X86-64 Assembly Book

I have written earlier blog posts about my diversion from studying Oracle to studying computer science. Here are some relevant posts: url1,url2,url3,url4. After finishing the math for computer science online class and book that I was working on I stared working through a book about assembly language programming. I bought the book in a frenzy of enthusiasm about studying computer science for fun. But then I had to decide if I was going to work through the assembly book now and delay moving on to the algorithms class using Python that I had intended to do next. I intended to use the math that I studied to prepare me for the algorithms class. But, assembly is a nice low-level hardware oriented thing to study and a contrast from the math, computer science theory, and higher level Python scripting language. So, I decided to delay my study of algorithms and work on assembly.  I’m working on the last exercise of the 16th chapter out of 19 in the book and thought I would start this blog post to document my experience so that others might benefit from it. The book is Ray Seyfarth’s  “Introduction to 64 Bit Assembly Language Programming for Linux and OS X“. I have saved my work on the exercises on GitHub: repository.

I want to let people know what type of environment and tools that I used so they can compare notes with my experience if they decide to work through the book. I started out on an Oracle Enterprise Linux 7 virtual machine running under VirtualBox on my laptop. Oracle’s Linux is a version of Red Hat Linux. I believe that I had to compile YASM and the author’s ebe tool. It has been a while but I think I had to search around a bit to get the right packages so that they would compile and I had some parts of ebe that never worked correctly. Starting with Chapter 9 Exercise 2 I switched from YASM, the assembler recommended by the author of the book to NASM, a more commonly used assembler. I switched because I hit a bug in YASM. So, chapter 9 exercise 1 and earlier were all YASM. Also, after the early lessons I moved from the GUI ebe debugger to the command line gdb debugger. I wanted to get more familiar with gdb anyway since I use Linux for my job and might need gdb to help resolve problems. After I got a new laptop I switched to using Centos 7 on VirtualBox. I was able to install nasm and gdb using yum from standard repositories and did not have to do any manual compilation of development tools in my new environment. So, if you choose to work through the book you could go the nasm and gdb route that I ended up with if you have challenges installing and using ebe and yasm. There are some minor differences between nasm and yasm but they are pretty easy to figure out using the nasm manual.

There are connections between x86-64 assembly programming on Linux and my job working with Oracle databases. At work, 64 bit Linux virtual machines are our standard Oracle database platform. They are also the building blocks of the cloud. You see a lot of 64 bit Linux on Amazon Web Services, for example. So, I’m really kind of doing assembly language for fun since it is so impractical as a programming language, but at the same time I’m doing it on the platform that I use at work. Maybe when I’m looking at a dump in an Oracle trace file on Linux it will help me to know all the registers. If I’m working with some open source database like MySQL it can’t hurt to know how to debug in gdb and compile with gcc.

The assembly language book also connected with my passion for performance tuning. The author had some interesting things to say about performance. He kind of discouraged people from thinking that they could easily improve upon the performance of the GCC C compiler with all of its optimizations. It was interesting to think about the benefits of SIMD and how you might write programs to work better with pipelining and the CPU’s cache. It was kind of like Oracle performance tuning except you were looking at just CPU and lower level factors. But you still have tests to prove out your assumptions and you have to try to build tests to show that what you think is so will really hold up. Chapter 16 Exercise 1 is a good example of SIMD improving performance. I started with a simple C version that ran in 3.538 seconds. An AVX version of the subroutine did 8 floating point operations at a time and ran the same function in 2.1057 seconds. Here are some of the AVX instructions just for fun:

    vmovups ymm0,[x_buffer]      ; load 8 x[i] values
    vmovups ymm1,[rsi+r10*4]     ; load 8 x[j] values
    vsubps ymm0,ymm1             ; do 8 x[i]-x[j] ops
    vmulps ymm0,ymm0             ; square difference

Generally, x86-64 assembly ended up feeling a lot like C. The book has you use a variety of calls from the C library so in the later chapters the assembly programs had calls to a lot of the functions that you use in C such as printf, scanf, strlen, strcmp, and malloc. Like C it was common to get segmentation faults without a lot to go on about what caused it. Still the back trace (bt) command in gdb leads you right to the instruction that got the error so it some ways it was easier to diagnose segmentation faults in assembly than I remember it being in C. It brought back memories of taking C in college and puzzling over segmentation faults and bad pointers. It also made me think of the time in a previous job when I progressed from C to C++ to Java. I came out of school having done a fair amount of C programming. Then I read up on C++ and object-oriented programming. But C++ still had the segmentation faults. Then I found Java and thought it was great because it gave you more meaningful error messages than segmentation fault. Now, I have embraced Python recently because of the edX classes that I took and because of the ways I have used it at work. Working with assembly has kind of taking me back down the chain of ease of use from Python to C to assembly. I can’t see using C or assembly for every day use but most of the software that we use is written in C so it seems reasonable to have some familiarity with C and the lower level assembly that lies beneath it.

