On Caching and Evangelizing SQL

http://www.oracle.com/technetwork/issue-archive/2011/11-sep/o51asktom-453438.html

 

Our technologist caches scalar subqueries, votes for SQL, and recommends technology and community.

One of the talks I gave recently at the Oracle Benelux User Group (OBUG) conference in Belgium was regarding techniques—some SQL tricks, if you will—you can use when writing SQL. One of the techniques I discussed at length was scalar subqueries and how they can be used to minimize the number of times a PL/SQL function is called from SQL. This is important because the overhead of going from SQL to PL/SQL over and over again can be quite expensive, and the scalar subquery caching database feature can dramatically decrease this overhead.

What exactly is a scalar subquery? It is a subquery in a SQL statement that returns exactly one column and zero rows or one row. That single column can be a complex object type, so it can consist of many attributes, but the subquery returns a single scalar value (or NULL if the subquery returns zero records). A scalar subquery can be used anywhere a literal could have been used. For example, you can obviously create the query

select deptno, dname,
      'Hello world'
 from dept;

You can also (since Oracle8i Database Release 8.1.5, anyway) include a scalar subquery:

select deptno, dname,
      (select count(*) 
         from emp 
        where emp.deptno = dept.deptno) 
  from dept;

Note that the complete last query is semantically equivalent to

select dept.deptno, dept.dname,
count(emp.empno)
from dept left outer join emp
on (dept.deptno = emp.deptno )
group by dept.deptno, dept.dname;

Regarding these last two queries, if your goal is to optimize your query for initial response time, you may want to use the former query, with the scalar subquery, because Oracle Database would have to get just the first row from DEPT, run the scalar subquery (select count...) against EMP to get the count, and then return it. The process for returning that first row would be very fast. The latter query, which includes an outer join, would likely require two full table scans, a hash outer join, and an aggregation—and require all these processes to be completed before the first row could be returned.

So, now that you know what a scalar subquery is, you need to understand the scalar subquery caching feature of Oracle Database. In the above query with a scalar subquery, the database would actually extract the scalar subquery and rewrite it to use bind variables. In effect, the database would be executing

(select count(*) from emp
where emp.deptno = ?)

for each row in the DEPT table. Given that DEPTNO is unique in DEPT, the database would have to physically execute that scalar subquery for each row in DEPT. But what if you are not querying the DEPT table but, rather, another table that includes a DEPTNO column? Perhaps a PROJECTS table with records that have PROJECT_NAME and DEPTNO columns? Then a query such as

select project_name, deptno,
      (select count(*)
         from emp
        where emp.deptno =
              projects.deptno)
  from projects;

will have to execute the scalar subquery at least as many times as there are unique DEPTNO values in the PROJECTS table. Note that I said, “at least as many times”—the database does not have to execute the scalar subquery for every row in the PROJECTS table if it caches some of the results and reuses them, which is what happens with scalar subquery caching.

When you’re using a scalar subquery, Oracle Database will set up a small in-memory hash table for the subquery and its results each time it runs the query. So, when you run the previous query, Oracle Database sets up in memory a hash table that looks like this:

Select count(*) from emp where emp.deptno = :deptno
:deptno	Count(*)

Oracle Database will use this hash table to remember the scalar subquery and the inputs to it—just :DEPTNO in this case—and the output from it. At the beginning of every query execution, this cache is empty, but suppose you run the query and the first PROJECTS row you retrieve has a DEPTNO value of 10. Oracle Database will assign the number 10 to a hash value between 1 and 255 (the size of the hash table cache in Oracle Database 10g and Oracle Database 11g currently) and will look in that hash table slot to see if the answer exists. In this case, it will not, so Oracle Database must run the scalar subquery with the input of 10 to get the answer. If that answer (count) is 42, the hash table may look something like this:

Select count(*) from emp where emp.deptno = :deptno
:deptno	Count(*)

You’ll have saved the DEPTNO value of 10 and the answer (count) of 42 in some slot—probably not the first or last slot, but whatever slot the hash value 10 is assigned to. Now suppose the second row you get back from the PROJECTS table includes a DEPTNO value of 20. Oracle Database will again look in the hash table after assigning the value 20, and it will discover “no result in the cache yet.” So it will run the scalar subquery, get the result, and put it into the hash table cache. Now the cache may look like this:

Select count(*) from emp where emp.deptno = :deptno
:deptno	Count(*)
…	…
10	42

Now suppose the query returns a third row and it again includes a DEPTNO value of 10. This time, Oracle Database will see DEPTNO = 10, find that it already has that value in the hash table cache, and will simply return 42 from the cache instead of executing the scalar subquery. In fact, it will never have to run that scalar subquery for the DEPTNO values of 10 or 20 again for that query—it will already have the answer.

