Simple MySQL Range Scan Test

I was trying to tune a MySQL query this week. I ran the same query against Oracle with the same data and got a much faster runtime on Oracle. I couldn’t get MySQL to do a range scan on the column that Oracle was doing it on. So, I just started barely scratching the surface with a simple test of when MySQL will use an index versus a full table scan in a range query. In my test MySQL always uses an index except on extreme out of range conditions. This is funny because in my real problem query it was the opposite. But I might as well document what I found for what it’s worth. I haven’t blogged much lately.

Here is my testcase and its output:

This is on 8.0.26 as part of an AWS Aurora MySQL RDS instance with 2 cores and 16 gigabytes of RAM.

I created a simple test table and put 10485760 rows in it:

create table test 
b integer,

The value of b is always 1 and a ranges from 1 to 10878873.

This query uses a range query using the index:

a > -2147483648;

This query uses a full table scan:

a > -2147483649;

The full scan is slightly faster.

Somehow when you are 2147483650 units away from the smallest value of a the MySQL optimizer suddenly thinks you need a full scan.

There are a million more tests I could do like things with a million variables, but I thought I might as well put this out there. I’m not really any the wiser but it is a type of test that might be worth mentioning.


P.S. I did find this document:

-2147483648 is the smallest value for a 4 byte int type so that is probably why the behavior changes at -2147483649. Not sure what that information is worth!

P.P.S I think this explains why -2147483649 leads to the full scan in this case maybe:

“Beginning with MySQL 8.0.16, comparisons of columns of numeric types with constant values are checked and folded or removed for invalid or out-of-rage values”

Since -2147483648 is the smallest value for an int then with -2147483649 the optimizer removes the condition.

Posted in Uncategorized | 3 Comments

Archivelog Space Needed Python Script

I wrote a script called to help size our Oracle archive log filesystems to support replication tools such as Fivetran, DMS, or GoldenGate which need a certain number of hours or days of archive log history at all times. In many cases we had backups that removed all the archive logs after they had been backed up once. So, that is essentially zero hours of history. If we only backed up once per day, it really peaked out at 24 hours of history, but the minimum was zero. Our replication products need 24 hours minimum in some cases. In other cases, we needed different numbers of hours. Also, the log backups and deletes run multiple times a day and on different schedules on some systems.

I based on a simplistic idea which I know is not perfect but so far it has been helpful. So, I thought I would share it here in case others can benefit. I would love any feedback, suggestions and criticism of the approach and implementation. The idea was to query V$ARCHIVED_LOG and see how full the filesystem would get if these same logs came in at the same times but with different retention times and different archive log filesystem sizes. I could try different settings and see the highest percentage that the filesystem hit.

I think this approach is imperfect because the past behavior recorded in V$ARCHIVED_LOG may not represent future behavior very well as things change. Also, the way I wrote it I assumed that the archive logs are laid down instantaneously. I.e. If the row in V$ARCHIVED_LOG has COMPLETION_TIME of 2/24/2024 16:11:15 then I assume that the filesystem gets (BLOCKS+1)*BLOCK_SIZE bytes fuller in that second. I also assume that the backups instantly remove all the logs which are beyond the retention.

I tested the script using the current settings for the archive log filesystem size and log backup run times and retention settings. I just compared the output percent full with reality. It was close but not exact. So, when I went to apply this for real, I padded the filesystem sizes so the expected percent full was less than 50%. So far so good in reality. I would like to build an emergency script that automatically clears out old logs if the filesystem gets full but so far, I have not. We do have alerting on archive log filesystem space getting too full.

