Saturday, February 25, 2012

Looking for query plan determination information

Hi,

I was just helping a coworker optimize a query. He had two versions:
one which used UNION for each value for which he was tallying results
and another query which used GROUP BY. Here is an aproximation of what
they were:

Query #1:
---
SELECT 12 AS [Row],
ISNULL(SUM(CASE WHEN T.my_date BETWEEN @.week_start_date AND
DATEADD(d, 1, @.week_start_date) THEN 1 ELSE 0 END), 0) AS [Monday],
ISNULL(SUM(CASE WHEN T.my_date BETWEEN DATEADD(d, 1,
@.week_start_date) AND DATEADD(d, 2, @.week_start_date) THEN 1 ELSE 0
END), 0) AS [Tuesday]
FROM My_Table T
INNER JOIN Another_Table T2 ON T2.col1 = T.col1
WHERE T.my_date BETWEEN @.week_start_date AND @.week_end_date
AND T.col2 = 5
UNION
SELECT 13 AS [Row],
ISNULL(SUM(CASE WHEN T.my_date BETWEEN @.week_start_date AND
DATEADD(d, 1, @.week_start_date) THEN 1 ELSE 0 END), 0) AS [Monday],
ISNULL(SUM(CASE WHEN T.my_date BETWEEN DATEADD(d, 1,
@.week_start_date) AND DATEADD(d, 2, @.week_start_date) THEN 1 ELSE 0
END), 0) AS [Tuesday]
FROM My_Table T
INNER JOIN Another_Table T2 ON T2.col1 = T.col1
WHERE T.my_date BETWEEN @.week_start_date AND @.week_end_date
AND T.col2 = 6

Query #2:
---
SELECT R.row_num AS [Row],
ISNULL(SUM(CASE WHEN T.my_date BETWEEN @.week_start_date AND
DATEADD(d, 1, @.week_start_date) THEN 1 ELSE 0 END), 0) AS [Monday],
ISNULL(SUM(CASE WHEN T.my_date BETWEEN DATEADD(d, 1,
@.week_start_date) AND DATEADD(d, 2, @.week_start_date) THEN 1 ELSE 0
END), 0) AS [Tuesday]
FROM My_Table T
INNER JOIN Another_Table T2 ON T2.col1 = T.col1
INNER JOIN Report_Rows R ON R.col2 = T.col2
WHERE T.my_date BETWEEN @.week_start_date AND @.week_end_date
GROUP BY ALL R.row_num
ORDER BY R.row_num

The Report_Rows table in this case would have had two rows mapping row
12 to a column value of 5 and row 13 to a column value of 6. The
second query was performing horribly until I noticed the ALL keyword
in the GROUP BY, which I didn't think was necessary. When I removed
that it performed more like I expected it to perform.

Before I had noticed that I was scouring over the query plans and
couldn't figure out why in one instance the query optimizer chose to
join My_Table and Another_Table, yet when the ALL keyword was there it
chose to return all of the records from Another_Table (a rather large
table) and join it to the Report_Rows table before then joining to
My_Table, which had the date criteria in the WHERE clause.

So, if you've read this far without giving up...

1. Why would the ALL keyword cause this? I understand the
functionality of ALL, but I still don't see why that caused the
reordering of the joins.

2. (more importantly) Are there any good resources that you know of
that explain how the query optimizer choices its query paths? Do the
"Inside SQL Server" books go into that much detail? Any good online
resources?

