Bobby Durrett's DBA Blog

I am not a very good blogger. Only four posts in 2024. I will need to pick it up because I just increased my spend on AWS for this blog. This site kept going down and I finally spent a few minutes looking at it and found that it was running out of memory. I was running on a minimal t2.micro EC2 which has 1 virtual CPU and 1 gigabyte of memory with no swap. So, rather than just add swap I bumped it up to a t3.small and paid for a 3-year reserved instance. It was about $220, nothing outrageous. Worth it to me. I still have two years left on the t2.micro reserved instance that I was using for the blog so I need to find a use for it, but I don’t regret upgrading. I spent a little money to make the site more capable so now I must write more posts!

I have three Machine Learning books to work through. After finishing my edX ML class with Python I picked up a book that covered the same topics and used the same Python libraries and I have been steadily working through it: Machine Learning with PyTorch and Scikit-Learn. I am on chapter 9 and want to get up to chapter 16 which covers Transformers. My edX class got up to the material in chapter 15 so the book would add to what was taught in the class. Earlier chapters also expand on what was taught in the class. The cool thing about the book is the included code. Lots of nice code examples to refer to later. Even the Matplotlib code for the graphs could be very helpful to me. I’ve gotten away from using Matplotlib after using it in my PythonDBAGraphs scripts.

My birthday is the day after Christmas, so I get all my presents for the year at the end of December. I got two ML books for Christmas/Birthday. Probably the first I will dive into after I finish the PyTorch/Scikit-Learn book is Natural Language Processing with Python. This is available for free on the NLTK web site. I have the original printed book. I think once I get through the Transformers chapter of the PyTorch book it makes sense to look at natural language processing since that is what ChatGPT and such is all about. I have played with some of this already, but I like the idea of working through these books.

The second book that I got in December as a present seems more technical and math related, although the author claims to have kept the math to a minimum. It is Pattern Recognition and Machine-Learning. This might be a slower read. It seems to be available for free as a PDF. ChatGPT recommended this and the NLTK book when I was chatting with it about my desire to learn more about Artificial Intelligence and Machine Learning.

I have all these conversations with ChatGPT about whether it makes sense for me as an Oracle database specialist to learn about machine learning. It assures me that I should, but it might be biased in favor of ML. Oracle 23ai does have AI in its name and does have some machine learning features. But it could just be a fad that blows over after the AI bubble bursts as many have before it. I can’t predict that. My current job title is “Technical Architect”. I work on a DBA team and I’m in the on-call rotation like everyone else, but my role includes learning about new technology. Plus, I think that I personally add value because of some of the computer science background I had in school and that I have been refreshing in recent years. Plus, I need to get some level of understanding of machine learning for my own understanding regardless of how much we do or don’t use it for my job. Just being a citizen of the world with a computer science orientation I feel is enough motivation to get up to speed on recent AI advancements. Am I misguided to think I should study machine learning?

Despite all this talk about AI, I am still interested in databases. I have this folder on my laptop called “Limits of SQL Optimization”. I had these grandiose ideas about writing interesting blog posts about what SQL could or couldn’t do without human intervention. Maybe SQL is a little like AI because the optimizer does what a human programmer would have to do without it. I’m interested in it all really. I like learning about how computer things work. I’ve spent my career so far working with SQL statements and trying to understand how the database system processes them. I thought about playing with MySQL’s source code. I downloaded it and compiled it but that’s about it. Plus, some day we will have Oracle 23ai and I’ll have to figure out how to support it in our environment, even if we do not use its AI features. Anyway, I’m sure I will have some non-AI database things to post about here.

To wrap up I think I may have some things to post on this blog in 2025, so it’s worth the extra expense to keep it running. Could be some more machine learning/artificial intelligence coming as I work through my books. I still have databases on the brain. Wish you all a great new year.

Bobby

In my previous post I mentioned that I took a machine learning class based on Python and a library called PyTorch. Since the class ended, I have been working on a useful application of the PyTorch library and machine learning ideas to my work with Oracle databases. I do not have a fully baked script to share today but I wanted to show some things I am exploring about the relationship between the current date and time and database performance metrics such as host CPU utilization percentage. I have an example that you can download here: datetimeml2.zip

There is a relationship between the day of the week and the hour of the day and database performance metrics on many Oracle database systems. This is a graph from my hostcpu.py script that shows CPU utilization on a production Oracle database by date and hour of the day:

During the weekdays the CPU peaks at a certain hour and on the weekends, there are some valleys. So, I thought I would use PyTorch to model this relationship. Here is what the graph looks like of actual host CPU used versus predicted by PyTorch:

It’s the same source database but an earlier date. The prediction is close. I guess the big question after I got to this point was, so what? I’m not sure exactly what to do with it now that I have a model of the database CPU. I guess if the CPU is at 100% for an entire day instead of going up and down, I should throw an alert? What if CPU stays near 0% for a day during the week? It must be helpful to have a prediction of the host CPU but exactly how to alert on deviations from the predicted value is still a work in progress.

