As I learn more about the CBO and Banner, I start to wonder about "system tuning". I like Cary Millsap's "Method R" concept; I was thinking about how to adapt that in the narrow field of sql statements. If, for example, I turn set statistics_level = ALL, I know I can generate rowsource statistics at the cost of perhaps 5% overhead to my system. Doing this in a development environment, I could utilize Wolfgang Breitling's "Cardinality Feedback" to find sql statements that are not performing optimally, and maybe even rank them by degradation. Maybe.
Why would I want to do that? Good question. In and of itself, it is a hard sell if my only purpose is to go on a witch hunt and try to tune the system purely from the sql. However, I think the bigger advantage is to use the various sql statements as teachable moments, a way to develope real-life examples for our DBAs that demonstrate how our application(s) behave, and how to attack "slow" statements. The fortunate side-affect is that we end up making some things faster.
Just an idea. I have not been able to drill down into 11g's automatic tuner yet, but I have to believe it is using something similar to cardinality feedback. If memory serves, I believe it does spend some extra overhead reconsidering previous sql plans in an attempt to find those that are not optimal (or that crazy phrase "suboptimal"). Now I have to go back and read up on that again just to refresh myself....
Thursday, September 18, 2008
Wednesday, August 27, 2008
Character Sets
I have been tasked with presenting on the topic of converting the character set for our Banner system databases. We have been through the exercise several times, which has been quite an eye-opening experience in itself (The DataPump Index work-around, dbms_pclxutil.build_part_index, creating tablespaces). Essentially, the vendor has decided in all its god-like wisdom to force all customers to start using a multibyte Unicode character set that will support international characters (AL32UTF8) when they upgrade to the latest version 8. Yea, not very practical for us at a Big Ten school.
Anyway, I have taken this thing to heart and have been learning a lot about codepoints and various facts about character sets. We will be converting from WE8ISO8859P1 to AL32UTF8, a single-byte ISO 8859-1 West European somewhat basic (eg, simple) codepoint to the latest'n'greatest multibyte grandaddy of Universal character sets. The big issue is making sure that all the data stays the same, even if physically it is stored a little different. Along the way, I have had to come to terms with phrases such as "Lossy Data" and "Truncation Data". April Sims has been a great help in answering some questions and getting me to focus my searches. In the end, Metalink note 252352.1 laid out an example that finally opened my eyes for Lossy Data.
I was having a hard time understanding why one must go through WE8MSWIN1252 first to resolve lossy data. And what was meant by the notation in Metalink Note 260192.1:
Why? Well, the Euro example in note 252352.1 is great. To paraphrase, the Euro is stored as codepoint 128 in WE8ISO8859P1. In this case, that character set essentially has an empty hole there, no previous definitions for that codepoint. So if one were to attempt to select it in some kind of "native" WE8ISO8859P1 environment, there would be nothing. So how did it get there? It is passed in from a different character set. And if using that same character set on the client side, you can easily retrieve the correct symbol.
Since WE8WSWIN1252 is a superset of WE8ISO8859P1 (meaning it has all the same codepoints for existing symbols, and then some), and it is most likely that a client in the US inserted data with a US Microsoft character set, the codepoint will be mapped to a valid symbol. Since that codepoint is "valid" in WE8MSWIN1252, it can then be converted to any other character set that also supports that symbol.
Where this falls apart (the "likly" part) is if the client using a "foreign" character set inserts a symbol that is mapped to a completely different symbol (via the codepoint) in our target client character set. In that case, you are screwed. =) Unless you would want to convert that data manually for each case of incompatible character sets. Not I.
Truncation data is a little easier to grasp. Metalink note 119119.1 mentions the English pound sign (£):
In our case, the fix is handled by setting nls_length_semantics=CHAR (opposed to the default BYTE). We also found out through experimentation that nls_length_semantics is completely ignored for SYS and SYSTEM objects; in fact, the data dictionary is always with BYTE semantics. Interesting, eh?