Thanks!
-Tom."Thomas R. Hummel" <tom_hummel@.hotmail.com> wrote in message
news:a2c0eeb8.0309160626.901177@.posting.google.com ...
> Hi,
> I was just helping a coworker optimize a query. He had two versions:
> one which used UNION for each value for which he was tallying results
> and another query which used GROUP BY. Here is an aproximation of what
> they were:
> Query #1:
> ---
> SELECT 12 AS [Row],
> ISNULL(SUM(CASE WHEN T.my_date BETWEEN @.week_start_date AND
> DATEADD(d, 1, @.week_start_date) THEN 1 ELSE 0 END), 0) AS [Monday],
> ISNULL(SUM(CASE WHEN T.my_date BETWEEN DATEADD(d, 1,
> @.week_start_date) AND DATEADD(d, 2, @.week_start_date) THEN 1 ELSE 0
> END), 0) AS [Tuesday]
> FROM My_Table T
> INNER JOIN Another_Table T2 ON T2.col1 = T.col1
> WHERE T.my_date BETWEEN @.week_start_date AND @.week_end_date
> AND T.col2 = 5
> UNION
> SELECT 13 AS [Row],
> ISNULL(SUM(CASE WHEN T.my_date BETWEEN @.week_start_date AND
> DATEADD(d, 1, @.week_start_date) THEN 1 ELSE 0 END), 0) AS [Monday],
> ISNULL(SUM(CASE WHEN T.my_date BETWEEN DATEADD(d, 1,
> @.week_start_date) AND DATEADD(d, 2, @.week_start_date) THEN 1 ELSE 0
> END), 0) AS [Tuesday]
> FROM My_Table T
> INNER JOIN Another_Table T2 ON T2.col1 = T.col1
> WHERE T.my_date BETWEEN @.week_start_date AND @.week_end_date
> AND T.col2 = 6
> Query #2:
> ---
> SELECT R.row_num AS [Row],
> ISNULL(SUM(CASE WHEN T.my_date BETWEEN @.week_start_date AND
> DATEADD(d, 1, @.week_start_date) THEN 1 ELSE 0 END), 0) AS [Monday],
> ISNULL(SUM(CASE WHEN T.my_date BETWEEN DATEADD(d, 1,
> @.week_start_date) AND DATEADD(d, 2, @.week_start_date) THEN 1 ELSE 0
> END), 0) AS [Tuesday]
> FROM My_Table T
> INNER JOIN Another_Table T2 ON T2.col1 = T.col1
> INNER JOIN Report_Rows R ON R.col2 = T.col2
> WHERE T.my_date BETWEEN @.week_start_date AND @.week_end_date
> GROUP BY ALL R.row_num
> ORDER BY R.row_num
> The Report_Rows table in this case would have had two rows mapping row
> 12 to a column value of 5 and row 13 to a column value of 6. The
> second query was performing horribly until I noticed the ALL keyword
> in the GROUP BY, which I didn't think was necessary. When I removed
> that it performed more like I expected it to perform.
> Before I had noticed that I was scouring over the query plans and
> couldn't figure out why in one instance the query optimizer chose to
> join My_Table and Another_Table, yet when the ALL keyword was there it
> chose to return all of the records from Another_Table (a rather large
> table) and join it to the Report_Rows table before then joining to
> My_Table, which had the date criteria in the WHERE clause.
> So, if you've read this far without giving up...
> 1. Why would the ALL keyword cause this? I understand the
> functionality of ALL, but I still don't see why that caused the
> reordering of the joins.
> 2. (more importantly) Are there any good resources that you know of
> that explain how the query optimizer choices its query paths? Do the
> "Inside SQL Server" books go into that much detail? Any good online
> resources?
> Thanks!
> -Tom.

It's almost impossible (at least for me) to know why the optimizer chose a
particular plan without knowing the table structures, indexes and amount of
data, and even with that knowledge, it may not be clear at all. So I can't
say much about your first question, but I can definitely recommend Inside
SQL Server 2000 for a great explanation of what the optimizer considers when
it produces a query plan. There's a lot of detail, including how to go about
using query plans to tune individual queries. Another useful book is
Advanced Transact SQL for SQL Server 2000, which also explains many of the
examples with reference to their query plans.

Simon|||Thomas R. Hummel (tom_hummel@.hotmail.com) writes:
> Before I had noticed that I was scouring over the query plans and
> couldn't figure out why in one instance the query optimizer chose to
> join My_Table and Another_Table, yet when the ALL keyword was there it
> chose to return all of the records from Another_Table (a rather large
> table) and join it to the Report_Rows table before then joining to
> My_Table, which had the date criteria in the WHERE clause.

I can only echo Simon's reply that without table definitions etc, this
is difficult to tell. In fact, even with all information available,
this might be difficult to tell. Understanding the output of a cost-
based optimizer is by no means an easy task.

> 2. (more importantly) Are there any good resources that you know of
> that explain how the query optimizer choices its query paths? Do the
> "Inside SQL Server" books go into that much detail? Any good online
> resources?

Certainly, you learn a great deal from Kalen's book. But I also like to
add that that experience counts a lot too. And some creative thinking.
The basic thing to understand is why a table scan may be better than
an index seek. This is something which also can be extended to joins.
That is a scan + merge/hash join may be faser than seek + loop join.

But then there are all such wild things which includes parallelism that
I find myself understanding only fragments of.

--
Erland Sommarskog, SQL Server MVP, sommar@.algonet.se

Books Online for SQL Server SP3 at
http://www.microsoft.com/sql/techin.../2000/books.asp|||Thank you both for the input. I had tried to duplicate the effect with
test tables, but as you know, the query optimizer takes a lot into
account and I couldn't find an example that would be practical for
posting here.

I will give the two books that Simon suggested a more thorough read.
I've seen adverts for Kalen's online webinars as well, so perhaps I'll
look into those.

> But then there are all such wild things which includes parallelism that
> I find myself understanding only fragments of.
I hate to think of something that is complex enough that you have
trouble understanding it Erland... ;-)

Thanks again!
-Tom.

No comments:

Post a Comment