Vineet,

In case you forgot, can you please log that JIRA case? If we have a lengthy 
design discussion without creating an action item, we are wasting everyone’s 
time.

Julian

> On Nov 1, 2016, at 11:00 AM, Julian Hyde <[email protected]> wrote:
> 
> Alexander & Vineet,
> 
> One further comment about “NOT IN”. SQL in general is fairly close to 
> relational algebra, but “NOT IN” is one of the places where the gap is 
> widest. “NOT IN” is difficult in general to execute efficiently, because of 
> the problem of NULL values (at Oracle, we always recommended to users to 
> rewrite as NOT EXISTS if possible). The gap between SQL and relational 
> algebra is apparent when a short SQL query becomes a complex RelNode tree.
> 
> There is a silver lining: the RelNode tree, being relational algebra, has 
> well-behaved semantics. Once you’re in RelNode land, you can freely apply 
> transformation rules to make it efficient.
> 
> Vineet,
> 
> If the planner rules produce a plan that is sub-optimal I wouldn’t call it a 
> bug but a missing feature. (It would be a bug if the planner over-reached and 
> created a plan that gave wrong results, so I always tend to be conservative 
> about adding rules.)
> 
> Usually it’s OK if we make a mess in SQL-to-RelNode conversion. A few 
> redundant projects and filters are no problem, and can be easily removed 
> later with rules that reduce constants and propagate predicates throughout 
> the tree. But for the general case of NOT IN, we have to add a self-join to 
> deal with the possibility that the key has NULL values. After constant 
> reduction has kicked in and we have realized that NULL key values are not 
> possible, it is not easy to remove that self-join.
> 
> Here is a very simple query where this happens:
> 
> sqlline> !connect 
> jdbc:calcite:model=core/src/test/resources/hsqldb-model.json sa ""
> sqlline> !set outputformat csv
> sqlline> explain plan for select * from scott.emp where deptno not in (
>> select deptno from scott.dept where deptno = 20);
> 'PLAN'
> 'EnumerableCalc(expr#0..11=[{inputs}], expr#12=[0], expr#13=[=($t8, $t12)], 
> expr#14=[false], expr#15=[IS NOT NULL($t11)], expr#16=[true], expr#17=[IS 
> NULL($t7)], expr#18=[null], expr#19=[<($t9, $t8)], expr#20=[CASE($t13, $t14, 
> $t15, $t16, $t17, $t18, $t19, $t16, $t14)], expr#21=[NOT($t20)], 
> proj#0..7=[{exprs}], $condition=[$t21])
>  EnumerableJoin(condition=[=($7, $10)], joinType=[left])
>    EnumerableCalc(expr#0..9=[{inputs}], EMPNO=[$t2], ENAME=[$t3], JOB=[$t4], 
> MGR=[$t5], HIREDATE=[$t6], SAL=[$t7], COMM=[$t8], DEPTNO=[$t9], c=[$t0], 
> ck=[$t1])
>      EnumerableJoin(condition=[true], joinType=[inner])
>        JdbcToEnumerableConverter
>          JdbcAggregate(group=[{}], c=[COUNT()], ck=[COUNT($0)])
>            JdbcFilter(condition=[=(CAST($0):INTEGER NOT NULL, 20)])
>              JdbcTableScan(table=[[SCOTT, DEPT]])
>        JdbcToEnumerableConverter
>          JdbcTableScan(table=[[SCOTT, EMP]])
>    JdbcToEnumerableConverter
>      JdbcAggregate(group=[{0, 1}])
>        JdbcProject(DEPTNO=[$0], i=[true])
>          JdbcFilter(condition=[=(CAST($0):INTEGER NOT NULL, 20)])
>            JdbcTableScan(table=[[SCOTT, DEPT]])
> '
> 1 row selected (0.067 seconds)
> 
> Note that there are two scans of DEPT, but one is sufficient because DEPTNO 
> is never null. In the JdbcAggregate, c always equals ck, and therefore the 
> CASE can be simplified, and therefore the scan of DEPT that produces c and ck 
> can be dropped, but Calcite rules cannot deduce that fact.
> 
> Can you please log a JIRA case for this? See if you can find some other 
> queries (maybe using IN rather than NOT IN, or whose key columns are not so 
> obviously NOT NULL) and include these in the JIRA case also.
> 
> I doubt we can fix using a planner rule. The best solution may be to use 
> RelMetadataQuery.getPulledUpPredicates() to simplify the CASE before we add 
> the join.
> 
> Julian
> 
> 
>> On Nov 1, 2016, at 8:49 AM, Alexander Shoshin <[email protected]> 
>> wrote:
>> 