If you run the script, you can see the arguments:

Arguments: oracle-username oracle-password tns-name configfile-name

Config file is text file with this format:

archivelog filesystem size in bytes
number of backups per day
one line per backup with 24-hour:minutes:seconds archivlog-retention-hours

for example:

02:15:00 168
06:15:00 168
10:45:00 168
14:15:00 168
18:15:00 168
22:15:00 168

The output is something like this:

2024-02-20 08:55:57 1.67%  
2024-02-20 09:10:02 1.68%  
2024-02-20 10:00:29 1.69%  
2024-02-20 11:00:20 1.7%  
2024-02-20 11:37:32 1.7%  
2024-02-20 12:01:17 1.68%  
2024-02-20 12:09:05 1.68%  
2024-02-20 12:43:53 1.69%  
2024-02-20 12:55:52 1.69%  
Max percent used archivelog filesystem: 46.15%
Date and time of max percent: 2023-12-24 11:52:17

When your archive log filesystem is too small the Max percent is over 100%.

It’s not perfect or fancy but it is available if someone finds it useful.


p.s. The script uses cx_Oracle so you will need to install that.

Posted in Uncategorized | 5 Comments

User Privileges Script

I keep running into situations on Oracle databases where I need to dump out the privileges an Oracle user has. I have been just manually putting together SQL statements like:

select * from dba_role_privs where grantee='MYUSER';
select * from dba_sys_privs where grantee='MYUSER';
select * from dba_tab_privs where grantee='MYUSER';
select * from dba_users where username='MYUSER';

This captures the three kinds of grants the user could have in the first three queries and the last query just shows if the user exists and things like whether it is locked. Really this simple set of queries is good enough in most cases.

But I had also wrote a script that would show all the system and object grants that were included in the roles. Because you can have roles granted to roles, you must loop through all the roles until you get down to the base system and object privileges. I rewrote this logic from scratch several times until I finally convinced myself to make a script and save it on my GitHub site. The current version of the script is here:


The interesting part of the script is where we keep looping through the roles in table my_role_privs deleting each role and then inserting the role’s system, object, and role privileges into the my_sys_privs, my_tab_privs, and my_role_privs tables. Eventually you run out of roles to delete and the loop finishes. I guess this works because you cannot have a circular role grant situation:

SQL> create role a;

Role created.

SQL> create role b;

Role created.

SQL> grant a to b;

Grant succeeded.

SQL> grant b to a;
grant b to a
ERROR at line 1:
ORA-01934: circular role grant detected

In the past I have put a loop counter in the code just in case there was something circular or a really long chain of roles, but this version does not have it.

To make the output useful I put it in three sections. The first section just has the direct grants and corresponds to the first three queries listed above.

Privileges granted directly to user MYUSER
Role privileges for user MYUSER

System privileges for user MYUSER

Summarized table privileges for user MYUSER

OWNER                PRIVILEGE                                  COUNT(*)
-------------------- ---------------------------------------- ----------
SYS                  EXECUTE                                           1
Detailed table privileges for user MYUSER

PRIVILEGE                                OWNER                TABLE_NAME
---------------------------------------- -------------------- -----------
EXECUTE                                  SYS                  DBMS_RANDOM

I put counts of each type of object grants in case there was a bunch. I called them table privileges because view is named dba_tab_privs but I really should have called them object privileges because they can be grants on objects which are not tables.

The second section has the output of the loop showing all the system and object privileges implied by the role grants as well as those granted directly to the user:

Privileges granted through a role or directly to user MYUSER
System privileges for user MYUSER

Summarized table privileges for user MYUSER

OWNER                PRIVILEGE                                  COUNT(*)
-------------------- ---------------------------------------- ----------
AUDSYS               EXECUTE                                           1
GSMADMIN_INTERNAL    EXECUTE                                           1
OUTLN                SELECT                                            3
SYS                  DELETE                                           11
SYS                  EXECUTE                                         169
SYS                  FLASHBACK                                        14
SYS                  INSERT                                           12
SYS                  READ                                             15
SYS                  SELECT                                         4759

Detailed table privileges for user MYUSER

PRIVILEGE                                OWNER                TABLE_NAME
---------------------------------------- -------------------- ------------------------
DELETE                                   SYS                  AUX_STATS$
DELETE                                   SYS                  DBA_REGISTRY_SQLPATCH
DELETE                                   SYS                  EXPIMP_TTS_CT$
DELETE                                   SYS                  INCEXP
DELETE                                   SYS                  INCFIL

I use this a lot of times to see if a user has CREATE SESSION either directly or through a role so that I will know whether the user can login.