Anyway, I have three more chapters to go but thought I would put out this update now while I am thinking about it. I may tweak this post later or put out a follow-up, but I hope that it is useful to someone. If you feel inclined to study 64 bit x86 assembly on Linux I think that you will find the Ray Seyfarth book a good resource for you. If you want to talk to me about my experience feel free to leave at comment on this post or send an email.


Posted in Uncategorized | Leave a comment

Finished reading the Snowflake database documentation

I just finished reading the Snowflake database documentation and I thought I would blog about my impressions. I have not spent a lot of time using Snowflake so I can not draw from experience with the product. But, I read all the online documentation so I think I can summarize some of the major themes that I noticed. Also, I have read a lot of Oracle database documentation in the past so I can compare Oracle’s documentation to Snowflake’s. Prior to reading the online documentation I also read their journal article which has a lot of great technical details about the internals. So, I intend to include things from both the article and the documentation.

My observations fall into these major categories:

  • Use of Amazon Web Services instead of dedicated hardware for data warehouse
  • Ease of use for people who are not database administrators
  • Use of S3 for storage with partitioning and columnar storage
  • Lots of documentation focused on loading and unloading data
  • Lots of things about storing and using JSON in the database
  • Limited implementation of standard SQL – no CREATE INDEX
  • Computer science terms – from the development team instead of marketing?
  • Role of database administrator – understand architecture and tuning implications
  • Focus on reporting usage of resources for billing
  • JavaScript or SQL stored functions
  • Down side to parallelism – high consumption of resources
  • High end developers but still a new product
  • Maybe specialized purpose – not general purpose database

First let me say that it is very cool how the Snowflake engineers designed a data warehouse database from scratch using Amazon Web Services (AWS). I have worked on a couple of different generations of Exadata as well as HP’s Neoview data warehouse appliance and a large Oracle RAC based data warehouse so I have experience with big, expensive, on site data warehouse hardware. But Snowflake is all in Amazon’s cloud so you don’t have to shell out a lot of money up front to use it. It makes me think that I should try to come up with some clever use of AWS to get rich and famous. All of the hardware is there waiting for me to exploit it and you can start small and then add hardware as needed. So, to me Snowflake is a very neat example of some smart people making use of the cloud. Here is a pretty good page about Snowflake’s architecture and AWS: url

You would not think that I would be happy about a database product that does not need database administrators since I have been an Oracle database administrator for over 20 years. But, it is interesting how Snowflake takes some tasks that DBAs would do working with an on site Oracle database and makes them easy enough for a less technical person to do them.  There is no software to install because Snowflake is web-based. Creating a database is a matter of pointing and clicking in their easy to use web interface. Non-technical users can spin up a group of virtual machines with enormous CPU and memory capacity in minutes. You do not setup backup and recovery. Snowflake comes with a couple of built-in recovery methods that are automatically available. Also, I think that some of the redundancy built into AWS helps with recovery. So, you don’t have Oracle DBA tasks like installing database software, creating databases, choosing hardware, setting up memory settings, doing RMAN and datapump backups. So, my impression is that they did a good job making Snowflake easier to manage. Here is a document about their built-in backup and recovery: url

Now I get to the first negative about Snowflake. It stores the data in AWS’s S3 storage in small partitions and a columnar data format. I first saw this in the journal article and the documentation reinforced the impression: url1,url2. I’ve used S3 just enough to upload a small file to it and load the data into Snowflake. I think that S3 is AWS’s form of shared filesystem. But, I keep thinking that S3 is too slow for database storage. I’m used to solid state disk storage with 1 millisecond reads and 200 microsecond reads across a SAN network from a storage device with a large cache of high-speed memory. Maybe S3 is faster than I think but I would think that locally attached SSD or SSD over a SAN with a big cache would be faster. Snowflake seems to get around this problem by having SSD and memory caches in their compute nodes. They call clusters of compute nodes warehouses, which I think is confusing terminology. But from the limited query testing I did and from the reading I think that Snowflake gets around S3’s slowness with caching. Caching is great for a read only system. But what about a system with a lot of small transactions? I’ve seen Snowflake do very well with some queries against some large data sets. But, I wonder what the down side is to their use of S3. Also, Snowflake stores the data in columnar format which might not work well for lots of small insert, update, and delete transactions.