What happens if the number of unique DEPTNO values exceeds the size of the hash table? What if there are more than 255 values? Or, more generally, if more than one DEPTNO value is assigned to the same slot in the hash table, what happens in a hash collision?

The answer is the same for all these questions and is rather simple: Oracle Database will not be able to cache the second or nth value to that slot in the hash table. For example, what if the third row returned by the query contains the DEPTNO = 30 value? Further, suppose that DEPTNO = 30 is to be assigned to exactly the same hash table slot as DEPTNO = 10. The database won’t be able to effectively cache DEPTNO = 30 in this case—the value will never make it into the hash table. It will, however, be “partially cached.” Oracle Database still has the hash table with all the previous executions, but it also keeps the last scalar subquery result it had “next to” the hash table. That is, if the fourth row also includes a DEPTNO = 30 value, Oracle Database will discover that the result is not in the hash table but is “next to” the hash table, because the last time it ran the scalar subquery, it was run with an input of 30. On the other hand, if the fourth row includes a DEPTNO = 40 value, Oracle Database will run the scalar subquery with the DEPTNO = 40 value (because it hasn’t seen that value yet during this query execution) and overwrite the DEPTNO = 30 result. The next time Oracle Database sees DEPTNO = 30 in the result set, it’ll have to run that scalar subquery again.

So, all this discussion so far was a setup—a prelude, if you will—for what I really wanted to write about: how to reduce the number of times a PL/SQL function invoked from SQL is called.

Suppose you have a PL/SQL function like the one in Listing 1. This function will simply increment a counter each time it is called. The counter value will be stored in the CLIENT_INFO column of V$SESSION, and the function will then return the length of its input.

Code Listing 1: PL/SQL function f

SQL> create or replace function f( x in varchar2 )
     return number
  2  as
  3  begin
  4     dbms_application_info.set_client_info
        (userenv('client_info')+1 );
  5     return length(x);
  6  end;
  7  /
Function created.
 

If you execute a simple query like the one in Listing 2 (note that the STAGE table is simply a copy of ALL_OBJECTS), you can see that the function was called once per row even though the inputs to the function were repeated over and over again. If you use a scalar subquery—if you replace “f(owner)” with “(select f(owner) from dual)”—you will notice a massive reduction in calls to the function, as shown in Listing 3.

Code Listing 2: Demonstrating once-per-row repetition in function f

SQL> begin
   2   :cpu := dbms_utility.get_cpu_time;
   3   dbms_application_info.set_client_info(0);
   4 end;
   5 /
PL/SQL procedure successfully completed.

SQL> select owner, f(owner) from stage;
...
72841 rows selected.

SQL> select
  2  dbms_utility.get_cpu_time-:cpu cpu_hsecs,
  3  userenv('client_info')
  4  from dual;

 CPU_HSECS   USERENV('CLIENT_INFO')
 ---------   -----------------------
       118                     72841

Code Listing 3: Reducing function calls with scalar subquery

SQL> begin
   2   :cpu := dbms_utility.get_cpu_time;
   3   dbms_application_info.set_client_info(0);
   4 end;
   5 /
PL/SQL procedure successfully completed.

SQL> select owner, (select f(owner) from dual) f from stage;
...
72841 rows selected.

SQL> select
  2  dbms_utility.get_cpu_time-:cpu cpu_hsecs,
  3  userenv('client_info')
  4  from dual;

 CPU_HSECS   USERENV('CLIENT_INFO')
 ---------   -----------------------
        29                        66

As you can see in Listing 3, the function went from 72,841 calls down to 66! And the CPU time dropped dramatically as well. That reduction in CPU time resulted from not calling the function (and therefore not making the context switch from SQL to PL/SQL and not invoking DBMS_APPLICATION_INFO and LENGTH so often).

Now, some of you might be thinking about the possibility of marking the function deterministic, because it is, in fact, deterministic. Wouldn’t that reduce the number of function calls as well? The short answer is “yes,” but the longer answer is “yes, but not as well as scalar subquery caching can.”