Lastly, I included a couple of details about the user at the end:

Account status, last password change for user ZBL6050

-------------------------------- -------------------
OPEN                             2023-10-10 11:01:01

You need to give the user that runs userprivs.sql SELECT on sys.user$ to get the last password changed date and time. Otherwise, this query returns an error.

I mainly use this script to validate if a user has the correct permissions and if they can log in, so putting this information at the end in addition to the grant information above just fills in some details I would have to query anyway. I.e., Is the user locked? How long since they changed their password?

I thought about bringing down some statement about CREATE SESSION here. As it is written now, I have to visually scan the system privileges for CREATE SESSION to get the full picture on the user’s ability to login. It might be nice to add a column “Has CREATE SESSION” to this screen.

There are probably fancier scripts and tools to do all this, but this is what I have been using and the reasoning behind it. Maybe it will be useful to others and a reminder to myself to document it here.


Posted in Uncategorized | Leave a comment

MongoDB Atlas LangChain Vector Store

I was in a short programming contest at work for three days last week. My team got second place! We used a LangChain vector store in a MongoDB Atlas cluster so I thought I would at least document the links we referred to and videos I watched for others who are interested.

First I watched this video about LangChain:

I recommend watching the whole thing but the part about VectorStores starts at 25:22.

I got Python 3.9 setup in an Amazon EC2 instance and ran through these steps with the FAISS database:

Then I watched this video about the new MongoDB Atlas Vector Search feature:

This video is totally worth watching. I got a ton out of it. After watching the video I redid the VectorStore example but with MongoDB Atlas as the database:

I got in a discussion with ChatGPT about why they call them “vectors” instead of “points”. A vector is just an array or list of floating point numbers. In math this could be a point in some multi-dimensional space. ChatGPT didn’t seem to realize that software does use the vectors as vectors in a math sense sometimes. The MongoDB Atlas index we used cosine similarity which must be related to the vectors with some common starting point like all zeroes pointing towards the point represented by the list of numbers in the “vector”.

When I created the search index in MongoDB Atlas I forgot to name it and it did not work since the code has the index name. For the sample the index name has to be langchain_demo. By default index name is “default”.

LangChain itself was new to me. I watched the first video all the way through but there is a lot I did not use or need. I had played with OpenAI in Python already following the Python version of this quick start:

I edited the example script and played with different things. But I had never tried LangChain which sits on top of OpenAI and simplifies and expands it.

The project we worked on for the contest implemented the architecture documented at 28:57 in the MongoDB video above. If you look at the MongoDB Atlas vector store example this “information flow” would take the output from docsearch.similarity_search(query) and send it through OpenAI to summarize. If you take the piece of the President’s speech that is returned by the similarity search and past it into OpenAI’s playground the result looks like this:

So, our programming project involved pulling in documents that were split up into pieces and then retrieve a piece based on a similarity query using the vector store and then ran that piece through OpenAI to generate a readable English summary.

Posted in Uncategorized | Leave a comment

Install MySQL 5.7.38 from source on Oracle Linux 8

This is a follow up to my earlier post about installing various MySQL versions from source:

So, I am an Oracle database administrator. I have been for about 30 years starting with Oracle 7. But for the past few years we have been supporting MySQL in Amazon Web Services’ RDS version. Unfortunately, we have been forced to embrace the Aurora version of MySQL also which is proprietary to Amazon and does not have the full source code. But we still have several vanilla MySQL RDS instances that we support.

Working with Oracle for so many years I have tried to learn about its internals – how does it really work? This is difficult because I do not have access to Oracle’s source code and even if I did, I probably do not have enough years remaining in my life to ever fully understand it. Still, the idea of working with the MySQL community edition version that we have the full source code to has always intrigued me. This is similar to my fascination with the antique computer game Nethack that I play in a character-based mode exploring a dungeon and fighting monsters. It is a good game, but more importantly, I have the full source code.