Metalink Notes:
Others:
Oracle® Database Globalization Support Guide 10g Release 2 (10.2)
Anyway, I have taken this thing to heart and have been learning a lot about codepoints and various facts about character sets. We will be converting from WE8ISO8859P1 to AL32UTF8, a single-byte ISO 8859-1 West European somewhat basic (eg, simple) codepoint to the latest'n'greatest multibyte grandaddy of Universal character sets. The big issue is making sure that all the data stays the same, even if physically it is stored a little different. Along the way, I have had to come to terms with phrases such as "Lossy Data" and "Truncation Data". April Sims has been a great help in answering some questions and getting me to focus my searches. In the end, Metalink note 252352.1 laid out an example that finally opened my eyes for Lossy Data.
I was having a hard time understanding why one must go through WE8MSWIN1252 first to resolve lossy data. And what was meant by the notation in Metalink Note 260192.1:
if you have a WE8ISO8859P1 database and lossy then changing your WE8ISO8859P1 to WE8MSWIN1252 will most likly solve you lossy [sic]
Why? Well, the Euro example in note 252352.1 is great. To paraphrase, the Euro is stored as codepoint 128 in WE8ISO8859P1. In this case, that character set essentially has an empty hole there, no previous definitions for that codepoint. So if one were to attempt to select it in some kind of "native" WE8ISO8859P1 environment, there would be nothing. So how did it get there? It is passed in from a different character set. And if using that same character set on the client side, you can easily retrieve the correct symbol.
Since WE8WSWIN1252 is a superset of WE8ISO8859P1 (meaning it has all the same codepoints for existing symbols, and then some), and it is most likely that a client in the US inserted data with a US Microsoft character set, the codepoint will be mapped to a valid symbol. Since that codepoint is "valid" in WE8MSWIN1252, it can then be converted to any other character set that also supports that symbol.
Where this falls apart (the "likly" part) is if the client using a "foreign" character set inserts a symbol that is mapped to a completely different symbol (via the codepoint) in our target client character set. In that case, you are screwed. =) Unless you would want to convert that data manually for each case of incompatible character sets. Not I.
Truncation data is a little easier to grasp. Metalink note 119119.1 mentions the English pound sign (£):
The pound sign for example is indeed 163 ( A3 in hexadecimal) in the WE8ISO8859P1 and WE8MSWIN1252 charactersets, but in AL32UTF8 the pound sign is code 49827 (C2 A3 in hex).
In our case, the fix is handled by setting nls_length_semantics=CHAR (opposed to the default BYTE). We also found out through experimentation that nls_length_semantics is completely ignored for SYS and SYSTEM objects; in fact, the data dictionary is always with BYTE semantics. Interesting, eh?
Metalink Notes:
- 260192.1 Changing WE8ISO8859P1/ WE8ISO8859P15 or WE8MSWIN1252 to (AL32)UTF8
- 225938.1 Database Character Set Healthcheck
- 252352.1 Euro Symbol Turns up as Upside-Down Questionmark
- 279068.1 Usage of Character Set Scanner Utility (csscan) version 1.1
- 137127.1 Character Sets, Code Pages, Fonts and the NLS_LANG Value
- 241047.1 The Priority of NLS Parameters Explained
- 158577.1 NLS_LANG Explained (How does Client-Server Character Conversion Work?)
- 179133.1 The correct NLS_LANG in a Windows Environment
- 119119.1 AL32UTF8 / UTF8 (Unicode) Database Character Set Implications
Others:
Oracle® Database Globalization Support Guide 10g Release 2 (10.2)
Friday, August 01, 2008
Software Configuration Manager
I learned that Oracle Support has a couple Oracle blogs up:
http://blogs.oracle.com/supportportal/
http://blogs.oracle.com/Support/
I think I was just checking on them when I noticed that the supportPortal was being a little more aggressive looking for more feedback. Over the years, I have participated in many of the Survey's that Oracle sends you after you close an SR, and have made no small number of comments (both good and bad) to a growing number of contacts I have established within Oracle, so this seemed like yet another opportunity to voice my opinion. I think I do that much more when I get the impression Oracle is actually listening; I don't share the same thoughts here because, well.... I highly doubt Oracle is reading this. =) No offense.
