I was thinking the NATURAL JOIN
example you just used
SELECT * FROM my_table NATURAL JOIN (
SELECT group_col, MAX(sort_col) sort_col
FROM my_table
GROUP BY group_col
) t
If you shift to another type of JOIN
and impose WHERE
, ordering can come and go without warning in spite of the ill-advised reliance on undocumented behavior of the GROUP BY
.
For this example, I will
- use Windows 7
- use MySQL 5.5.12-log for Windows
- create some sample data
- impose a
LEFT JOIN
without a WHERE
clause
- impose a
LEFT JOIN
with a WHERE clause
For the DB Environment
mysql> select version();
+------------+
| version() |
+------------+
| 5.5.12-log |
+------------+
1 row in set (0.00 sec)
mysql> show variables like '%version_co%';
+-------------------------+------------------------------+
| Variable_name | Value |
+-------------------------+------------------------------+
| version_comment | MySQL Community Server (GPL) |
| version_compile_machine | x86 |
| version_compile_os | Win64 |
+-------------------------+------------------------------+
3 rows in set (0.00 sec)
mysql>
Using this script to generate sample data
DROP DATABASE IF EXISTS eggyal;
CREATE DATABASE eggyal;
USE eggyal
CREATE TABLE groupby
(
id int not null auto_increment,
num int,
primary key (id)
);
INSERT INTO groupby (num) VALUES
(floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp())),
(floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp())),
(floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp())),
(floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp()));
INSERT INTO groupby (num) SELECT num FROM groupby;
SELECT * FROM groupby;
and these two queries for testing the GROUP BY
subsequent use;
SELECT * FROM groupby A LEFT JOIN
(
SELECT num, MAX(id) id
FROM groupby
GROUP BY num
) B USING (id);
SELECT * FROM groupby A LEFT JOIN
(
SELECT num, MAX(id) id
FROM groupby
GROUP BY num
) B USING (id) WHERE B.num IS NOT NULL;
Let's test the durability of the GROUP BY
's results;
STEP 01 : Create the Sample Data
mysql> DROP DATABASE IF EXISTS eggyal;
Query OK, 1 row affected (0.09 sec)
mysql> CREATE DATABASE eggyal;
Query OK, 1 row affected (0.00 sec)
mysql> USE eggyal
Database changed
mysql> CREATE TABLE groupby
-> (
-> id int not null auto_increment,
-> num int,
-> primary key (id)
-> );
Query OK, 0 rows affected (0.07 sec)
mysql> INSERT INTO groupby (num) VALUES
-> (floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp())),
-> (floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp())),
-> (floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp())),
-> (floor(rand() * unix_timestamp())),(floor(rand() * unix_timestamp()));
Query OK, 8 rows affected (0.06 sec)
Records: 8 Duplicates: 0 Warnings: 0
mysql> INSERT INTO groupby (num) SELECT num FROM groupby;
Query OK, 8 rows affected (0.05 sec)
Records: 8 Duplicates: 0 Warnings: 0
mysql> SELECT * FROM groupby;
+----+------------+
| id | num |
+----+------------+
| 1 | 269529129 |
| 2 | 387090406 |
| 3 | 1126864683 |
| 4 | 411160755 |
| 5 | 29173595 |
| 6 | 266349579 |
| 7 | 1244227156 |
| 8 | 6231766 |
| 9 | 269529129 |
| 10 | 387090406 |
| 11 | 1126864683 |
| 12 | 411160755 |
| 13 | 29173595 |
| 14 | 266349579 |
| 15 | 1244227156 |
| 16 | 6231766 |
+----+------------+
16 rows in set (0.00 sec)
STEP 02 : Use LEFT JOIN
without a WHERE
clause
mysql> SELECT * FROM groupby A LEFT JOIN
-> (
-> SELECT num, MAX(id) id
-> FROM groupby
-> GROUP BY num
-> ) B USING (id);
+----+------------+------------+
| id | num | num |
+----+------------+------------+
| 1 | 269529129 | NULL |
| 2 | 387090406 | NULL |
| 3 | 1126864683 | NULL |
| 4 | 411160755 | NULL |
| 5 | 29173595 | NULL |
| 6 | 266349579 | NULL |
| 7 | 1244227156 | NULL |
| 8 | 6231766 | NULL |
| 9 | 269529129 | 269529129 |
| 10 | 387090406 | 387090406 |
| 11 | 1126864683 | 1126864683 |
| 12 | 411160755 | 411160755 |
| 13 | 29173595 | 29173595 |
| 14 | 266349579 | 266349579 |
| 15 | 1244227156 | 1244227156 |
| 16 | 6231766 | 6231766 |
+----+------------+------------+
16 rows in set (0.00 sec)
mysql>
STEP 03 : Use LEFT JOIN
with a WHERE
clause
mysql> SELECT * FROM groupby A LEFT JOIN
-> (
-> SELECT num, MAX(id) id
-> FROM groupby
-> GROUP BY num
-> ) B USING (id) WHERE B.num IS NOT NULL;
+----+------------+------------+
| id | num | num |
+----+------------+------------+
| 16 | 6231766 | 6231766 |
| 13 | 29173595 | 29173595 |
| 14 | 266349579 | 266349579 |
| 9 | 269529129 | 269529129 |
| 10 | 387090406 | 387090406 |
| 12 | 411160755 | 411160755 |
| 11 | 1126864683 | 1126864683 |
| 15 | 1244227156 | 1244227156 |
+----+------------+------------+
8 rows in set (0.00 sec)
mysql>
ANALYSIS
Looking at the aforementioned results, here are two questions:
- Why does a
LEFT JOIN
keep an ordering by id
?