Who knows? Maybe when we are working with MySQL we will run across a bug, and I will be able to dig into the part of the code that has the bug and it will help me figure out how to solve the problem. With Oracle it is always a guess what is going on. It can be an educated guess based on what Oracle reveals through various traces and logs. But maybe even though I do not have the time and motivation to be a hardcore MySQL internals developer there could be a situation where having the source code will help.

So, that leads me to want to download the exact version of the MySQL source code that we are using on AWS RDS and compile it, link it, install it on a test VM on the chance that someday I will need a working test MySQL database of the same version as one that is having a problem in production.

Things have changed since my 5/22/2019 post about setting up this kind of environment. At that time, I was working with an Oracle Linux 7 VM running on my work laptop with all its firewalls and Zscaler and all between my VM and the internet. Today I am using Oracle Linux 8 running on a VM on a personal laptop which is on my home network so there is nothing in the way of my downloading things from the internet like Linux rpm packages. One of the side effects of COVID-19 is that I am working from home full time. Also, not due to COVID, my middle daughter moved out leaving her bedroom open to be converted to an office. So, I have two “desks” setup with two laptops and big monitors on both. My home laptop sits to my left and I can run VirtualBox VMs on it without being on our corporate network. This is great for testing where I just need to setup a technology and I do not need or even want access to something on our corporate network.

So, with all this prelude let me tell you some of the things I had to do to get the MySQL 5.7.38 source code compiled on my Oracle Linux 8 VM.

I cloned a OEL 8 VM that I already had and then updated the packages. I think I used yum instead of dnf which was dumb, but it worked.

Once I had an up-to-date OEL 8 environment I had to get the source tree for MySQL with the right commit point for 5.7.38. I was following this document:

I ran these git commands:

git clone

git checkout 5.7

git pull

This gets 5.7.38:

git checkout c94ce78

I installed Boost and Cmake like what I did in the earlier post. I got the rest of the development tools like gcc, make, bison in place using this dnf command:

dnf groupinstall "Development Tools"

Then I had to get the cmake command to work:

cmake . -DWITH_BOOST=/home/bobby/boost_1_59_0

I had to install several packages before I could get this to run without errors. I had to enable the “CodeReady Builder” repository in the file oracle-linux-ol8.repo:

[root@localhost yum.repos.d]# diff oracle-linux-ol8.repo oracle-linux-ol8.repo.06212023
< enabled=1
> enabled=0

name=Oracle Linux 8 CodeReady Builder ($basearch) - Unsupported

Once that repository was enabled, I was able to add the needed packages.

dnf install ncurses-devel
dnf install libtirpc-devel
dnf install rpcgen

Eventually I realized that after messing with adding the new repository and packages I needed to go back and clean everything up and run cmake again:

make clean
rm CMakeCache.txt
cmake . -DWITH_BOOST=/home/bobby/boost_1_59_0
su -
make install

Other than that, it is just the normal steps to create the database and run it, which I think is documented in my earlier post and in the MySQL docs.

I thought it couldn’t hurt to document the things I had to do if nothing else for myself. I use this blog as a reference for myself, so it is not just something for other people to read. Anyway, I’m glad I could get this down and maybe someone else will benefit.


Posted in Uncategorized | Leave a comment

Full Database Restore and Recovery

Friday, we had to do a full restore and recovery of a production Oracle database. We used a script like this:

rman target / <<EOF

run {
allocate channel 'dev_0' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_1' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_2' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_3' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_4' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_5' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';

restore database;
recover database
 delete archivelog maxsize 100 G;



We ran it like this:

nohup ./ > restoredb.txt &

The only tricky thing was getting the allocate channel commands right to work with our backup system (HP Data Protector). We had old examples of past recoveries, and we looked in Data Protector to see how the backup job was setup and what output the last backup produced.