I thought it was weird that out of the relatively small amount of documentation Snowflake devoted a lot of it to loading and unloading data. I have read a lot of Oracle documentation. I read the 12c concepts manual and several other manuals while studying for my OCP 12c certification. So, I know that compared to Oracle’s documentation Snowflake’s is small. But, I kept seeing one thing after another about how to load data. Here are some pages: url1,url2. I assume that their data load/unload statements are not part of the SQL standard so maybe they de-emphasized documenting normal SQL constructs and focused on their custom syntax. Also, they can’t make any money until their customers get their data loaded so maybe loading data is a business priority for Snowflake. I’ve uploaded a small amount of data so I’m a little familiar with how it works. But, generally, the data movement docs are pretty hard to follow. It is kind of weird. The web interface is so nice and easy to use but the upload and download syntax seems ugly. Maybe that is why they have some much documentation about it?

Snowflake also seems to have a disproportionate amount of documentation about using JSON in the database. Is this a SQL database or not? I’m sure that there are Oracle manuals about using JSON and of course there are other databases that combine SQL and JSON but out of the relatively small Snowflake documentation set there was a fair amount of JSON. At least, that is my impression reading through the docs. Maybe they have customers with a lot of JSON data from various web sources and they want to load it straight into the database instead of extracting information and putting it into normal SQL tables. Here is an example JSON doc page: url

Snowflake seems to have based their product on a SQL standard but they did not seem to fully implement it. For one thing there is no CREATE INDEX statement. Uggh. The lack of indexes reminds me strongly of Exadata. When we first got on Exadata they recommended dropping your indexes and using Smart Scans instead. But, it isn’t hard to build a query on Exadata that runs much faster with a simple index. If you are looking up a single row with a unique key a standard btree index with a sequential, i.e. non-parallel, query is pretty fast. The lack of CREATE INDEX combined with the use of S3 and columnar organization of the data makes me think that Snowflake would not be great for record at a time queries and updates. Of course, an Oracle database excels at record at a time processing so I can’t help thinking that Snowflake won’t replace Oracle with its current architecture. Here is a page listing all the things that you can create, not including index: url

Snowflake sprinkled their article and documentation with some computer science terms. I don’t recall seeing these types of things in Oracle’s documentation. For example, they have a fair amount of documentation about HyperLogLog. What in the world? HyperLogLog is some fancy algorithm for estimating the number of rows in a large table without reading every row. I guess Oracle has various algorithms under the covers to estimate cardinality. But they don’t spell out the computer science term for it. At least that’s my impression. And the point of this blog post is to give my impressions and not to present some rigorous proof through extensive testing. As a reader of Oracle manuals I just got a different feel from Snowflake’s documentation. Maybe a little more technical in its presentation than Oracle’s. It seems that Snowflake has some very high-end software engineers with a lot of specialized computer science knowledge. Maybe some of that leaks out into the documentation. Another example, their random function makes reference to the name of the underlying algorithm: url. Contrast this with Oracle’s doc: url. Oracle just tells you how to use it. Snowflake tells you the algorithm name. Maybe Snowflake wants to impress us with their computer science knowledge?