I thought it would be helpful to talk about the date and time inputs to this model. If you look at datetimeoscpu.sql in the zip it has this SQL for getting the date and time values:

select 
to_char(sn.END_INTERVAL_TIME,'D') day_of_week,
to_char(sn.END_INTERVAL_TIME,'DD') day_of_month,
to_char(sn.END_INTERVAL_TIME,'DDD') day_of_year,
to_char(sn.END_INTERVAL_TIME,'HH24') hour_of_day,
to_char(sn.END_INTERVAL_TIME,'MM') month,
to_char(sn.END_INTERVAL_TIME,'Q') quarter,
to_char(sn.END_INTERVAL_TIME,'WW') week_of_year,
to_char(sn.END_INTERVAL_TIME,'W') week_of_month,
to_char(sn.END_INTERVAL_TIME,'YYYY') year,
...

I ended up ignoring year because it is not a cyclical value. The rest have a range like 1 to 7 for day of week or 1 to 31 for day of month. Having all eight of these is probably overkill. I could just focus on day of week and hour of day and forget the other six. We have 6 weeks of AWR history so I’m not sure why I care about things like month, quarter, day of year because I don’t have multiple years of history to find a pattern.

My input data looks like this:

4 17 199 03 07 3 29 3 2024            15.4210261
4 17 199 04 07 3 29 3 2024            15.5799532
4 17 199 05 07 3 29 3 2024            26.2080227
4 17 199 06 07 3 29 3 2024            24.1532019
4 17 199 07 07 3 29 3 2024            23.2947767
4 17 199 08 07 3 29 3 2024            25.5198717
4 17 199 09 07 3 29 3 2024             26.029638
4 17 199 10 07 3 29 3 2024            27.1160204
4 17 199 11 07 3 29 3 2024            27.7076496

Each line represents an AWR snapshot. The first 8 are the cyclical date and time input values or X. The last value is the host CPU utilization percentage or Y. The point of the program is to create a model based on this data that will take in the 8 values and put out a predicted CPU percentage. This code was used to make the predictions for the graph at the end of the train.py script:

predictions = model(new_X)

When I first started working on this script it was not working well at all. I talked with ChatGPT about it and discovered that cyclical values like hour of day would work better with a PyTorch model if they were run through sine and cosine to transform them into the range -1 to 1. Otherwise PyTorch thinks that an hour like 23 is far apart from the hour 0 when really, they are adjacent. Evidently if you have both cosine and sine their different phases help the model use the cyclical date and time values. So, here is the code of the function which does sine and cosine:

def sinecosineone(dttmval,period):
    """
    Use both sine and cosine for each of the periodic
    date and time values like hour of day or
    day of month
    
    """
    # Convert dttmval to radians
    radians = (2 * np.pi * dttmval) / period
    # Apply sine and cosine transformations
    sin_dttmval = np.sin(radians)
    cos_dttmval = np.cos(radians)
    return sin_dttmval, cos_dttmval

def sinecosineall(X):
    """
    Column         Number
    
    day_of_week      7
    day_of_month     31
    day_of_year      366
    hour_of_day      24
    month            12
    quarter          4
    week_of_year     52
    week_of_month    5
    
    """
...

The period is how many are in the range – 24 for hour of the day. Here is the hour 23 and hour 0 example:

>>> sinecosineone(23.0,24.0)
(-0.25881904510252157, 0.9659258262890681)
>>> sinecosineone(0.0,24.0)
(0.0, 1.0)

Notice how the two values are close together for the two hours that are close in time. After running all my input data through these sine and cosine procedures they are all in the range -1.0 to 1.0 and these were fed into the model during training. Once I switched to this method of transforming the data the model suddenly became much better at predicting the CPU.

You can play with the scripts in the zip. I really hacked the Python script together based on code from ChatGPT, and I think one function from my class. It isn’t pretty. But you are welcome to it.

I’m starting to get excited about PyTorch. There are a bunch of things you must learn initially but the fundamental point seems simple. You start with a bunch of existing data and train a model that can be used as a function that maps the inputs to the outputs just as they were in the original data. This post was about an example of inputting cyclical date and time values like time of day or day of week and outputting performance metrics like CPU used percentage. I don’t have a perfect script that I have used to great results in production, but I am impressed by PyTorch’s ability to predict database server host CPU percent based on date and time values. More posts to come, but I thought I would get this out there even though it is imperfect. I hope that it is helpful.