Some storage work that we were doing accidentally damaged a bunch of data files. Fortunately, our control files and redo logs were intact, so we were able to fully recover and did not lose any data.


Posted in Uncategorized | Leave a comment

LOB Space Scripts

I said in my previous post that I put LOB space scripts in my GitHub repository and I wanted to explain a little more here. I have two databases that were growing rapidly and the top segment in each was a LOB segment. For one database the top LOB was a CLOB and for the other it was a BLOB. In both cases there were many inserts and deletes against the tables with the largest LOB segment. I was trying to find out if space was wasted and unusable in the LOB segments. Best I can tell these applications are reusing space when LOBs are deleted. These were both BasicFiles LOBs in 11.2 Oracle. And for the CLOB the character set was such that each character used one byte. Also, both databases had an 8192-byte block size.

A lot of this work was inspired by this Oracle support document:

LOB space not released after delete (Doc ID 2285007.1)

The title made me think the LOB space was wasted after the deletes. But this article had an interesting disclaimer highlighted in a box:

“Note: be aware that the unused space after deletion can be reused in the
LOB segment for further insert after retention time is passed.”

As far as I can tell, Oracle’s disclaimer is correct, at least in our situation. Unused LOB space left after a deletion is reused.

Here is a description of the LOB scripts in my GitHub repository:

spacetest.sql – This was my first script. I used various versions of this script to test different metrics that I could find about LOBs. This was helpful because I started very simple and tried to understand what each source of information was telling me. I really struggled to understand what AVG_ROW_LEN from USER_TABLES was telling me until I realized that I was hitting this bug or one like it:

Bug 14651892 AVG_ROW_LEN computed incorrectly on LOB when AUTO_SAMPLE_SIZE is used

It also took me a long time to understand why DBMS_SPACE.SPACE_USAGE only reported full or unformatted blocks for LOB segments. LOBs are all or nothing for a block. Each LOB that is not stored inline for a row fully occupies one or more blocks. With all these observations I throw in the caveat that this is the best I can tell based on my observations on my system.

The database with the CLOB uses this character set: WE8MSWIN1252.

This Oracle document says that character sets like this have a single byte per character in a CLOB:

CLOBs and NCLOBs character set storage in Oracle Release 8i, 9i, 10g and higher (Doc ID 257772.1)

“CLOBs when using fixed width character set NLS_CHARACTERSET are stored in the NLS_CHARACTERSET character set on disk. Examples are WE8MSWIN1252 …”

This matters because dbms_lob.getlength returns the size of a CLOB in characters. In some character sets like AL32UTF8 you have to multiply dbms_lob.getlength’s output by 2 to get the bytes for the CLOB.

Given all the information I got from playing with spacetest.sql I created the next two scripts to feed information into the final space script.

blobinlinecutoff.sql – Takes an integer from 1 to 8191 and creates a row in a table with a LOB of that many bytes in size and it outputs whether that LOB is an inline LOB or not. I used this to find the cutoff point between LOBs that fit in a row and ones that did not. I found the cutoff to be 3964 bytes in all the scenarios I tried. I use this cutoff to ignore inline LOBs when counting the total LOB space in the column’s LOB segment.

inoneblock.sql – Takes an integer from 1 to 8191 and creates a row in a table with a LOB of that many bytes in size and it outputs whether that LOB fits fully in one block. Many LOBs are larger than one block which is only 8K for me so I wanted to calculate exactly how many blocks each LOB would take. I got 8132 in my tests which means 60 bytes per block are unusable for LOB storage. So, I would need to divide the LOB size in bytes by 8132 and round the resulting number up to the next integer to get the number of blocks occupied by the LOB.

I plugged the results from blobinlinecutoff.sql and inoneblock.sql into my final script:

lobspace.sql – Outputs information about LOB space used to show how much is allocated but not used by current LOBs. This space could either be available for reuse by future inserts or it could be wasted due to some bug.

Notice that the top of the script has the input from the previous two:

  inlinecutoff number := 3964;
  usableperblock number := 8132;

You also put the table owner, name, and LOB column name at the top as parameters.

