TL;DR: Some Very Simple tests of Assertions in Orale 26ai. I want to know if this feature has any impact and I want to look a little bit closer or even "under the hood".
This SQL-Assertion feature is still quite fresh, and I feel a bit like that boy looking at an item in development...
<smallprint
Warning: Do Not Trust This Information Yet, and Do Not Rely on it.
My tests here are very simple, blunt. And I would not claim any "solid finding" here yet. I also suspect several thing will happen that make the blog below go "out of date". Please note the caveats...
1- Oracle will further develop the feature and change some of the items I describe here below. For example, I expect new wait-events to appear that can pinpoint SQL-Assertions.
2- The real geeks (ChristofL, KamilS, StefanK) will soon start using oracle-tracefiles on this feature, and learn more about the "inside" (the famous 10046 and 10053 traces, google them if you are interested). I will sit back and wait for others to do the Real Work, while I just stick to SQL.
Also Please note that this sort of "digging inside" is generally only useful if you need to go troubleshooting, or if you are just plain-Curious (like me).
Hence, what I found below is just "scratching the surface" and is mainly written for my own curiousity. And Maybe for use when I need to know if an assertion is causing problems in a real system. But that day is still somewhere in the future...
Just SQL...
As I am a fairly Simple DBA, I try to do all my work with just SQL (SQL*Plus, SQLcl, SQLDeveloper, T.o.a.d., etc...) and I have tried, for all my years as a DBA, to refrain (mostly) from diving into trace files or stack-traces or linux-level debugging. This blog will do "SQL only", which might keep it more accessible for the normal ppl out there...
/smallprint>
Background: SQL-Assertions, what can I see... ?
The previous blog described my first-use of SQL-Assertions (link). After testing a very simple use-case for the new feature (link), and sort-of demonstrating that it works as advertised, I tried to take a closer look.
Firstly, I wanted to see if+how the Assertions created "extra effort", extra load on my system, and possibly How can I determine that any extra load coming from "SQL assertions". And I think I found a way to show that extra effort.
Secondly, I found some statement on the "new objects" in my schema. It is probably all related, so I tried examine some of it. Here is what I found so far.
(and as my notes.txt grows, and my experimental scripts increase, there may be material for more than 1 blog, but not sure how useful this digging into a fresh feature really is...)
My Research Question: Is there extra Load?
First thing I noticed was there were some tables in my schema that I had not created myself, and they seemed related to the SQL-Assertions. Check for yourself with USER_TABLES... But since I could not read or manipulate the data in those tables, I left them alone at first. I suspect that, if they remain visible in future versions, there will be blogs + traces about those tables in the future.
Because Assertions will have to check things conditions "on the fly" during DML, I expect there to be some extra effort, extra load on my database. Hence I set out to measure the "work done by the database".
For my "research", I the started with two things:
1. Session-statistics and the "time model" to have my session report on "effort done"
2. Checking the SQL done by the session from V$SQL, and see if I can spot a difference with and without Assersions.
In the course of looking for those SQL stmnts, I also found some of the "under the covers" SQL caused by SQL-Assertioons.
Experimental setup: Just the Asserion, then run with and without.
From the original test-script in previous blog (see link above), I created a script that would insert significant data to trigger the assertion.
I saved the data inserted from the table A_FND, originally 19 records, in a helper-table, generating 608 records. This data I could use to "fire the assertion" many times.
I than proceeded to insert data into A_FND, and tried to measure the effort. To keep the action repeatable, I used "rollback" to make sure the table didnt grow larger during the experiments.
Then it was just a matter of running the insert several times, with and without the Assertion created. I tried to run that on a database that was as "clean as possible", notably by a flush of the shared_pool. And each run used a "fresh connection" so ethe v$stats of the session would relect only the inserts I did.
Remember that on tests like this you have to take into account the "startup-effect" as there is the overhead of parsing a stmnt for the first time, hence every run was repeated 3 times and every stmnt was done 5x inside every run (scripts in links below)
Run it.. (get the scripts, and try for them yourself!)
After several attempts to put a good, scripted, test together, I ended up with the scripts linked blow. You should be able to try these for yourself. I am curious to see if the outcome differs much from what I found.
The scripts should be run in the order presented. And you can repeat the last two scripts, tst_as3b.sql and tst_as3c.sql, as often as you like.
When running the, you will see that the "session statistics" (using mystat.sql) from insert-statments done "With" the Assertion switched on, are measurably higher.
Note Again, there was a startup-effect: the very first time running the whole set of scripts on a fresh database, the numbers tended to be higher, but after running tst_as3b for several times, they are repeatable, and I would consider them reliable.