So let me blur the lines a little. First, I think Oracle is doing a much better job at collecting client feedback. However, the more critical part is what they are doing with that feedback. Unfortunately, for us end users, those comments seem to disappear into a black box, never to surface again (much like Enhancement Requests). I chatted with some really nice folks at Oracle (Kristal and Cherryl) and they both reassured me that Oracle is taking our comments seriously, and in fact hinted that the ongoing evolution of Metalink 3 is largely due to user feedback, and even newer things are coming down the pike.
Taking the invitation from the SupportPortal blog seriously, I called up for an appointment and got hooked up right away. I think the lady running the feedback session was working at home (heard a small child in the background), and she had some setup work to accomplish before we got down to it. But 90 minutes later I had made numerous observations about many aspects of the Software Configuration Manager and Metalink 3 in general. It was refreshing (and exhausting) to not only state some things I either liked or did not, but also to demonstrate via a recorded web session. Very satisfying; I recommend this to anyone who deals with metalink on a frequent basis.
So, what did I say? In general, I like the eye-candy of the new interface, but it is also a little slow and sometimes downright confusing. For instance, there is no way to bookmark bugs. The options to configure the layout are a nice thought, but still a little limited; it reminds me of how limited blogger is compared to WordPress, if that helps at all. Don't get me wrong, I think it is quite handy to configure the layout, but I really look forward to improvements. The key point, Software Configuration Manager, is a great concept. Unfortunately, I have some political roadblocks to fully utilizing SCM with my employer, which makes it a little frustrating when attempting to comment on the SCM because it is not SCM's fault at all. =) We require that we either have separate CSI numbers, or implement some sort of access to control such that not all our end-users have equal access to a site-wide license. But if we ignore that for a little while, I think SCM is awesome and has a lot of potential. Just the fact that it helps speed up SRs significantly is a big bonus. I have heard some misgivings about allowing Oracle to retrieve information in such a fashion, even if only passively, but I personally do not have an issue with it.
I do not like how some of the questions are completely inane. For instance, 99% of my SRs are for the Oracle Server - Enterprise Edition. 99% of the time I forget what exactly my "Product" is called and have to hunt for it amongst a list of several hundred product options, most of which we are not even licensed for. Why? What about a list of 10 most recent products for my CSI? What about a list that only contains products I am licensed for? And when attempting to find a category that somewhat matches the subject of an SR, I hate having to select one from the list of 30 or so. I want a searchable list. And a fast search at that; the LOVs in the patch pages really suck.
I concluded by saying that it would be vitally important to me see how Oracle is using this feedback. I have repeated this message to my Oracle contacts and with nearly all feedback I give to Oracle. I long to see what difference I make. Don't you?
Kudos to the teams involved in the feedback gathering operations. And BIG PROPS to Cheryl, Joey and Kristal (all of Oracle).
http://blogs.oracle.com/supportportal/
http://blogs.oracle.com/Support/
I think I was just checking on them when I noticed that the supportPortal was being a little more aggressive looking for more feedback. Over the years, I have participated in many of the Survey's that Oracle sends you after you close an SR, and have made no small number of comments (both good and bad) to a growing number of contacts I have established within Oracle, so this seemed like yet another opportunity to voice my opinion. I think I do that much more when I get the impression Oracle is actually listening; I don't share the same thoughts here because, well.... I highly doubt Oracle is reading this. =) No offense.
So let me blur the lines a little. First, I think Oracle is doing a much better job at collecting client feedback. However, the more critical part is what they are doing with that feedback. Unfortunately, for us end users, those comments seem to disappear into a black box, never to surface again (much like Enhancement Requests). I chatted with some really nice folks at Oracle (Kristal and Cherryl) and they both reassured me that Oracle is taking our comments seriously, and in fact hinted that the ongoing evolution of Metalink 3 is largely due to user feedback, and even newer things are coming down the pike.