Bobby

Purpose – Document AI Resources

In this post, I will describe the AI training resources that I am using in my quest to catch up with the current state of the art. My goal is to document the resources that I used for my own reference and to benefit others while explaining my reasoning along the way.

After studying AI back in the 1980s, I have not kept my AI skills up to date until recently when I started learning about newer things like ChatGPT. AI is in the news and every technical person should learn something about it for their own understanding and to help their career. Because I had a good overall understanding of the state of AI years ago, I think of it as “catching up” with the present-day standard. So, my approach makes sense for me given my history. But even for people who are not as old as I am or who don’t have the AI experience from their past the resources below could still be useful because of their overall quality.

Revisit AI Overview – 6.034

The first thing I did was watch the lecture videos for Patrick Winston‘s 2010 MIT 6.034 class. I had used Winston’s textbook in my college AI class in the 1980s. So, I thought that a class taught my Winston might follow a similar outline to what I once knew about AI and help jog my memory and catch me up with what has changed in the last 30+ years. Just watching the lecture videos and not doing the homework and reading limited how much I got from the class, but it was a great high-level overview of the different areas of AI. Also, since it was from 2010 it was good for me because it was closer to present day than my 1980s education, but 14 years away from the current ChatGPT hoopla. One very insightful part of the lecture videos is lectures 12A and 12B which are about neural nets and deep neural nets. Evidently in 2010 Professor Winston said that neural nets were not that promising or could not do that much. Fast forward 5 years and he revised that lecture with two more modern views on neural nets. Of course, ChatGPT is based on neural nets as are many other useful AI things today. After reviewing the 6.034 lecture videos I was on the lookout for a more in depth, hands on, class as the next step in my journey.

edX Class With Python – 6.86x

I had talked with ChatGPT about where to go next in my AI journey and it made various suggestions including books and web sites. I also searched around using Google. Then I noticed an edX class about AI called “Machine Learning with Python-From Linear Models to Deep Learning,” or “6.86x”. I was excited when I saw that edX had a useful Python-based AI class. I had a great time with the two earlier Python edX classes which I took in 2015. They were:

• 6.00.1x – Introduction to Computer Science and Programming Using Python
• 6.00.2x – Introduction to Computational Thinking and Data Science

This blog has several posts about how I used the material in those classes for my work. I have written many Python scripts to support my Oracle database work including my PythonDBAGraphs scripts for graphing Oracle performance metrics. I have gotten a lot of value from learning Python and libraries like Matplotlib in those free edX classes. So, when I saw this machine learning class with Python, I jumped at the chance to join it with the hope that it would have hands-on Python programming with libraries that I could use in my regular work just as the 2015 classes had.

Math Prerequisites

I almost didn’t take 6.86x because the prerequisites included vector calculus and linear algebra which are advanced areas of math. I may have taken these classes decades ago, but I haven’t used them since. I was afraid that I might be overwhelmed and not able to follow the class. But since I was auditing the class for free, I felt like I could “cheat” as much or as little as I wanted because the score in the class doesn’t count for anything. I used my favorite algebra system, Maxima, when it was too hard to do the math by hand. I had some very interesting conversations with ChatGPT about things I didn’t understand. My initial fears about the math were unfounded. The class helped you with the math as much as possible to make it easier to follow. And, at the end of the day since I wasn’t taking this class for any kind of credit, the points don’t matter. (like the Drew Carey improv show). What matters was that I got something out of it. I finished the class this weekend and although the math was hard at times I was able to get through it.

How to Finish Catching Up to 2024

It looks like 6.86x is 4-5 years old so it caught me up to 2019. I have been thinking about where to go from here to get all the way to present day. Working on this class I found two useful Python resources that I want to pursue more:

• PyTorch – Neural Nets
• Hugging Face – Large Language Models

Neural Nets with PyTorch

I walked away from both 6.034 and 6.86x realizing how important neural nets are to AI today and so I want to dig deeper into PyTorch which is a top Python library for neural nets originally developed by Facebook. With the importance of neural nets and the experience I got with PyTorch from the class it only makes sense to dig deeper into PyTorch as a tool that I can use in my database work. As I understand it Large Language Models such as those behind ChatGPT are built using tools like PyTorch. Plus, neural nets have many other applications outside of LLMs. I hope to find uses for PyTorch in my database work and post about them here.

Large Language Models with Hugging Face

Everyone is talking about ChatGPT and LLMs today. I ran across the Hugging Face site during my class. I can’t recall if the class used any of the models there or if I just ran across them as I was researching things. I would really like to download some of the models there and play with updating them and using them. I have played with LLM’s before but have not gotten very far. I tried out OpenAI’s API programming doing completions. I played with storing vectors in MongoDB. But I didn’t have the background that I have now from 6.86x to understand what I was working with. If I can find the time, I would like to both play with the models from Hugging Face and revisit my earlier experiments with OpenAI and MongoDB. Also, I think Oracle 23ai has vectors so I might try it out. But one thing at a time! First, I really want to dig into PyTorch and then mess with Hugging Face’s downloadable models.