Reading the Snowflake docs made me think about the role of a database administrator with Snowflake. Is there a role? Of course, since I have been an Oracle DBA for over 20 years I have a vested interest in keeping my job. But, it’s not like Oracle is going away. There are a bazillion Oracle systems out there and if all the new students coming into the work force decide to shy away from Oracle that leaves more for me to support the rest of my career. But, I’m not in love with Oracle or even SQL databases or database technology. Working with Oracle and especially in performance tuning has given me a way to use my computer science background and Oracle has challenged me to learn new things and solve difficult problems. I could move away from Oracle into other areas where I could use computer science and work on interesting and challenging problems. I can see using my computer science, performance tuning, and technical problem solving skills with Snowflake. Companies need people like myself who understand Oracle internals – or at least who are pursuing an understanding of it. Oracle is proprietary and complicated. Someone outside of Oracle probably can not know everything about it. It seems that people who understand Snowflake’s design may have a role to play. I don’t want to get off on a tangent but I think that people tend to overestimate what Oracle can do automatically. With large amounts of data and challenging requirements you need some human intervention by people who really understand what’s going on. I would think that the same would be true with Snowflake. You need people who understand why some queries are slow and how to improve their performance. There are not as many knobs to turn in Snowflake. Hardly any really. But there is clustering: url1,url2,url3. You also get to choose which columns fit into which tables and the order in which you load the data, like you can on any SQL database. Snowflake exposes execution plans and has execution statistics: url1,url2,url3. So, it seems that Snowflake has taken away a lot of the traditional DBA tasks but my impression is that there is still a role for someone who can dig into the internals and figure out how to make things go faster and help resolve problems.

Money is the thing. There are a lot of money related features in the Snowflake documentation. You need to know how much money you are spending and how to control your costs. I guess that it is inevitable with a web-based service that you need to have features related to billing. Couple examples: url1,url2

Snowflake has SQL and JavaScript based user defined functions. These seem more basic than Oracle’s PL/SQL. Here is a link: url

There are some interesting things about limiting the number of parallel queries that can run on a single Snowflake warehouse (compute cluster). I’ve done a fair amount of work on Oracle data warehouses with people running a bunch of parallel queries against large data sets. Parallelism is great because you can speed up a query by breaking its execution into pieces that the database can run at the same time. But, then each user that is running a parallel query can consume more resources than they could running serially. Snowflake has the same issues. They have built-in limits to how many queries can run against a warehouse to keep it from getting overloaded. These remind me of some of the Oracle init parameters related to parallel query execution. Some URLs: url1,url2,url3 In my opinion parallelism is not a silver bullet. It works great in proofs of concepts with a couple of users on your system. But then load up your system with users from all over your company and see how well it runs then. Of course, one nice thing about Snowflake is that you can easily increase your CPU and memory capacity as your needs grow. But it isn’t free. At some point it becomes worth it to make more efficient queries so that you don’t consume so many resources. At least, that’s my opinion based on what I’ve seen on Oracle data warehouses.

I’m not sure if I got this information from the article or the documentation or somewhere else. But I think of Snowflake as new. It seems to have some high-end engineers behind it who have worked for several years putting together a system that makes innovative use of AWS. The limited manual set, the technical terms in the docs, the journal article all make me think of a bunch of high-tech people working at a startup. A recent Twitter post said that Snowflake now has 500 customers. Not a lot in Oracle terms. So, Snowflake is new. Like any new product it has room to grow. My manager asked me to look into technical training for Snowflake. They don’t have any. So, that’s why I read the manuals. Plus, I’m just a manual reader by nature.

My impression from all of this reading is that Snowflake has a niche. Oracle tries to make their product all things to all people. It has every feature but the kitchen sink. They have made it bloated with one expensive add-on option after another. Snowflake is leaner and newer. I have no idea how much Snowflake costs, but assuming that it is reasonable I can see it having value if companies use it where it makes sense. But I think it would be a mistake to blindly start using Snowflake for every database need. You probably don’t want to build a high-end transactional system on top of it. Not without indexes! But it does seem pretty easy to get a data warehouse setup on Snowflake without all the time-consuming setup of an on premise data warehouse appliance like Exadata. I think you just need to prepare yourself for missing features and for some things not to work as well as they do on a more mature database like Oracle. Also, with a cloud model you are at the mercy of the vendor. In my experience employees have more skin in the game than outside vendors. So, you sacrifice some control and some commitment for ease of use. It is a form of outsourcing. But, outsourcing is fine if it meets your business needs. You just need to understand the pros and cons of using it.

To wrap up this very long blog post, I hope that I have been clear that I’m just putting out my impressions without a lot of testing to prove that I’m right. This post is trying to document my own thoughts about Snowflake based on the documentation and my experience with Oracle. There is a sense in which no one can say that I am wrong about the things that I have written as long as I present my honest thoughts. I’m sure that a number of things that I have written are wrong in the sense that testing and experience with the product would show that my first impressions from the manuals were wrong. For example, maybe I could build a transactional system and find that Snowflake works better than I thought. But, for now I’ve put out my feeling that it won’t work well and that’s just what I think. So, the post has a lot of opinion without a ton of proof. The links show things that I have observed so they form a type of evidence. But, with Oracle the documentation and reality don’t always match so it probably is the same with Snowflake. Still, I hope this dump of my brain’s thoughts about the Snowflake docs is helpful to someone. I’m happy to discuss this with others and would love any feedback about what I have written in this post.