Taking the invitation from the SupportPortal blog seriously, I called up for an appointment and got hooked up right away. I think the lady running the feedback session was working at home (heard a small child in the background), and she had some setup work to accomplish before we got down to it. But 90 minutes later I had made numerous observations about many aspects of the Software Configuration Manager and Metalink 3 in general. It was refreshing (and exhausting) to not only state some things I either liked or did not, but also to demonstrate via a recorded web session. Very satisfying; I recommend this to anyone who deals with metalink on a frequent basis.
So, what did I say? In general, I like the eye-candy of the new interface, but it is also a little slow and sometimes downright confusing. For instance, there is no way to bookmark bugs. The options to configure the layout are a nice thought, but still a little limited; it reminds me of how limited blogger is compared to WordPress, if that helps at all. Don't get me wrong, I think it is quite handy to configure the layout, but I really look forward to improvements. The key point, Software Configuration Manager, is a great concept. Unfortunately, I have some political roadblocks to fully utilizing SCM with my employer, which makes it a little frustrating when attempting to comment on the SCM because it is not SCM's fault at all. =) We require that we either have separate CSI numbers, or implement some sort of access to control such that not all our end-users have equal access to a site-wide license. But if we ignore that for a little while, I think SCM is awesome and has a lot of potential. Just the fact that it helps speed up SRs significantly is a big bonus. I have heard some misgivings about allowing Oracle to retrieve information in such a fashion, even if only passively, but I personally do not have an issue with it.
I do not like how some of the questions are completely inane. For instance, 99% of my SRs are for the Oracle Server - Enterprise Edition. 99% of the time I forget what exactly my "Product" is called and have to hunt for it amongst a list of several hundred product options, most of which we are not even licensed for. Why? What about a list of 10 most recent products for my CSI? What about a list that only contains products I am licensed for? And when attempting to find a category that somewhat matches the subject of an SR, I hate having to select one from the list of 30 or so. I want a searchable list. And a fast search at that; the LOVs in the patch pages really suck.
I concluded by saying that it would be vitally important to me see how Oracle is using this feedback. I have repeated this message to my Oracle contacts and with nearly all feedback I give to Oracle. I long to see what difference I make. Don't you?
Kudos to the teams involved in the feedback gathering operations. And BIG PROPS to Cheryl, Joey and Kristal (all of Oracle).
Thursday, June 26, 2008
create tablespace ddl
I recently went looking for ways to recreate tablespaces, as we prepare to migrate our production database to a different character set (necessitated by the vendor, grrr). DBMS_METADATA is really cool, and I like it alot. Data Pump essentially uses DBMS_METADATA (I believe), so I thought one of those cool tools would be the way to go. Close, but not quite. If you do a complete export/import of your database, you end up reorganizing all your data, in essence defragging the whole thing. Which is good (nay, GREAT!). But Data Pump only lets you create tablespaces with the original datafile sizes.
Granted, considering that this is Production, prudence would say that one will have tested this several times (as we are currently doing) before actually running the process in Production. And in doing so, one will know exactly how large the files should be, and can in fact extract the ddl (metadata api, data pump, TOAD, etc) from the test database. As a simple exercise, I still want a way to generate slightly customized tablespace ddl. So I ended up writing my own.
Some of the key things I wanted was to be to dynamically adjust the initial sizes of the datafiles. Also, for the sake of portability, I added a runtime parameter for maxsize. I am still torn about LMT being locked into UNIFORM, but since that is all we use here, I am not too worried for us. =) But I do want LMT and ASSM to be default; for 10g, I see no reason to turn these off at all.
Blogspot is going to horribly screw up the formatting for this, and I have not found a way around it. Sorry.
Granted, considering that this is Production, prudence would say that one will have tested this several times (as we are currently doing) before actually running the process in Production. And in doing so, one will know exactly how large the files should be, and can in fact extract the ddl (metadata api, data pump, TOAD, etc) from the test database. As a simple exercise, I still want a way to generate slightly customized tablespace ddl. So I ended up writing my own.