Recap

I introduced several resources for learning about artificial intelligence. MIT’s OCW class from 2010, 6.034, has lecture videos from a well-known AI pioneer. 6.86x is a full-blown online machine learning class with Python that you can audit for free or take for credit. Python library PyTorch provides cutting edge neural net functionality. Hugging Face lets you download a wide variety of large language models. These are some great resources to propel me on my way to catching up with the present-day state of the art in AI, and I hope that they will help others in their own pursuits of AI understanding.

Bobby

P.S. Here are a couple of fun screenshots from the last project in my 6.86x class. Hopefully it won’t give too much away to future students.

This is a typical run using PyTorch to train a model for the last assignment.

This is the graphical output showing when the training converged on the desired output.

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:

https://www.bobbydurrettdba.com/uploads/mysqlindexuserangequeries.zip

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 
(a integer NOT NULL AUTO_INCREMENT,
b integer,
PRIMARY KEY (a));

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

This query uses a range query using the index:

select
sum(b)
from
test
where
a > -2147483648;

This query uses a full table scan:

select
sum(b)
from
test
where
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.

Bobby

P.S. I did find this document:

https://dev.mysql.com/doc/refman/8.0/en/integer-types.html

-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:

https://dev.mysql.com/doc/refman/8.0/en/where-optimization.html

“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.

I wrote a script called archivelogspace.py 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 archivelogspace.py 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:

8795958804480
6
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.

Bobby

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

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:

userprivs.sql

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

GRANTED_ROLE
--------------------
DBA
RESOURCE
	
System privileges for user MYUSER

PRIVILEGE
----------------------------------------
SELECT ANY TABLE
UNLIMITED TABLESPACE
	
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

PRIVILEGE
----------------------------------------
ADMINISTER ANY SQL TUNING SET
ADMINISTER DATABASE TRIGGER
ADMINISTER RESOURCE MANAGER
...
	
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

ACCOUNT_STATUS                   LAST_PASSWORD_CHNG
-------------------------------- -------------------
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.

Bobby

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:

https://python.langchain.com/docs/modules/data_connection/vectorstores/

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:

https://python.langchain.com/docs/integrations/vectorstores/mongodb_atlas

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:

https://platform.openai.com/docs/quickstart/build-your-application

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.

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

https://www.bobbydurrettdba.com/2019/05/22/mysql-source-installs-for-each-rds-version/

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:

https://dev.mysql.com/doc/refman/5.7/en/installing-development-tree.html

I ran these git commands:

git clone https://github.com/mysql/mysql-server.git

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
20c20
< enabled=1
---
> enabled=0

[ol8_codeready_builder]
name=Oracle Linux 8 CodeReady Builder ($basearch) - Unsupported
baseurl=https://yum$ociregion.$ocidomain/repo/OracleLinux/OL8/codeready/builder/$basearch/
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-oracle
gpgcheck=1
enabled=1

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
make
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.

Bobby

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/libob2oracle8_64bit.so,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_1' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/libob2oracle8_64bit.so,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_2' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/libob2oracle8_64bit.so,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_3' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/libob2oracle8_64bit.so,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_4' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/libob2oracle8_64bit.so,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';
allocate channel 'dev_5' type 'sbt_tape'
 parms 'SBT_LIBRARY=/opt/omni/lib/libob2oracle8_64bit.so,ENV=(OB2BARTYPE=Oracle8,OB2APPNAME=MYDB,OB2BARLIST=VTL_myhost_MYDB_oral1)';

restore database;
recover database
 delete archivelog maxsize 100 G;

}

exit
EOF

We ran it like this:

nohup ./restoredb.sh > 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.

Bobby

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:

LOB (MYLOBCOLUMN) STORE AS BASICFILE (
  TABLESPACE  MYTABLESPACE
  ENABLE      STORAGE IN ROW
  CHUNK       8192
  RETENTION
  CACHE
  LOGGING

Here are the settings for the CLOB:

LOB (MYLOBCOLUMN2) STORE AS BASICFILE (
  TABLESPACE  MYTABLESPACE2
  ENABLE      STORAGE IN ROW
  CHUNK       8192
  RETENTION
  NOCACHE
  LOGGING

Some blog posts that I studied for this:

https://asktom.oracle.com/pls/apex/asktom.search?tag=reclaimreuse-lob-space

https://jonathanlewis.wordpress.com/2017/01/26/basicfile-lobs/

https://jonathanlewis.wordpress.com/2022/07/20/lob-space-2/

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.

Bobby