Posted in Uncategorized | Leave a comment

Submitted two abstracts for Oracle OpenWorld 2017

I submitted two abstracts for Oracle OpenWorld 2017. I have two talks that I have thought of putting together:

  • Python for the Oracle DBA
  • Toastmasters for the Oracle DBA

I want to do these talks because they describe two things that I have spent time on and that have been valuable to me.

I have given several recent talks about Delphix. Kyle Hailey let me use his slot at Oaktable World in 2015 which was at the same time as Oracle OpenWorld 2015. Right after that I got to speak at Delphix Sync which was a Delphix user event. More recently I did a Delphix user panel webinar.

So, I’ve done a lot of Delphix lately and that is because I have done a lot with Delphix in my work. But, I have also done a lot with Python and Toastmasters so that is why I’m planning to put together presentations about these two topics.

I probably go to one conference every two years so I’m not a frequent speaker, but I have a list of conferences that I am thinking about submitting these two talks to, hoping to speak at one. These conferences are competitive and I’ve seen that better people than me have trouble getting speaking slots at them. But here is my rough idea of what I want to submit the talks to:

I’ve never gone to RMOUG but I think it is in Denver so that is a short flight and I have heard good things.

Also, we have our own local AZORA group in Phoenix. Recently we have had some really good ACE Director/Oak Table type speakers, but I think they might like to have some local speakers as well so we will see if that will work out.

If all else fails I can give the talks at work. I need to start working on the five speeches in my Toastmasters “Technical Presentations” manual which is part of the Advanced Communication Series. I haven’t even cracked the book open, so I don’t know if it applies but it seems likely that I can use these two talks for a couple of the speech projects.

Anyway, I’ve taken the first steps towards giving my Python and Toastmasters speeches. Time will tell when these will actually be presented, but I know the value that I have received from Python and Toastmasters and I’m happy to try to put this information out there for others.


Posted in Uncategorized | 2 Comments

How to find the object that caused ORA-08103 error

A developer told me that two package executions died with ORA-08103 errors and he didn’t know which object caused the errors.

I found two trace files that had the following contents:

*** SESSION ID:(865.1201) 2017-04-17 10:17:09.476
OBJD MISMATCH typ=6, seg.obj=21058339, diskobj=21058934, dsflg=100000, dsobj=21058339, tid=21058339, cls=1

*** SESSION ID:(595.1611) 2017-04-17 10:17:35.395
OBJD MISMATCH typ=6, seg.obj=21058340, diskobj=21058935, dsflg=100000, dsobj=21058340, tid=21058340, cls=1

Bug 13844883 on Oracle’s support site gave me the idea to look up the object id for the diskobj part of the trace as the current object id. So, I needed to look up 21058934 and 21058935. I used this query to find the objects:

select * from dba_objects where DATA_OBJECT_ID in

This pointed to two index partitions that had been rebuilt while the package was running. I’m pretty sure this caused the ORA-08103 error. So, if you get an ORA-08103 error find diskobj in the trace file and look it up as DATA_OBJECT_ID in dba_objects.


Posted in Uncategorized | Leave a comment

Optimizer bug fix makes a query run more than 3 times slower

I’m working on an to upgrade and found a handful of queries that run more than 3 times longer on than The data and optimizer statistics are very similar on the two test databases. I’m pretty sure that an optimizer bug fix caused this difference. So, the irony is that a fix to the optimizer that we get with the upgrade to the very stable release is causing a 3x slowdown in the query.