Some of the key things I wanted was to be to dynamically adjust the initial sizes of the datafiles. Also, for the sake of portability, I added a runtime parameter for maxsize. I am still torn about LMT being locked into UNIFORM, but since that is all we use here, I am not too worried for us. =) But I do want LMT and ASSM to be default; for 10g, I see no reason to turn these off at all.
Blogspot is going to horribly screw up the formatting for this, and I have not found a way around it. Sorry.
REM create_tablespace_ddl.sql
REM Charles Schultz
REM 25-Jun-2008
REM
REM Generates tablespace DDL
REM Assumes the following:
REM - Excluding ('SYSTEM','SYSAUX','USERS','TEMP','TOOLS','UNDOTS')
REM - All LMT
REM - All ASSM
REM - All datafiles AUTOEXTEND, initial size = half of current, or size of initial extent, maxsize defaults to 4000M
REM
REM Retains INITIAL_EXTENT settings for LMT
REM
REM Builds 5 inline materialized views
REM - tablespace_specs: gets the data from dba_tablespaces and dba_data_files, ordering by file_id
REM - first_ddl: generates the initial "CREATE TABLESPACE" statement, along with the first datafile
REM - last_ddl: generates the tablespace options (LMT, ASSM)
REM - mid_ddl: generates statements for any datafiles that are not the first
REM - tablespace_ddl: Combines all ddl, ordered by tablespace name and section
REM
REM Limitations
REM - Does not allow for BIGFILES, but this is a trivial change
REM - Does not allow for LMT Autoallocate. Again, not a difficult modification
PROMPT This script will generate tablespace ddl, making the following assumptions:
PROMPT - All tablespaces are Locally Managed (LMT) with UNIFORM sizes
PROMPT - All tablespaces have automatic segment space management (ASSM)
PROMPT - All datafiles autoextend
PROMPT
accept sqlfile default 'create_tablespaces.sql' char prompt 'Enter name for generated sql file [create_tablespaces.sql]: '
accept initial_size default '50' char prompt 'Percent by which to shrink datafile initial sizes, 0 being none [50]: '
accept max_size default '4000' char prompt 'MAXSIZE of autoextend datafiles, in megabytes [4000]: '
set head off newpage 0 pages 0 verify off feed off lines 200
spool &SQLFILE
with tablespace_specs as
(select row_number() over (partition by tb.tablespace_name order by df.file_id) rn_up,
row_number() over (partition by tb.tablespace_name order by df.file_id desc) rn_down,
tb.BIGFILE,
tb.TABLESPACE_NAME,
df.FILE_NAME,
greatest(df.BYTES*((100-greatest(least(&INITIAL_SIZE,100),0))/100),tb.INITIAL_EXTENT) file_size,
df.INCREMENT_BY*tb.block_size INCREMENT_BY,
tb.INITIAL_EXTENT
from dba_tablespaces tb,dba_data_files df
where tb.TABLESPACE_NAME = df.TABLESPACE_NAME
and tb.tablespace_name not in ('SYSTEM','SYSAUX','USERS','TEMP','TOOLS','UNDOTS')
),
first_ddl as
(select tablespace_name,
1 section_num,
'create tablespace '||tablespace_name||' datafile '''||file_name||''' size '||file_size||' autoextend on next '||increment_by||' maxsize '||&MAX_SIZE||'M' text
from tablespace_specs where rn_up = 1),
last_ddl as
(select tablespace_name,
3 section_num,
'EXTENT MANAGEMENT LOCAL UNIFORM SIZE '||INITIAL_EXTENT||' SEGMENT SPACE MANAGEMENT AUTO;' text
from tablespace_specs where rn_down = 1),
mid_ddl as
(select tablespace_name,
2 section_num,
','''||file_name||''' size '||file_size||' autoextend on next '||increment_by||' maxsize '||&MAX_SIZE||'M' text
from tablespace_specs where rn_up > 1),
tablespace_ddl as
(select text,tablespace_name, section_num from first_ddl
union
select text,tablespace_name, section_num from mid_ddl
union
select text,tablespace_name, section_num from last_ddl
order by tablespace_name, section_num)
select text from tablespace_ddl
/
spool off
Wednesday, June 11, 2008
Stragg
Tom Kyte popularized the STRAGG function, a sometimes critical piece missing from Oracle's arsenal of cool commands, has been bandied about quite a bit:
http://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:21961626004
We were recently tasked with using a STRAGG-like function but with dynamic delimiters. Tom Kyte's standard STRAGG does not handle that. Enter Adrian Billington, who took it one step further (look for "10g goes one further"). I really liked his example, and it caused me to research the COLLECTION functionality a little deeper. A little - I do not claim to have mastered it by any means.