- Why in the world did using a
WHERE
impose a reordering ?
- Was it during the JOIN phase ?
- Did the Query Optimizer look ahead at the ordering of the subquery or ignore it ?
No one foresaw any of these effects because the behavior of explicit clauses was relied upon by the implicit behavior of the Query Optimizer.
CONCLUSION
From my perspective, corner cases can only be of an external nature. In light of this, developers must be willing to fully evaluate the results of a GROUP BY
in conjunction with the following twelve(12) aspects:
- aggregate functions
- subquery usage
JOINs
clauses
WHERE
clauses
- sort order of results with no explicit
ORDER BY
clause
- query results using older GA releases of MySQL
- query results using newer beta releases of MySQL
- the current SQL_MODE setting in
my.cnf
- the operating system the code was compiled for
- possibly the size of join_buffer_size with respect to its effect on the Query Optimizer
- possibly the size of sort_buffer_size with respect to its effect on the Query Optimizer
- possibly the storage engine being used (MyISAM vs InnoDB)
Here is the key thing to remember : Any instance of MySQL that works for your query in a specific environment is itself a corner case. Once you change one or more of the twelve(12) evaluation aspects, the corner case is due to break, especially given the first nine(9) aspects.
I initially thought you were on to something here. Working assumption was along the lines that perhaps the buffer pool wasn't immediately flushed as it requires "some work" to do so and why bother until the memory was required. But...
Your test is flawed.
What you're seeing in the buffer pool is the pages read as a result of re-attaching the database, not the remains of the previous instance of the database.
And we can see that the buffer pool was not totally blown away by the
detach/attach. Seems like my buddy was wrong. Does anyone disagree or
have a better argument?
Yes. You're interpreting physical reads 0
as meaning there were not any physical reads
Table 'DatabaseLog'. Scan count 1, logical reads 782, physical reads
0, read-ahead reads 768, lob logical reads 94, lob physical reads 4,
lob read-ahead reads 24.
As described on Craig Freedman's blog the sequential read ahead mechanism tries to ensure that pages are in memory before they're requested by the query processor, which is why you see zero or a lower than expected physical read count reported.
When SQL Server performs a sequential scan of a large table, the
storage engine initiates the read ahead mechanism to ensure that pages
are in memory and ready to scan before they are needed by the query
processor. The read ahead mechanism tries to stay 500 pages ahead of
the scan.
None of the pages required to satisfy your query were in memory until read-ahead put them there.
As to why online/offline results in a different buffer pool profile warrants a little more idle investigation. @MarkSRasmussen might be able to help us out with that next time he visits.
Best Answer
This might happen if the access path to retrieve the data happens to produce data already ordered e.g. an index scan in key order. Or if the plan contains an explicit sort in order to provide data in order for operators that require this, such as merge join or stream aggregate. Without an explicit
order by
no order is guaranteed but SQL Server won't go out of its way to prevent results being output in order either.For the first query you can see different ordering if SQL Server uses an allocation order scan.
So make sure the table is at least 64 pages in size and that the allocation order isn't the same as key order then run the query at read uncommitted isolation level.
The read uncommitted isn't required for the above in the case that a table lock is taken, there is a hint for that but a less obvious way would be to change the above table definition to
And then hopefully you will see disordered results without having to change the isolation level.
For the second query you want a hash aggregate rather than a stream aggregate. You can force this with a query hint or less obviously with a plan guide but this is more likely to be chosen organically if there are relatively few distinct groups compared to the size of the table.