For my testing I’m using the gather_plan_statistics hint and this query to dump out the plan after executing the query:

select * from table(dbms_xplan.display_cursor(null,null,’ALLSTATS’));

I used an outline hint to force the plan to run under and then I looked at the estimated and actual row counts to find a discrepancy. I found one table with estimated row counts that did not look correct on but made sense on

| Id  | Operation                              | Name                      | Starts | E-Rows | A-Rows |   A-Time   | Buffers |  OMem |  1Mem |  O/1/M   |
|* 29 |           TABLE ACCESS BY INDEX ROWID  | MY_TABLE                  |     17 |      1 |      0 |00:00:07.34 |   96306 |       |       |          |
|* 30 |            INDEX RANGE SCAN            | MY_TABLE_PK               |     17 |     16 |    102 |00:00:01.20 |   96255 |       |       |          |

| Id  | Operation                              | Name                      | Starts | E-Rows | A-Rows |   A-Time   | Buffers |  OMem |  1Mem |  O/1/M   |
|* 29 |           TABLE ACCESS BY INDEX ROWID  | MY_TABLE                  |     17 |      8 |      0 |00:00:07.44 |   96306 |       |       |          |
|* 30 |            INDEX RANGE SCAN            | MY_TABLE_PK               |     17 |     17 |    102 |00:00:01.22 |   96255 |       |       |          |

After the range scan in step 30 in the plan in you have an estimate of 16 rows but the table access in step 29 has an estimate of only 1. In the estimate for step 29 is 8 rows.  Given the optimizer statistics, the optimizer should have estimated 8 rows step 29 in It appears that fixed a bug like this.

Here are the predicates for step 29 in the plan:

29 – filter((“PCP”.”MY_FLAG”=’M’ OR “PCP”.”MY_FLAG”=’Y’))

So the column could have value M or Y. The optimizer statistics have 4 distinct values for the column and no histogram. So the optimizer should assume that 1/4 of the rows meet each criteria. So the optimizer should have estimated 1/4 + 1/4 = 1/2 of the rows from step 30 meet the criteria in step 29. So, 17/2 = 8, rounding down. But in it seems that they multiplied the rows from step 30 by 1/4 two times making it 16*1/4*1/4 = 1. It seems that in the optimizer multiplied by 1/4 twice instead of adding them and then multiplying. There is a known bug related to OR conditions in where clauses:

Bug 10623119 – wrong cardinality with ORs and columns with mostly nulls (Doc ID 10623119.8)

Our database includes this bug fix but I don’t know if this fix caused the difference in behavior that I saw. It seems possible that it did.

The interesting thing is that the real row count for step 29 is 0. So, the pre-bug fix plan in actually estimated the row count more accurately by accident. It estimated 1 and the real count was 0. The correct estimate should have been 8, but that is not as close to 0 as 1 is. I think we just happened to have a few queries where the bug resulted in a more accurate estimate than a properly functioning optimizer. But, I’m only looking at the queries whose performance is worse after the upgrade. There may have been other queries that performed better because of this bug fix.

I ended up passing this and a similar query back to a senior SQL developer and he took one look at the query and described it as “ugly”. He fixed both of them in no time so that both queries now run just as fast or faster on than they did on

So, the original query ran faster when the optimizer was not working properly. A human developer simplified the query and then it ran faster when the optimizer was working properly. Maybe the moral of the story is to build simpler and cleaner SQL queries to begin with and if you find a query whose performance declines with better optimizer information then consider improving the query so that it works well with the better functioning optimizer.


Update: I messed around with cardinality hints some more. The problem with blog posts is that once I’ve written one I start second guessing myself. Did I cover all of my bases? I finally found a cardinality hint on the problem table that forced the plan back to the plan. But the weird thing is that I had to hint that the number of rows on the table was larger than reality. My hint was like this: cardinality(pcp,1000000). I expected that a smaller cardinality would change the plan!

The good thing about this test is that it brought me back to why I focused on this table in the first place. Most of the time in the query execution centered around this one table and its index range scan. Now I know that messing with the cardinality hint on this table changes the plan back I feel good about the idea that this table’s row count has something to do with the plan change in But, I’m not sure how to tie this change back to a specific bug fix.

Yet another update:

I ran the query on with the parameter optimizer_features_enable=’′ and the plan reverted to the plan. At least I know that the slight differences in statistics between the two databases aren’t causing this issue. It is something about I also forced the plan under and its cost was less than the plan that was choosing so that suggests that the optimizer is not even trying the better plan. It would have chosen the lower cost plan. I tried setting optimizer_max_permutations=1000000000000 to get it to try more plans but it still didn’t choose the lower cost plan. So, I guess that my observations about the difference is the row counts on the table do not explain the slower plan. I am not sure how to diagnose why is not trying the lower cost plan. It could be that this is a bug that introduced.

Still I guess all of this research reinforces the main point. I’m tearing my hair out trying to apply all of this arcane Oracle SQL tuning information to understand why the queries ran slower on But, a good SQL developer rewrote the queries in minutes so maybe I do not need to spend more time on it?