Unfortunately for us, we seem to be hitting bug 4381035 (not exactly sure about that, but the best hit I can find). This is under 10.2.0.2.
BANDEV_SQL > CREATE OR REPLACE TYPE oracle.stragg_type AS TABLE OF VARCHAR2(4000);
/
2
Type created.
BANDEV_SQL > CREATE OR REPLACE FUNCTION oracle.stragg (
nt_in IN stragg_type,
delimiter_in IN VARCHAR2 DEFAULT ','
) RETURN varchar2 IS
v_idx PLS_INTEGER;
v_str varchar2(32000);
v_dlm VARCHAR2(10);
BEGIN
v_idx := nt_in.FIRST;
WHILE v_idx IS NOT NULL LOOP
v_str := v_str || v_dlm || nt_in(v_idx);
v_dlm := delimiter_in;
v_idx := nt_in.NEXT(v_idx);
END LOOP;
RETURN v_str;
END stragg;
/
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Function created.
BANDEV_SQL > drop table test;
Table dropped.
BANDEV_SQL > create table test (a char(2), b char(1));
Table created.
BANDEV_SQL > insert into test values ('aa','1');
1 row created.
BANDEV_SQL > insert into test values ('aa','2');
1 row created.
BANDEV_SQL > insert into test values ('aa','3');
1 row created.
BANDEV_SQL > select a, oracle.stragg(CAST(COLLECT(b) as ORACLE.STRAGG_TYPE)) b from test group by a;
select a, oracle.stragg(CAST(COLLECT(b) as ORACLE.STRAGG_TYPE)) b from test group by a
*
ERROR at line 1:
ORA-00932: inconsistent datatypes: expected - got -
This was a bummer. After a bunch of reading and experimenting, I finally hit upon something that seems to be working:
BANDEV_SQL > select a, oracle.stragg(cast(multiset(select b from test) as oracle.STRAGG_TYPE),'|') b from test group by a;
A
--
B
--------------------------------------------------------------------------------------------------------------------------------------------
aa
1|2|3
I am not exactly sure why MULTISET is working better than COLLECT in this case; it seems to me that both should be producing nested tables.
http://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:21961626004
We were recently tasked with using a STRAGG-like function but with dynamic delimiters. Tom Kyte's standard STRAGG does not handle that. Enter Adrian Billington, who took it one step further (look for "10g goes one further"). I really liked his example, and it caused me to research the COLLECTION functionality a little deeper. A little - I do not claim to have mastered it by any means.
Unfortunately for us, we seem to be hitting bug 4381035 (not exactly sure about that, but the best hit I can find). This is under 10.2.0.2.
BANDEV_SQL > CREATE OR REPLACE TYPE oracle.stragg_type AS TABLE OF VARCHAR2(4000);
/
2
Type created.
BANDEV_SQL > CREATE OR REPLACE FUNCTION oracle.stragg (
nt_in IN stragg_type,
delimiter_in IN VARCHAR2 DEFAULT ','
) RETURN varchar2 IS
v_idx PLS_INTEGER;
v_str varchar2(32000);
v_dlm VARCHAR2(10);
BEGIN
v_idx := nt_in.FIRST;
WHILE v_idx IS NOT NULL LOOP
v_str := v_str || v_dlm || nt_in(v_idx);
v_dlm := delimiter_in;
v_idx := nt_in.NEXT(v_idx);
END LOOP;
RETURN v_str;
END stragg;
/
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Function created.
BANDEV_SQL > drop table test;
Table dropped.
BANDEV_SQL > create table test (a char(2), b char(1));
Table created.