The output looks like this on my BLOB column table:

Table owner = MYOWNER
Table name = MYTABLE
LOB column name = MYLOBCOLUMN
Number of rows in table = 380553608
Number of rows with lob in table row = 338491004
Number of rows with lob in lob segment = 42062604
Total lob segment size = 1496535 megabytes
Total size of full lob segment blocks = 1473933.135288238525390625 megabytes
Total lob space used in lob segment = 1462133.555019378662109375 megabytes
Percentage of full blocks used = 99%

PL/SQL procedure successfully completed.

Elapsed: 07:46:50.26

I found that this database is not deleting all the rows that it is supposed to, so it is fully using all the space freed by the deletes and is continually adding space.

But the output from the other database (the one with the CLOB) looked different:

Table owner = MYOWNER2
Table name = MYTABLE2
LOB column name = MYLOBCOLUMN2
Number of rows in table = 66360290
Number of rows with lob in table row = 54200592
Number of rows with lob in lob segment = 12159698
Total lob segment size = 913999 megabytes
Total size of full lob segment blocks = 906413.57425689697265625 megabytes
Total lob space used in lob segment = 373804.97021484375 megabytes
Percentage of full blocks used = 41%

PL/SQL procedure successfully completed.

Elapsed: 00:45:22.57

Notice that the percentage of full blocks used in the first database was 99% but in this one it is 41%. This made me really wonder if deleted space was really being freed for use by inserts. But digging further I found that we only keep 7 days of history on this table and back in November we hit a peak of activity which expanded the LOB segment. Since then, we have stopped adding new space to the LOB segment’s tablespace. It appears that we have plenty of space free for this LOB segment to absorb a new batch of inserts because more than half of the space in the LOB segment is free for reuse.

Here are the settings for the BLOB:

  CHUNK       8192

Here are the settings for the CLOB:

  CHUNK       8192

Some blog posts that I studied for this:

I wanted to put this out there to help others but also myself. I do not want to forget some of the things I learned in the process. Also, if anyone out there has any feedback on this including any errors that I made it would be great to hear it.


Posted in Uncategorized | 2 Comments

LOB Scripts in my GitHub Repository

I do not have time to write a post explaining these now, but I wanted to note that I am posting LOB space related scripts here:


Posted in Uncategorized | Leave a comment

December 26 Plan Change

On December 26th, which is a holiday for my company this year, our team got a page that a plan had changed on an important HP Unix platform Oracle database. The new plan was inefficient but did not cause any harm to the application. I added the SQL_ID value for the query to our plan change monitor exception list so that it would not page us again when it changes to slower plans because the slower plans were not a threat to the system. Normally I would just move on but this time I thought I would dig into why the query changed plan and see if there is anything we could do to prevent similar changes. If the new plan was a problem, I typically would force the faster plan with a SQL Profile. I was going to do that this morning, but the system was busy, and the profile script was hung up for a few minutes on a library cache lock, so I just backed it out. I do not really need to intervene since it is not hurting the system. We have enough CPU capacity to handle the slow plan. But I was still curious if I could find the underlying cause of the plan change and learn any lessons from it. That is what this post is about.

Like many queries that change plan and set off alerts, this problem query uses bind variables. The PL/SQL package that includes the queries passes in values in the where clause instead of using constants. I.e., it looks like this:

    div = p_div
    AND cust_nbr = p_cust_nbr

and not this:

    div = 123
    AND cust_nbr = 456

You can see here where the query changed to the bad plan:

--------------- --------------------- ---------------- ------------------
      504369030 25-DEC-22 04.00.45 AM              838         .478386635
      504369030 25-DEC-22 07.00.40 PM             1599         .669989368
      504369030 25-DEC-22 11.00.15 PM             1044         .595122605
      504369030 26-DEC-22 01.00.16 AM              891         .558159371
      504369030 26-DEC-22 02.00.43 AM              473         .453122622
     2693825138 26-DEC-22 02.00.43 AM              311         3.61250804
     2693825138 26-DEC-22 03.00.07 AM              779         2.91877792
     2693825138 26-DEC-22 04.00.10 AM             1076           7.274671
     2693825138 26-DEC-22 05.00.41 AM             1218         11.1912258

The bad plan is averaging 11 milliseconds instead of less than 1 millisecond so “bad” is relative. It still is not slow enough to affect the users.