I guess if we run into a query that changes to a worse plan in the upgrade we can always try setting optimizer_features_enable=’′ in a hint like this


Posted in Uncategorized | 5 Comments

_small_table_threshold=1000000 results in > 5x query speedup

Today I sped a query up by over 5 times by setting _small_table_threshold=1000000.

Here is the query elapsed time and a piece of the plan showing its behavior before setting the parameter:

Elapsed: 00:28:41.67

| Id  | Operation                                              | Name                   | Starts | E-Rows | A-Rows |   A-Time   | Buffers | Reads  |  OMem |  1Mem |  O/1/M   |
|  69 |              PARTITION RANGE ITERATOR                  |                        |   9125 |      1 |   9122 |00:13:02.42 |    3071K|   3050K|       |       |          |
|* 70 |               TABLE ACCESS FULL                        | SIS_INV_DTL            |   9125 |      1 |   9122 |00:13:02.25 |    3071K|   3050K|       |       |          |

I think that this part of the plan means that the query scanned a range of partitions  9125 times resulting in over three million physical reads. These reads took about 13 minutes. If you do the math it works out to between 200-300 microseconds per read. I have seen similar times from repeated reads from a storage server that has cached the data in memory. I have seen this with a SAN and with Delphix.

Here is my math for fun:

>>> 1000000*((60*13)+2.25)/3050000

About 256 microseconds per read.

I ran this query again and watched the wait events in Toad’s session browser to verify that the query was doing a bunch of direct path reads. Even though the query was doing full scans on the partition range 9000 times the database just kept on doing direct path reads for 13 minutes.

So, I got the idea of trying to increase _small_table_threshold. I was not sure if it would work with parallel queries. By the way, this is on on HP-UX Itanium platform. So, I tried

alter session set "_small_table_threshold"=1000000;

I ran the query again and it ran in under 5 minutes. I had to add a comment to the query to get the plan to come back cleanly. So, then I reran the query again and I guess because of caching it came back in under 2 minutes:

First run:

Elapsed: 00:04:28.83

Second run:

Elapsed: 00:01:39.69

| Id  | Operation                                              | Name                   | Starts | E-Rows | A-Rows |   A-Time   | Buffers | Reads  |  OMem |  1Mem |  O/1/M   |
|  69 |              PARTITION RANGE ITERATOR                  |                        |   9125 |      1 |   9122 |00:00:45.33 |    3103K|      0 |       |       |          |
|* 70 |               TABLE ACCESS FULL                        | SIS_INV_DTL            |   9125 |      1 |   9122 |00:00:45.27 |    3103K|      0 |       |       |          |

The second execution did zero physical reads on these partitions instead of the 3 million that we had without the parameter!

So, it seems that if you have a query that keeps doing full scans on partitions over and over it can run a lot faster if you disable direct path read by upping _small_table_threshold.


Posted in Uncategorized | 2 Comments

Jetpack 4.8 outage

I just got my bobbydurrettdba.com blog back online. It seems that an automatic update of Jetpack to version 4.8 took it down last night. I thought I would post a quick comment on how I resolved it because it is slightly different from what the Jetpack support site says.

When I tried to get into my blog today I just got a white screen saying there was a problem. I could not get into the wp-admin screen to do anything. Then I found out that Jetpack 4.8 had been pushed out and had brought down a lot of WordPress sites. Then I found out that Jetpack 4.8.1 had just come out today to fix it.

Here is the url for the 4.8.1 fix to this issue: fix

The fix points to the manual plugin install url which talks about deleting your /wp-content/plugins/jetpack folder and then manually installing by downloading the 4.8.1 jetpack zip, unzipping it, and ftping it up to /wp-content/plugins.

But all I had to do, after backing up /wp-content/plugins/jetpack was to rename it to /wp-content/plugins/jetpack.old. Then I was able to get into my site and to update the plugin to 4.8.1 through the normal web-based process.

Strangely enough the Jetpack plugin update removed my /wp-content/plugins/jetpack.old directory and replaced it with /wp-content/plugins/jetpack.

Here is what the directory looked like after I ran the update from my blogs admin page:

I never did ftp the 4.8.1 jetpack directory over though I had it unzipped and ready to go on my laptop.