I also tested against both Oracle an oracle provided container-image of v23.26.1 and against the image of gvenzl/orale-free:slim, also v23.26.1. Both yield similar resulting number that differ only in single-percentages.
There is Extra effort going on "With" the SQL-Assertion, and it is Detectable, Measurable. I consider that a good sign, because it means we found a way to "attribute" effort to the assertion.
The outcome that I found most interesting is Here:
Notice the difference in effort between "with" and "without" the SQL-Assertion in place. The numbers are "per execution" (e.g. ela_p_x = elapsed time per execution).
When The Assertion is defined (With..), the insert of 608 rows takes about 3x longer in CPU-tine and in ELApsed-time in microseconds. And "with" uses .. 2606 buffer-gets instead of the 142. There is a clear indication of extra effort, and it is visible "on the statement" in V$SQL, not hidden in some background-process.
In total time+effort: both statements are still well inside a single Second, and inside what I would call "acceptably efficient" for processing 608 records. Hence on a normal system, a user would probably not notice any difference. But on a busy system, the overhead for SQL-Assertions might be noticeable in AWR or in db-time consumed, e.g. capacity used.
I want to put at least two comments on this:
- The extra effort would likely be A Lot More if we implemented this "constraint" in PL/SQL or in Triggers instead of in an SQL-Assertion.
- This feature is brand-new: Expect changes and improvements in the future.
Diving Deeper ? more SQL-stmtns.
I could not resist examining the "new tables" that I got from creating the assertion. And for quick search on their names in V$SQL I used the script tst_as3c.sql.
It so happens that, in my testcase, there are some 23-26 additional SQL statements generated when we do inserts on tables involved in an Assertion. Some consume very little, some Check for yourself.
Here is a sample:
Interesting to see those stmnts...
In future tests, I would be insteresting to see if those stmnts "cover" the extra effort found when using the Assertion. And it will be interesting to see how those may one day show up in an AWR or AHS.
For now I am just content to "see them". In future I will probably be able to more or less guesstimate the impact of SQL-Assertions on systems, and to put together unit-tests to spot possible problems.
And I hope to spot them "in the wild" if they ever cause Real Trouble.
Summary: Found it. And Happy so far.
I found the Extra Effort caused by SQL-Assertions.
With these tests, I was able to spot the extra effort generated by SQL-Asssertions on a simple test-system. And if these items (temp-tables, SQL-stmnts) behave similarly in future implementations, I will be able to spot them, and to "test them". And I will probably be able to help fix any problems they cause.
I should also add that I think these SQL-Assertions are a Real Cool Feature, which can potentially replace many triggers and "after-the-fact" check programs.
This can make The Database even Smarter.
#SmartDB.
-- -- -- -- -- End of this blog, for Now -- -- -- -- --
There are of course some appendices, notes, remarks. So here come the Appendices....
-- -- -- -- -- Appendix 1 : Links to Scripts -- -- -- -- --
This is what I used for this blog (and for the previous one). If you run them in this order, you should get comparable results:
demobld.sql : Build the demo tables EMP and DEPT.
tst_ass.sql : Script from previous blog, use this to prime the demo-test.
mystat.sql : Report statistics from v$mystat. Called repeatedly.
connscott.sql : Freshly connect scott/tiger. Called repeatedly.
tst_as3b.sql : The test-script for this blog: do tests, show results.
tst_as3c.sql : Lis the other SQL...
Note: if you run on Windows, the host-read command to pause the script will not work. Easiest is to just remove it or out-comment it and examine the spool-file, tst_as3b.lst, afterwards.
The containers/images I used:
With an Honorable Metion to Gerald Venzl for providing the super-dev-Friendly images via Docker (link).
(notice that those images, both from Oracle and from GVenzl, get updated over time, generally every few months - YVMBN: your version may be newer...)
-- -- -- -- -- Appendix 2: more future items -- -- -- -- --
There is always more to investigate...
- Split up the assertion in two separate ones to avoid the OR-conditions and see if that is more efficient.. (Interesting one!!)
- What is the impact on Delete- or Update stmnts ? (on Delete, for this particular assertion, there may not be an impact at all.. )
- Examine the additional SQL in detail, see if the effort for those (background-) stmnts adds up to the difference in time+effort of the original inserts.
- Run larger tests, and examine the effects in AWR reports: Is this the good way to spot "expensive assertions", possibly find problematic ones ?
- We could examine the ORA$SA$ tables in more details. But Is that Relevant ??
- .. You can probably think of more..
And remember, we are only geeking out.. and this is a very new Feature.. Expect more development some time soon.
Maybe Later !
-- -- -- -- -- End of this Blog, for Real -- -- -- -- --