I used my plan script to run the problem query with different constants in place of the two variables. I used my optimizer statistics scripts to find what the optimizer thought the range of possible values was for the two columns. Here are the ranges:

----------- ---- --------
DIV         228  5220    
CUST_NBR    1675 74394502

I tried a variety of variable value combinations in and out of this range and most of the values outside these ranges resulted in the bad plan having a lower cost than the good one. It was especially sensitive to the values of DIV. It turns out that there are many possible DIV values outside this range. They are just not on this table. So likely this query is often run with variable values that are out of the range of what the optimizer knows is in the table.

I tried to find a way to make the out-of-range values pick the better plan. I loaded this data on a 19c database but got the same results. I tried different kinds of histograms, including one on (DIV, CUST_NBR), but they didn’t help.

So what?

Maybe when you write a query that takes variables you should try out its performance with atypical values for the variable values. In simple equals conditions like those in my problem query maybe check how the query runs on values that are outside of those found in the table. Since the query’s plan is not fixed you must be prepared for the query to run with all the plans that the data passed into its variables could generate. Also, the database will stick with any of those plans for a long time unless something forces it to be recalculated. So, you would have to test the query with atypical values and once the plan is in memory run all your typical data through that plan to see if it is still fast enough. If the plan from the atypical data is too slow then you would have to change the query or the data to handle those cases or resort to something like a hint to lock in the plan that runs well with the more typical variable values.



I just want to say how hard it is to write a blog post. What a pain. But the interaction with others on the internet is very valuable.

Someone commented that I could look at v$sql_shared_cursor to find why 1anm65yacs6ky changed plan. As expected, it changed plan because statistics were gathered on the table. Often statistics gathering causes plans to be recreated (hard parsing). But it was interesting that the reason from v$sql_shared_cursor was “Rolling Invalidate Window Exceeded”. A quick Google search led me to this article:

It claims that there is a 5 hour window after statistics are gathered for plans to be reparsed. This makes sense because I know statistics are gathered on the table around 22:00 but the plan is flipping between 01:00 and 02:00 the next day. Very cool.

But it does not really explain why the plan changes. It only explains why 1anm65yacs6ky was hard parsed. The plan changed because out of range data was passed into the bind variables during the hard parse and this resulted in a plan that was inefficient for in range data. On December 26th we got unlucky during the hard parse. This morning, January 9th, we got lucky, and we flipped back to the good plan.

What makes blogging about this stuff so hard is that there is so much I want to say about this, but I just don’t know if I have the energy or ability to put it all down. We use bind variables to minimize parsing. With the good plan this query runs in a couple of milliseconds. If it had to be hard parsed every time it runs, 14,000 times per hour, it would be very wasteful and slow. But the cost of using bind variables is that a plan can get locked in that isn’t efficient for many of the values passed into the variables. It is a no-win scenario. The bigger picture is that I think Oracle’s optimizer, and probably any conceivable SQL optimizer, has limits. Somehow all these years Oracle and others have produced SQL RDBMSs that people find useful. But based on my Oracle experience it seems like they are imperfect but useful.

In the case of my December 26th query that changed plan, our business is functioning just fine with the query’s plan flipping back and forth. If I did not have my plan change monitor setup, I would not even know that the plan was changing because it is not causing an issue that is visible to the users. So, it offends my perfectionist tendencies to have a query run an inefficient plan, but from a business perspective the inefficiency and imperfection that is fundamental to the design of Oracle’s optimizer with the tradeoffs of minimizing hard parses with bind variables but suffering with inefficient plans until the next hard parse is acceptable because the business purpose of the application is still being accomplished. Maybe in the bigger picture Oracle’s SQL and SQL in general have been useful to people all these years because they do well enough to meet people’s needs despite their fundamentally imperfect execution of SQL queries.


