[jira] [Created] (KYLIN-3149) Calcite's ReduceExpressionsRule.PROJECT_INSTANCE not working as expected

["hongbin ma (JIRA)" <ji...@apache.org>]

hongbin ma created KYLIN-3149:
---------------------------------

             Summary: Calcite's ReduceExpressionsRule.PROJECT_INSTANCE not working as expected
                 Key: KYLIN-3149
                 URL: https://issues.apache.org/jira/browse/KYLIN-3149
             Project: Kylin
          Issue Type: Bug
    Affects Versions: v2.2.0
            Reporter: hongbin ma


for queries like:

{code:sql}
select TRANS_ID from kylin_sales group by cast (case 
WHEN  '1030101' = '1030101' then substring(COALESCE(OPS_USER_ID, '999999999999'), 1, 1)
when  '1030101' = '1030102' then substring(COALESCE(OPS_REGION, '999999999999'), 1, 1)  
when  '1030101' = '1030103' then substring(COALESCE(LSTG_FORMAT_NAME, '999999999999'), 1, 1)
when  '1030101' = '1030104' then substring(COALESCE(LSTG_FORMAT_NAME, '999999999999'), 1, 1)
end as varchar(256)), TRANS_ID;
{code}

the expected logical plan after volcano is:

{code}
EXECUTION PLAN BEFORE REWRITE
OLAPToEnumerableConverter
  OLAPProjectRel(TRANS_ID=[$1], ctx=[])
    OLAPLimitRel(ctx=[], fetch=[50000])
      OLAPAggregateRel(group=[{0, 1}], ctx=[])
        OLAPProjectRel($f0=[SUBSTRING(CASE(IS NOT NULL($9), $9, '999999999999'), 1, 1)], TRANS_ID=[$0], ctx=[])
          OLAPTableScan(table=[[DEFAULT, KYLIN_SALES]], ctx=[], fields=[[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]])
{code}

however the actual is:

{code}
EXECUTION PLAN BEFORE REWRITE
OLAPToEnumerableConverter
  OLAPLimitRel(ctx=[], fetch=[50000])
    OLAPProjectRel(TRANS_ID=[$1], ctx=[])
      OLAPAggregateRel(group=[{0, 1}], ctx=[])
        OLAPProjectRel($f0=[CAST(CASE(=('1030101', '1030101'), SUBSTRING(CASE(IS NOT NULL($9), $9, '999999999999'), 1, 1), =('1030101', '1030102'), SUBSTRING(CASE(IS NOT NULL($10), $10, '999999999999'), 1, 1), =('1030101', '1030103'), SUBSTRING(CASE(IS NOT NULL($2), $2, '999999999999'), 1, 1), =('1030101', '1030104'), SUBSTRING(CASE(IS NOT NULL($2), $2, '999999999999'), 1, 1), null)):VARCHAR(256) CHARACTER SET "UTF-16LE" COLLATE "UTF-16LE$en_US$primary"], TRANS_ID=[$0], ctx=[])
          OLAPTableScan(table=[[DEFAULT, KYLIN_SALES]], ctx=[], fields=[[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]])
{code}

looks like Calcite's ReduceExpressionsRule.PROJECT_INSTANCE not working as expected. If we dump the internal state of this VolcanoPlanner (org.apache.calcite.plan.volcano.VolcanoPlanner#dump), line 19-21 from the complete dump is attached:

{code}
	rel#337:Subset#1.OLAP.[], best=rel#339, importance=0.6561
		rel#339:OLAPProjectRel.OLAP.[](input=rel#303:Subset#0.OLAP.[],$f0=CAST(CASE(=('1030101', '1030101'), SUBSTRING(CASE(IS NOT NULL($9), $9, '999999999999'), 1, 1), =('1030101', '1030102'), SUBSTRING(CASE(IS NOT NULL($10), $10, '999999999999'), 1, 1), =('1030101', '1030103'), SUBSTRING(CASE(IS NOT NULL($2), $2, '999999999999'), 1, 1), =('1030101', '1030104'), SUBSTRING(CASE(IS NOT NULL($2), $2, '999999999999'), 1, 1), null)):VARCHAR(256) CHARACTER SET "UTF-16LE" COLLATE "UTF-16LE$en_US$primary",TRANS_ID=$0,ctx=), rowcount=100.0, cumulative cost={15.0 rows, 25.05 cpu, 0.0 io}
		rel#348:OLAPProjectRel.OLAP.[](input=rel#303:Subset#0.OLAP.[],$f0=SUBSTRING(CASE(IS NOT NULL($9), $9, '999999999999'), 1, 1),TRANS_ID=$0,ctx=), rowcount=100.0, cumulative cost={15.0 rows, 25.05 cpu, 0.0 io}
{code}

we see two rels with same cost:  #339 and #348, where #339 is created from LogicalProject = (OLAPProjectRule)=> OLAPProject, and #348 is created from LogicalProject =( ReduceExpressionsRule) => Reduced LogicalProject =(OLAPProjectRule)=> Reduced OLAPProject . Since ReduceExpressionsRule require Logical Project rather than OLAP Project, #339 is never reduced.

The worse thing is that cost of #339 and #348 are same. By current volcano planner algorithm  the first met rel will be chosen, so unexpected rel is chosen

A simple approach to fix this is to refine the rel choosing algorithm: when two rels are equal in cost, choose a "simpler" one. Since we don't have a perfect measurement of "simple", we simply choose the rel with smaller toString() length



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)