Roman Kondakov
>>>
>>>
>>> On 11.12.2019 17:11, Vladimir Ozerov wrote:
>>>> Roman,
>>>>
>>>> What is the advantage of Phoenix approach then? BTW, it looks like
>>> Phoenix
>>>> integration with Calcite never made it to production, did it?
>>>>
>>>> вт, 10 дек. 2019 г. в 19:50, Roman Kondakov <kondako...@mail.ru.invalid
>>>> :
>>>>
>>>>> Hi Vladimir,
>>>>>
>>>>> from what I understand, Drill does not exploit collation of indexes. To
>>>>> be precise it does not exploit index collation in "natural" way where,
>>>>> say, we a have sorted TableScan and hence we do not create a new Sort.
>>>>> Instead of it Drill always create a Sort operator, but if TableScan can
>>>>> be replaced with an IndexScan, this Sort operator is removed by the
>>>>> dedicated rule.
>>>>>
>>>>> Lets consider initial an operator tree:
>>>>>
>>>>> Project
>>>>> Sort
>>>>> TableScan
>>>>>
>>>>> after applying rule DbScanToIndexScanPrule this tree will be converted
>>> to:
>>>>>
>>>>> Project
>>>>> Sort
>>>>> IndexScan
>>>>>
>>>>> and finally, after applying DbScanSortRemovalRule we have:
>>>>>
>>>>> Project
>>>>> IndexScan
>>>>>
>>>>> while for Phoenix approach we would have two equivalent subsets in our
>>>>> planner:
>>>>>
>>>>> Project
>>>>> Sort
>>>>> TableScan
>>>>>
>>>>> and
>>>>>
>>>>> Project
>>>>> IndexScan
>>>>>
>>>>> and most likely the last plan will be chosen as the best one.
>>>>>
>>>>> --
>>>>> Kind Regards
>>>>> Roman Kondakov
>>>>>
>>>>>
>>>>> On 10.12.2019 17:19, Vladimir Ozerov wrote:
>>>>>> Hi Roman,
>>>>>>
>>>>>> Why do you think that Drill-style will not let you exploit collation?
>>>>>> Collation should be propagated from the index scan in the same way as
>>> in
>>>>>> other sorted operators, such as merge join or streaming aggregate.
>>>>> Provided
>>>>>> that you use converter-hack (or any alternative solution to trigger
>>>>> parent
>>>>>> re-analysis).
>>>>>> In other words, propagation of collation from Drill-style indexes
>>> should
>>>>> be
>>>>>> no different from other sorted operators.
>>>>>>
>>>>>> Regards,
>>>>>> Vladimir.
>>>>>>
>>>>>> вт, 10 дек. 2019 г. в 16:40, Zhenya Stanilovsky
>>>>> <arzamas...@mail.ru.invalid
>>>>>>> :
>>>>>>
>>>>>>>
>>>>>>> Roman just as fast remark, Phoenix builds their approach on
>>>>>>> already existing monolith HBase architecture, most cases it`s just a
>>>>> stub
>>>>>>> for someone who wants use secondary indexes with a base with no
>>>>>>> native support of it. Don`t think it`s good idea here.
>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> ------- Forwarded message -------
>>>>>>>> From: "Roman Kondakov" < kondako...@mail.ru.invalid >
>>>>>>>> To: dev@ignite.apache.org
>>>>>>>> Cc:
>>>>>>>> Subject: Adding support for Ignite secondary indexes to Apache
>>> Calcite
>>>>>>>> planner
>>>>>>>> Date: Tue, 10 Dec 2019 15:55:52 +0300
>>>>>>>>
>>>>>>>> Hi all!
>>>>>>>>
>>>>>>>> As you may know there is an activity on integration of Apache Calcite
>>>>>>>> query optimizer into Ignite codebase is being carried out [1],[2].
