calcite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Remus Rusanu <rrus...@hortonworks.com>
Subject Re: Cluster mismatch between RelNodes of a query and a materialized view
Date Thu, 09 Mar 2017 12:39:59 GMT
I think I have a solution that works, is not intrusive on base RelNode, and can be extended
by Calcite consumers using custom rel nodes.
A new RelShuttle is added to calcite, calcite.tools.RelCopier. This implements the LogicalXxx
copy operators. Hive extends this into HiveRelCopier which implements the HiveXxx copy operators.
This pattern already exists in Hive with HiveRelShuttle. I have updated the https://github.com/apache/calcite/pull/392
with the skeleton of the Calcite side changes, and is entirely contained in a shuttle, no
RelNode changes required. I will continue working on this to extend support for all LogicalXxx
operators, to add tests etc.

Thanks,
~Remus


On 3/8/17, 11:40 AM, "Julian Hyde" <jhyde@apache.org> wrote:

    > On Mar 8, 2017, at 11:27 AM, Ashutosh Chauhan <hashutosh@apache.org> wrote:
    > 
    > Fact that
    > RelFactories exist is evidence enough that LogicalRelNodes are not
    > sufficient outside of Calcite.
    
    Not true.
    
    Hive uses its own RelNodes far more than other projects do. The other projects do a significant
part of their planning in the logical domain, and only later move to the physical domain.
They use RelFactories so that they can re-use rules when they are in the physical domain.
    
    Materialized view substitution is a logical operation and it makes sense to do it in the
logical domain.
    
    I don’t recall the details of the Hive RelNodes but isn’t it true that in most cases
they have the same information as the corresponding Logical RelNode?
    
    We can find a solution to this problem for Hive RelNodes, but I still feel that Hive overuses
Hive RelNodes and that causes inefficiencies for both Hive and Calcite.
    
    Julian
    
    
    
    
    > On Mar 8, 2017, at 11:27 AM, Ashutosh Chauhan <hashutosh@apache.org> wrote:
    > 
    > Working on LogicalRelNodes doesnt solve problem for Hive (or any other
    > project which use RelFactories to generate custom rel node). Fact that
    > RelFactories exist is evidence enough that LogicalRelNodes are not
    > sufficient outside of Calcite. We have to solve this problem in that
    > context.
    > 
    > On Wed, Mar 8, 2017 at 10:50 AM, Remus Rusanu <rrusanu@hortonworks.com>
    > wrote:
    > 
    >> To make sure I understand, you’re saying that the cloning would take a
    >> tree of various Xxx operators (eg. HiveProject) and generate a tree of
    >> equivalent LogicalXxx  operators (HiveProject -> LogicalProject,
    >> HiveFilter->LogicalFilter, HiveAggregate ->  LogicalAggregate and so on).
    >> Is my understanding correct? Ignoring Hive issues, wouldn’t that loose
    >> information?
    >> For Hive specific case, the HiveXxx operator extend Xxx not LogicalXxx
    >> (not sure yet if this is an issue or not) but for sure they carry lots of
    >> extra info that would be lost if the clone would contain LogicalXxx instead
    >> of HiveXxx.
    >> 
    >> Thanks,
    >> ~Remus
    >> 
    >> On 3/8/17, 9:45 AM, "Julian Hyde" <jhyde@apache.org> wrote:
    >> 
    >>    Could you work on LogicalXxx rather than HiveXxx? I know the Hive team
    >> likes to do everything in terms of HiveXxx RelNodes but I’m not sure it has
    >> to be that way.
    >> 
    >>    Julian
    >> 
    >>> On Mar 8, 2017, at 9:37 AM, Remus Rusanu <rrusanu@hortonworks.com>
    >> wrote:
    >>> 
    >>> Agree on the RelOptCluster.
    >>> 
    >>> I noticed how adding new method to RelNode yields a gargantuan task
    >> (from the list of compile errors I got…). But I’m not sure a RelShuttle can
    >> handle this. For my test case the 3 nodes that need to be cloned are
    >> HiveScanTable, HiveFilter and HiveProject, all declared in Hive and not
    >> even extending AbstractRelNode but directly base RelNode. A RelShuttle
    >> wouldn’t know about these types, and wouldn’t be able to create them (short
    >> of using reflection and adhering to a strict constructor signature, which I
    >> think is much too fragile). What I did to get the ball rolling I added a
    >> default implementation in AbstractRelNode (basically assert ‘must be
    >> implemented by subclass’), this allowed me to test easily and, as a proof
    >> of concept, I have it working. But I reckon is fragile test wise, and new
    >> RelNode types wouldn’t know about the requirement to provide a copyTo
    >> implementation.
    >>> 
    >>> Thanks,
    >>> ~Remus
    >>> 
    >>> On 3/8/17, 9:05 AM, "Julian Hyde" <jhyde@apache.org> wrote:
    >>> 
    >>>   The argument should be a RelOptCluster, not a RelOptPlanner. The
    >> link from RelNode to planner is indirect currently (via cluster) and will
    >> be non-existent after CALCITE-1536.
    >>> 
    >>>   I question whether we need a new method. Putting an abstract
    >> method on RelNode is a huge burden because every RelNode sub-class needs to
    >> be fixed when people upgrade. Even a non-abstract method imposes a
    >> conceptual burden: more methods to understand.
    >>> 
    >>>   So, my approach would be to sub-class RelShuttle. It’s sufficient
    >> that it only works for LogicalXxx nodes.
    >>> 
    >>>   No need to copy RexNode expressions. They are immutable.
    >>> 
    >>>   Julian
    >>> 
    >>> 
    >>>> On Mar 8, 2017, at 4:14 AM, Remus Rusanu <rrusanu@hortonworks.com>
    >> wrote:
    >>>> 
    >>>> I created CALCITE-1681 https://issues.apache.org/
    >> jira/browse/CALCITE-1681 and I intent to work on it for finishing
    >> HIVE-15708.
    >>>> My current thinking is to create a RelCopier based on RelShuttle
    >> and add a new abstract RelNode.copyTo(RelOptPlanner) that each concrete Rel
    >> type must override. The Rex part is already handled by the existing
    >> RexCopier.
    >>>> 
    >>>> Thanks,
    >>>> ~Remus
    >>>> 
    >>>> On 3/6/17, 12:30 PM, "Julian Hyde" <jhyde@apache.org> wrote:
    >>>> 
    >>>>  Every RelNode belongs to a RelOptCluster, and basically there is
    >> one RelOptCluster created each time a query is prepared. When working with
    >> materialized views, the view’s query is represented as a tree of RelNodes,
    >> that tree is used for optimizing more than one query. When planning a
    >> particular query, the nodes of that query will have a different
    >> RelOptCluster than the nodes of the materialized view(s) they are matched
    >> against.
    >>>> 
    >>>>  How do we deal with this? Do we copy the nodes into the query’s
    >> cluster once we have found a match? If so, how? I couldn’t find a sub-class
    >> of RelVisitor or RelShuttle that copies trees to a different RelOptCluster.
    >>>> 
    >>>>  By the way, https://issues.apache.org/jira/browse/CALCITE-1536 <
    >> https://issues.apache.org/jira/browse/CALCITE-1536> aims to improve the
    >> RelNode life-cycle but I don’t think it will solve this problem.
    >>>> 
    >>>>  Julian
    >>>> 
    >>>> 
    >>> 
    >>> 
    >>> 
    >>> 
    >> 
    >> 
    >> 
    >> 
    >> 
    
    
    

Mime
View raw message