As you can see in the deterministic function in Listing 4, the number of calls is reduced, but only down to 8,316. The scalar subquery cache in this case was better than just marking the function as deterministic. (Note: Marking a function deterministic affects caching only in Oracle Database 10gand above; you will see no change in the number of calls in Oracle9i Database with deterministic functions.)

Code Listing 4: Deterministic function variation

SQL> create or replace function f( x in varchar2 )
     return number
  2  DETERMINISTIC
  3  as
  4  begin
  5    dbms_application_info.set_client_info
       (userenv('client_info')+1 );
  6    return length(x);
  7  end;
  8  /
Function created.

SQL> begin
   2   :cpu := dbms_utility.get_cpu_time;
   3   dbms_application_info.set_client_info(0);
   4 end;
   5 /
PL/SQL procedure successfully completed.

SQL> select owner, f(owner) from stage;
...
72841 rows selected.

SQL> select
  2  dbms_utility.get_cpu_time-:cpu cpu_hsecs,
  3  userenv('client_info')
  4  from dual;

 CPU_HSECS   USERENV('CLIENT_INFO')
 ---------   -----------------------
        69                      8316

Going one step further, you might ask, “But what if you used the function result cache in Oracle Database 11g? Could you get the function calls down to zero?” The answer is, “Yes, sort of.” The function calls would go down to zero, but the context switching from SQL to PL/SQL would remain at a very high value: 72,841 times in this case.

Listing 5 shows a function result cache solution. The number of function calls is 32 (because I happen to have 32 schemas in my database), and the CPU time is 0.73 seconds—about the same as that of the DETERMINISTIC function and far above the CPU time of 0.29 seconds used by the 66 calls made by the scalar subquery.

Code Listing 5: Function result cache solution

SQL> create or replace function f( x in varchar2 )
     return number
  2  RESULT_CACHE
  3  as
  4  begin
  5    dbms_application_info.set_client_info
       (userenv('client_info')+1 );
  6    return length(x);
  7  end;
  8  /
Function created.

SQL> begin
   2   :cpu := dbms_utility.get_cpu_time;
   3   dbms_application_info.set_client_info(0);
   4 end;
   5 /
PL/SQL procedure successfully completed.

SQL> select owner, f(owner) from stage;
...
72841 rows selected.

SQL> select
  2  dbms_utility.get_cpu_time-:cpu cpu_hsecs,
  3  userenv('client_info')
  4  from dual;

 CPU_HSECS   USERENV('CLIENT_INFO')
 ---------   -----------------------
        73                        32

Furthermore, if you were to run this query again—as shown in Listing 6—you’d discover that it calls the function zero times, but the CPU time (0.63 seconds) is still longer than the CPU time for the scalar subquery cache example in Listing 3.

Code Listing 6: Rerunning the function result cache solution

SQL> begin
   2   :cpu := dbms_utility.get_cpu_time;
   3   dbms_application_info.set_client_info(0);
   4 end;
   5 /
PL/SQL procedure successfully completed.

SQL> select owner, f(owner) from stage;
...
72841 rows selected.

SQL> select
  2  dbms_utility.get_cpu_time-:cpu cpu_hsecs,
  3  userenv('client_info')
  4  from dual;

 CPU_HSECS   USERENV('CLIENT_INFO')
 ---------   -----------------------
        63                         0

This just shows that even if your function is deterministic, even if it is “result-cached,” there is a good reason to always wrap your function call in a SELECT FROM DUAL. I’ve gotten into the habit over the years of never coding

select *
  from t
 where column = plsql_function(..);

but rather including a simple scalar subquery

select *
  from t
 where column =
    (select plsql_function(...)
       from dual);

to get the performance benefit of the scalar subquery cache.

Is the Cursor Dying?

With subquery factoring, MODEL, PIVOT, MERGE with DELETE, “Materialize” and other hints, global temporary tables, DML (data manipulation language) error logging, and so many other powerful new SQL features, when is it necessary to use cursors (apart from using ref cursors for reports). Is the cursor dying?

Well, the simple UPDATE statement

update t set x = 5;

requires a cursor. In fact, every SELECT you execute, every DELETE you perform—pretty much everything—requires a cursor. So it is doubtful that cursors are “dying.”