>>>>>>>>
>>>>>>>> One of a bunch of problems in this integration is the absence of
>>>>>>>> out-of-the-box support for secondary indexes in Apache Calcite. After
>>>>>>>> some research I came to conclusion that this problem has a couple of
>>>>>>>> workarounds. Let's name them
>>>>>>>> 1. Phoenix-style approach - representing secondary indexes as
>>>>>>>> materialized views which are natively supported by Calcite engine [3]
>>>>>>>> 2. Drill-style approach - pushing filters into the table scans and
>>>>>>>> choose appropriate index for lookups when possible [4]
>>>>>>>>
>>>>>>>> Both these approaches have advantages and disadvantages:
>>>>>>>>
>>>>>>>> Phoenix style pros:
>>>>>>>> - natural way of adding indexes as an alternative source of rows:
>>> index
>>>>>>>> can be considered as a kind of sorted materialized view.
>>>>>>>> - possibility of using index sortedness for stream aggregates,
>>>>>>>> deduplication (DISTINCT operator), merge joins, etc.
>>>>>>>> - ability to support other types of indexes (i.e. functional
>>> indexes).
>>>>>>>>
>>>>>>>> Phoenix style cons:
>>>>>>>> - polluting optimizer's search space extra table scans hence
>>> increasing
>>>>>>>> the planning time.
>>>>>>>>
>>>>>>>> Drill style pros:
>>>>>>>> - easier to implement (although it's questionable).
>>>>>>>> - search space is not inflated.
>>>>>>>>
>>>>>>>> Drill style cons:
>>>>>>>> - missed opportunity to exploit sortedness.
>>>>>>>>
>>>>>>>> There is a good discussion about using both approaches can be found
>>> in
>>>>>>> [5].
>>>>>>>>
>>>>>>>> I made a small sketch [6] in order to demonstrate the applicability
>>> of
>>>>>>>> the Phoenix approach to Ignite. Key design concepts are:
>>>>>>>> 1. On creating indexes are registered as tables in Calcite schema.
>>> This
>>>>>>>> step is needed for internal Calcite's routines.
>>>>>>>> 2. On planner initialization we register these indexes as
>>> materialized
>>>>>>>> views in Calcite's optimizer using VolcanoPlanner#addMaterialization
>>>>>>>> method.
>>>>>>>> 3. Right before the query execution Calcite selects all materialized
>>>>>>>> views (indexes) which can be potentially used in query.
>>>>>>>> 4. During the query optimization indexes are registered by planner as
>>>>>>>> usual TableScans and hence can be chosen by optimizer if they have
>>>>> lower
>>>>>>>> cost.
>>>>>>>>
>>>>>>>> This sketch shows the ability to exploit index sortedness only. So
>>> the
>>>>>>>> future work in this direction should be focused on using indexes for
>>>>>>>> fast index lookups. At first glance FilterableTable and
>>>>>>>> FilterTableScanRule are good points to start. We can push Filter into
>>>>>>>> the TableScan and then use FilterableTable for fast index lookups
>>>>>>>> avoiding reading the whole index on TableScan step and then filtering
>>>>>>>> its output on the Filter step.
>>>>>>>>
>>>>>>>> What do you think?
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> [1]
>>>>>>>>
>>>>>>>
>>>>>
>>> http://apache-ignite-developers.2346864.n4.nabble.com/New-SQL-execution-engine-tt43724.html#none
>>>>>>>> [2]
>>>>>>>>
>>>>>>>
>>>>>
>>> https://cwiki.apache.org/confluence/display/IGNITE/IEP-37%3A+New+query+execution+engine
>>>>>>>> [3] https://issues.apache.org/jira/browse/PHOENIX-2047
>>>>>>>> [4] https://issues.apache.org/jira/browse/DRILL-6381
>>>>>>>> [5] https://issues.apache.org/jira/browse/DRILL-3929
>>>>>>>> [6] https://github.com/apache/ignite/pull/7115
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>