>> Julian, thank you for help.
>> 
>> I had a wrong picture of NULL values processing. So, it looks like there is 
>> some problem in my planner rules.
>> As for the AST, I was confused by the wrong Flink "explain()" function 
>> description :)
>> 
>> 
>> Regards,
>> Alexander
>> 
>> -----Original Message-----
>> From: Julian Hyde [mailto:[email protected]] 
>> Sent: Monday, October 31, 2016 10:43 PM
>> To: [email protected]
>> Subject: Re: Problems with abstract syntax tree
>> 
>> The behavior of NOT IN in SQL is complicated when there are NULL values 
>> around. In particular, if one "word" value from the sub-query is null, then 
>> the outer query must return 0 rows. (Why? Because "word NOT IN ('foo', 'bar' 
>> null)" would evaluate to UNKNOWN for every row.)
>> 
>> It is valid to deduce that "word" in the sub-query is never null, because of 
>> the "WHERE word = 'hello'" condition. I would have hoped that a constant 
>> reduction could do that, and then maybe the CASE expression can be 
>> simplified.
>> 
>> By the way, to be pedantic, what we are talking about here is the RelNode 
>> tree, the relational algebra, which comes out of the SqlToRelConverter. The 
>> AST is the SqlNode tree, which comes out of the parser and goes into the 
>> SqlToRelConverter.
>> 
>> On Mon, Oct 31, 2016 at 8:46 AM, Alexander Shoshin 
>> <[email protected]> wrote:
>>> Hello, everybody.
>>> 
>>> Trying to resolve an Apache Flink issue I got some troubles with Calcite. 
>>> Can you help me to understand is there a problem in Calcite or just in 
>>> wrong settings passed to Calcite functions?
>>> 
>>> I have a simple table "Words" with one column named "word" and a query with 
>>> NOT IN operator:
>>> val query = "SELECT word FROM Words WHERE word NOT IN (SELECT word FROM 
>>> Words WHERE word = 'hello')"
>>> 
>>> This query parsed by org.apache.calcite.sql.parser.SqlParser.parseStmt() 
>>> and then transformed to a relational tree by 
>>> org.apache.calcite.sql2rel.SqlToRelConverter.convertQuery(...).
>>> 
>>> As a result I see the following abstract syntax tree
>>> LogicalProject(word=[$0])
>>> LogicalFilter(condition=[NOT(CASE(=($1, 0), false, IS NOT NULL($5), true, 
>>> IS NULL($3), null, <($2, $1), null, false))])
>>>   LogicalJoin(condition=[=($3, $4)], joinType=[left])
>>>     LogicalProject($f0=[$0], $f1=[$1], $f2=[$2], $f3=[$0])
>>>       LogicalJoin(condition=[true], joinType=[inner])
>>>         EnumerableTableScan(table=[[Words]])
>>>         LogicalAggregate(group=[{}], agg#0=[COUNT()], agg#1=[COUNT($0)])
>>>           LogicalProject($f0=[$0], $f1=[true])
>>>             LogicalProject(word=[$0])
>>>               LogicalFilter(condition=[=($0, 'hello')])
>>>                 EnumerableTableScan(table=[[Words]])
>>>     LogicalAggregate(group=[{0}], agg#0=[MIN($1)])
>>>       LogicalProject($f0=[$0], $f1=[true])
>>>         LogicalProject(word=[$0])
>>>           LogicalFilter(condition=[=($0, 'hello')])
>>>             EnumerableTableScan(table=[[Words]])
>>> 
>>> which fails later during query plan optimization (while calling 
>>> org.apache.calcite.tools.Programs.RuleSetProgram.run()).
>>> 
>>> I think it might be because of a very complex abstract syntax tree 
>>> generated by Calcite. Shouldn't it be more simple? This one looks good for 
>>> me:
>>> LogicalProject(word=[$0])
>>> LogicalFilter(condition=[IS NULL($2)])
>>>   LogicalJoin(condition=[=($0, $1)], joinType=[left])
>>>     EnumerableTableScan(table=[[Words]])
>>>     LogicalProject($f0=[$0], $f1=[true])
>>>       LogicalProject(word=[$0])
>>>         LogicalFilter(condition=[=($0, 'hello')])
>>>           EnumerableTableScan(table=[[Words]])
>>> 
>>> And when I write a query using LEFT OUTER JOIN to receive this syntax tree 
>>> - the optimization works fine. And the query execution result is the same 
>>> as must be in case of using NOT IN. So am I wrong with a supposition about 
>>> bad abstract syntax tree or not? I will be glad to receive any comments.
>>> 
>>> Regards,
>>> Alexander
> 

Reply via email to