However, I think you may mean, “Is the need to procedurally process the output of a cursor dying?” Many times (many, many, many times), a ton of procedural code can be replaced with a single SQL statement. That has been true for a long time and is truer today than ever with features such as DML error logging.

So, yes, we should all be striving to erase as much procedural code as possible and go set-based. Does it mean the death of cursors? No, not at all. It means the reduction of procedural code—more code = more bugs; less code = fewer bugs.

But it also means that people have to read the documentation, understand the benefits, learn the full SQL language, and start using it.

I was once asked in a question-and-answer session what I thought was the most underutilized Oracle Database feature. I responded almost immediately with the answer “SQL.” The questioner then asked me, “But what do you mean—everyone uses SQL all the time.” My response was that everyone uses very, very simple SQL and avoids 99 percent of its actual capabilities.

Sequence Caching

I read about sequence caching, but I still do not understand the CACHE and NOCACHE settings. Could you explain CACHE and NOCACHE with an example?

Sure. In short, if you use NOCACHE, every time you call sequence.nextval, you will cause a table named SYS.SEQ$ to be updated and committed. Every single time. If, on the other hand, you use the default value, CACHE 20, you will cause SYS.SEQ$ to be updated and committed every 20th time you call sequence.nextval.

It is a performance thing. If you don’t cache the next couple of sequence values in the system global area (SGA), you will be doing an UPDATE and COMMIT every single time you call nextval. You can evaluate the performance impact this might have on your application—especially with large data loads—if the cache value is set too low. For example, let’s create three sequences

SQL> create sequence
nocache_seq NOCACHE;
Sequence created.

SQL> create sequence
default_cache_seq;
Sequence created.

SQL> create sequence
big_cache_seq cache 100000;
Sequence created.

and a table to test with

SQL> create table t ( x int );
Table created.
 

Next Steps

ASK Tom
Tom Kyte answers your most difficult technology questions. Highlights from that forum appear in this column.

READ more Tom
Oracle Database Concepts 11g Release 2 (11.2)
Expert Oracle Database Architecture: Oracle Database Programming 9i, 10g, and 11gTechniques and Solutions, Second Edition
Effective Oracle By Design
The Tom Kyte Blog

DISCUSS 
Oracle Database, Express Edition

FOLLOW Oracle Database
on Twitter
on Facebook
index scans

DOWNLOAD
Oracle Database 11g Release 2
Oracle Database, Express Edition 11g (Beta)

Now let’s execute the following block of code with SQL trace enabled for each of the three sequences—replace &1 with the actual sequence name (nocache_seq, default_cache_seq, or big_cache_seq) as you run them:

SQL> truncate table t;
Table truncated.

SQL> insert into t (x)
  2  with data(x)
  3  as
  4  (select 1 x
  5     from dual
  6   union all
  7   select x+1
  8     from data
  9    where x <= 100000)
 10  select &1.nextval
 11    from data
 12  /

100001 rows created.

The resulting TKPROF reports are pretty telling. For the NOCACHE sequence (nocache_seq), you see

update seq$
   set ...
 where obj#=:1

call     count  cpu   elapsed
---      -----  ----  -------
Parse        1  0.00     0.00
Execute 100001 33.22    40.23
Fetch        0  0.00     0.00
---      -----  ----  -------
total   100002 33.22    40.23

It took 40 seconds to update the SYS.SEQ$ table! Every time you needed a new sequence, you had to do an UPDATE and COMMIT. The total elapsed time in the TKPROF report for all the statements executed (not shown) was 54 seconds, and most of that time was spent maintaining the sequence.

Using the default cache sequence (default_cache_seq), you see very different numbers:

call     count  cpu   elapsed
---      -----  ----  -------
Parse        1  0.00     0.00
Execute   5001  1.60     1.89
Fetch        0  0.00     0.00
---      -----  ----  -------

total     5002  1.60     1.89

You are down to about 2 seconds of runtime for the sequence operation, and the total runtime (not shown) for all the statements executed was about 4.5 seconds (yes, 4.5 seconds).

Using the big cache sequence (big_cache_seq), you see

call     count  cpu   elapsed
---      -----  ----  -------
Parse        1  0.00     0.00
Execute      2  0.00     0.00
Fetch        0  0.00     0.00
---      -----  ----  -------

total        3  0.00     0.00

You basically spent no time maintaining the SYS.SEQ$ table, and the total runtime (not shown) for all the statements executed was around 2.1 seconds, according to the TKPROF report.