Anyway, I didn’t delete my jetpack directory.  I just renamed it to jetpack.old. Then I ran the normal plugin update process.


p.s. My site stats for today are not looking so good with the blog down all day:

Posted in Uncategorized | 4 Comments

Information Technology Jobs Posted

My company, US Foods, has posted a number of information technology jobs that are in Chicago (Rosemont)  or Phoenix (Tempe). Here is the web site:


Enter Information Technology as the Job Field to see all the posted IT jobs.


Posted in Uncategorized | Leave a comment

Using Delphix to support Oracle upgrade

I’m working on upgrading a very buggy unpatched Oracle database to a fully patched version. I’m using Delphix to support the upgrade and it has been a big help so far. This is on the HP-UX 11.31 Itanium platform.

The great thing about using Delphix to support an upgrade is that my very first pass through the upgrade scripts was with a full-sized clone of production. In the past I probably started with a tiny subset or even an out of the box demo database for my first upgrade pass and even when I got to QA it wasn’t a full test of a production upgrade. This time, my first test was with all the data and that was very cool.

The main example of how this helped is that we had a lot of data in the SYS.WRI$_OPTSTAT_HISTGRM_HISTORY and SYS.WRI$_OPTSTAT_HISTHEAD_HISTORY tables in production and this made the first upgrade of its clone take a long time. After two or three attempts at other ways to speed things up, I ended up applying patch 12683802 on an Oracle home and this allowed me to truncate these two tables.

Delphix helped me here because I had an unused Oracle home on a different host from the one I was doing the upgrade on. I didn’t want to apply a patch on the upgrade host because there were three other databases using the home and I didn’t want to bring them down or patch them. Delphix let me move the VDB that I was upgrading over to the host that had the unused home. Then I applied the patch there and ran the truncate using the dbms_stats.purge_stats(dbms_stats.purge_all) procedure call that the patch enabled.

Then I moved the VDB back to the host where I intended to do the upgrade, which already had the fully patched binaries installed, and ran the upgrade there. Pretty cool. I think I did the most recent upgrade in about 3.5 hours with the empty OPTSTAT tables.

By the way, doing an upgrade within Delphix is easy. You just bring the VDB up on the old oracle home, do the upgrade as you normally would include all the shutdown and startup commands, and then within the Delphix GUI you let Delphix know the VDB is now on a new Oracle home by shutting it down, choosing the new home, and bringing it back up. Piece of cake.


Posted in Uncategorized | 2 Comments

Simple Python for Oracle database tuning example

I ran across a stackoverflow question and it gave me an idea for a simpler use of Python to graph some Oracle database performance information. I looked at my PythonDBAGraphs scripts and I’m not sure that it is worth modifying them to try to simplify those scripts since I like what they do. But they may make people think that Python scripts to graph Oracle performance data are difficult to write.  But, I think if someone just wants to put together some graphs using Python, Matplotlib, and cx_Oracle they could do it more simply than I have in my PythonDBAGraphs scripts and it still could be useful.

Here is an example that looks at db file sequential read waits and graphs the number of waits per interval and the average wait time in microseconds:

import cx_Oracle
import matplotlib.pyplot as plt
import matplotlib.dates as mdates

con = cx_Oracle.connect(connect_string)
cur = con.cursor()

(after.total_waits-before.total_waits) "number of waits",
(after.total_waits-before.total_waits) "ave microseconds"
where before.event_name='db file sequential read' and
after.event_name=before.event_name and
after.snap_id=before.snap_id+1 and
after.instance_number=1 and
before.instance_number=after.instance_number and
after.snap_id=sn.snap_id and
after.instance_number=sn.instance_number and
(after.total_waits-before.total_waits) > 0
order by after.snap_id


datetimes = []
numwaits = []
avgmicros = []
for result in cur:

title="db file sequential read waits"

fig = plt.figure(title)
ax = plt.axes()


# Format X axis dates

ax.fmt_xdata = mdates.DateFormatter('%m/%d/%Y %H:%M')
datetimefmt = mdates.DateFormatter("%m/%d/%Y")

# Title and axes labels

plt.xlabel("Date and time")
plt.ylabel("num waits and average wait time")

# Legend

plt.legend(["Number of waits","Average wait time in microseconds"],
loc='upper left')


The graph it produces is usable without a lot of time spent formatting it in a non-standard way:

It is a short 68 line script and you just need matplotlib and cx_Oracle to run it. I’ve tested this with Python 2.


Posted in Uncategorized | 8 Comments