BANDEV_SQL > insert into test values ('aa','1');
1 row created.
BANDEV_SQL > insert into test values ('aa','2');
1 row created.
BANDEV_SQL > insert into test values ('aa','3');
1 row created.
BANDEV_SQL > select a, oracle.stragg(CAST(COLLECT(b) as ORACLE.STRAGG_TYPE)) b from test group by a;
select a, oracle.stragg(CAST(COLLECT(b) as ORACLE.STRAGG_TYPE)) b from test group by a
*
ERROR at line 1:
ORA-00932: inconsistent datatypes: expected - got -
This was a bummer. After a bunch of reading and experimenting, I finally hit upon something that seems to be working:
BANDEV_SQL > select a, oracle.stragg(cast(multiset(select b from test) as oracle.STRAGG_TYPE),'|') b from test group by a;
A
--
B
--------------------------------------------------------------------------------------------------------------------------------------------
aa
1|2|3
I am not exactly sure why MULTISET is working better than COLLECT in this case; it seems to me that both should be producing nested tables.
Friday, May 16, 2008
I am digging this learning thing
One of the mixed blessings of Oracle is that there is always something you can learn, no matter how good you are, or how much experience you have in the field. And I still have a lot to learn.
This week I had the unique pleasure of speaking directly with an Oracle Developer about an SR I had file in regards to query. In fact, this query is what lead me down the path of last week's discovery. We have a 3rd party ERP (Banner from SunGard), and one of the tables is GJBCOLR, a "collector" table for various jobs. My understanding is that it used like a scratch table, and jobs will routinely insert rows, work on them, then delete them. Sound like a TEMPORARY table at all? I am told that SunGard needs this to be a permanent table for some reason, and I am still trying to hunt down that reason. But I digress.
This table routinely has a high water mark. In the SR I filed for Oracle, we observed that the stats pegged it with 21 million consisten reads (HHWM), even though it really had about 290 rows. Sure, we can drop the HHWM (ie, truncate table, alter table shrink, etc) as explain in my previous post, but the table will just grow again due to how Banner uses it. So my next thought was to delete the stats and force dynamic sampling. To my horror, the CBO was able to find a low cost FTS, favored over a Index access by a factor of 10. According to my developer buddy, he highly suspects this is due to a bug in dynamic sampling, since the 10053 trace shows the dynamic stats put the number of blocks at an incredibly low 95. We are talking about OSEE 10.2.0.2, and he found a bug fix in 10.2.0.3 that is only slightly related but he firmly believes it probably affects the same thing. One of the workarounds I wanted to run past the developer is faking the stats, putting them enormously high so that indexes are always favored, and he tentatively agreed. So hopefully that will solve our problem.
However, realizing what a rare opportunity I had to be talking to a developer, I asked him a few questions about the optimizer and the particular query we were looking at. I am sure a month from now I will be smacking my head about not asking him other questions, but hey.... =)
The query plan looks like this:
I include the predicate information because that becomes important later on. Not knowing much about the FILTER operation other than that it works on two children, I asked him how it works. It is almost exactly like a nested loop operation. However, he specifically said the Hash Join is a source and the GJBCOLR subquery is a probe, which indicates to me a hash join, not a nested loop. But, like a nested loop, each result of the HJ child is compared to the GJBCOLR subquery, hence the subquery is executed 210K times. I asked why the HJ child is not second (as opposed to first), and he mentioned it had to do with how Oracle used part of the GJBCOLR subquery as a filter predicate. I am not entirely positive why that is the intended result, but I hope to chew on it a bit more. As you can see, the subquery was transformed into a EXISTS clause (not present in the original), and I believe that is part of the answer.
Now if only I could talk to a developer everytime I file an SR.....
This week I had the unique pleasure of speaking directly with an Oracle Developer about an SR I had file in regards to query. In fact, this query is what lead me down the path of last week's discovery. We have a 3rd party ERP (Banner from SunGard), and one of the tables is GJBCOLR, a "collector" table for various jobs. My understanding is that it used like a scratch table, and jobs will routinely insert rows, work on them, then delete them. Sound like a TEMPORARY table at all? I am told that SunGard needs this to be a permanent table for some reason, and I am still trying to hunt down that reason. But I digress.