I found two ways to prevent the SQL statement from changing plans when it is hard parsed with variable values that are out of the range of the column statistics. One is to add a new index, and the other is to add a cardinality hint.

I had no idea if the index would help but I noticed that the query touched the same table 3 times after looking up rows using the only index. I looked at the subselects on the query and picked a new index that would cover all the columns used by the subselects and be efficient. This eliminated all but the one table lookup that we had to have. For some reason this plan is chosen even if the data in the variables is out of range. Maybe the new index made the plan so efficient that out-of-range data would not cause a plan flip.

Just now I got the idea of trying a cardinality hint like this:

/*+ cardinality(PSD 10) */

I put this on every subselect that had the main table. This caused the plan to act as if the table would return 10 rows and it choose the good plan even if the data in the variables was out of range.

These two approaches work but they require some query tuning knowledge when you are rolling out new SQL statements. I think part of the challenge of using SQL databases like Oracle is that you need developers and DBAs who understand the optimizer well enough to understand the challenges of bind variables and plan changes and how to prevent them or fix them.


See this Oracle support document:

Limitations of the Oracle Cost Based Optimizer (Doc ID 212809.1)

There is also this bug which may prevent us from using Adaptive Cursor Sharing on this application which has all its queries wrapped in PL/SQL:


It looks like none of our SQL is using ACS:

  4  count(*) cnt
  5  from
  6  v$sql
  7  group by
 10  order by cnt desc;

I I        CNT
- - ----------
N N      22970

Maybe our application design suppresses the use of ACS which results in more bad plans being locked in on queries with bind variables. If we bypassed PL/SQL and just ran queries against the database from the Java front end, we might use ACS to prevent a bad plan getting locked in based on certain oddball variable values.


Tired of writing PPS and all that. I am probably running out of steam. I cannot find anything that really applies generally to our system based on this example. I found this weird part of the plan when the variable values are out of range:

| Id  | Operation                            | Name            | Rows  |
|   6 |       MERGE JOIN CARTESIAN           |                 |     1 |
|   7 |        TABLE ACCESS BY INDEX ROWID   | PROD_ATLN_DTL   |     1 |
|   8 |         INDEX RANGE SCAN             | PROD_ATLN_DTL_0 |     1 |
|   9 |        BUFFER SORT                   |                 |    36 |
|  10 |         TABLE ACCESS FULL            | XYZ_ENT         |    36 |

It seems like for Id 6 the Rows value should be 36 instead of 1. I think the optimizer treats 0 rows as 1 row when it displays the plan so maybe it is multiplying 0 * 36 and then displaying 1 instead of 0 for id 6. But other parts of the plan treat 0 as if it was 1 and add up the cost of the steps of the plan. Both the good and bad plan have costs that are way off for the out of range variable values so maybe when you get down to this level the errors inherent in the optimizer make plan comparisons meaningless. But then what strategy can you use with queries that use bind variables and that have out-of-range data values passed into them? You can use hints, etc. but what can you do without intervening for every query like this?


Looks like the cartesian joins were caused by the constants being in the query in two places.

In the innermost subselect it was this:

   WHERE     DIV = 111
         AND CUST_NBR = 222

Then in the final where clause:

 WHERE     PCD.DIV = 111
       AND PCD.CUST_NBR = 222

DIV and CUST_NBR where part of the joins all through the query so these constants were only needed once. Ultimately this query was poorly constructed, so the optimizer had trouble locking down a good plan. Maybe the bigger picture is that the better job we do designing our queries, indexes, etc. the less likely plans are to change for the worse.

Posted in Uncategorized | 2 Comments

US Foods IT Hiring

US Foods IT Jobs: url

My company is hiring a number of IT positions with more to come soon.


Posted in Uncategorized | Leave a comment