This table routinely has a high water mark. In the SR I filed for Oracle, we observed that the stats pegged it with 21 million consisten reads (HHWM), even though it really had about 290 rows. Sure, we can drop the HHWM (ie, truncate table, alter table shrink, etc) as explain in my previous post, but the table will just grow again due to how Banner uses it. So my next thought was to delete the stats and force dynamic sampling. To my horror, the CBO was able to find a low cost FTS, favored over a Index access by a factor of 10. According to my developer buddy, he highly suspects this is due to a bug in dynamic sampling, since the 10053 trace shows the dynamic stats put the number of blocks at an incredibly low 95. We are talking about OSEE 10.2.0.2, and he found a bug fix in 10.2.0.3 that is only slightly related but he firmly believes it probably affects the same thing. One of the workarounds I wanted to run past the developer is faking the stats, putting them enormously high so that indexes are always favored, and he tentatively agreed. So hopefully that will solve our problem.
However, realizing what a rare opportunity I had to be talking to a developer, I asked him a few questions about the optimizer and the particular query we were looking at. I am sure a month from now I will be smacking my head about not asking him other questions, but hey.... =)
The query plan looks like this:
--------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes |TempSpc| Cost (%CPU)|
--------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 210K| 50M| | 14353 (6)|
| 1 | SORT ORDER BY | | 210K| 50M| 106M| 14353 (6)|
|* 2 | FILTER | | | | | |
|* 3 | HASH JOIN RIGHT OUTER| | 210K| 50M| | 2412 (18)|
| 4 | TABLE ACCESS FULL | FTVDTYP | 54 | 378 | | 3 (0)|
| 5 | TABLE ACCESS FULL | FFBMAST | 210K| 49M| | 2387 (17)|
|* 6 | TABLE ACCESS FULL | GJBCOLR | 1 | 50 | | 29 (4)|
--------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - filter( EXISTS (SELECT /*+ */ 0 FROM "GENERAL"."GJBCOLR"
"GJBCOLR" WHERE "GJBCOLR_VALUE"=:B1 AND "GJBCOLR_ONE_UP_NO"=4549363 AND
"GJBCOLR_JOB"='FFRAGRP' AND "GJBCOLR_NUMBER"='01'))
3 - access("FFBMAST_OTHER_SRC_DTYP_NUM"="FTVDTYP_SEQ_NUM"(+))
6 - filter("GJBCOLR_VALUE"=:B1 AND "GJBCOLR_ONE_UP_NO"=4549363 AND
"GJBCOLR_JOB"='FFRAGRP' AND "GJBCOLR_NUMBER"='01')
I include the predicate information because that becomes important later on. Not knowing much about the FILTER operation other than that it works on two children, I asked him how it works. It is almost exactly like a nested loop operation. However, he specifically said the Hash Join is a source and the GJBCOLR subquery is a probe, which indicates to me a hash join, not a nested loop. But, like a nested loop, each result of the HJ child is compared to the GJBCOLR subquery, hence the subquery is executed 210K times. I asked why the HJ child is not second (as opposed to first), and he mentioned it had to do with how Oracle used part of the GJBCOLR subquery as a filter predicate. I am not entirely positive why that is the intended result, but I hope to chew on it a bit more. As you can see, the subquery was transformed into a EXISTS clause (not present in the original), and I believe that is part of the answer.
Now if only I could talk to a developer everytime I file an SR.....
Friday, May 09, 2008
Learned a couple things this week
We had a couple examples of how a high watermark affects FTS, especially when that FTS gets executed on the wrong side of a FILTER operation. Which got me to delving into the black magic of high water marks.
First off, Oracle only mentions them in passing. If you search for LHWM (low high water mark), you will get a couple hits in metalink for a bug and nothing in the documentation. So, what the heck is a low high watermark anyway? And a high high watermark?
Apparently, after a couple lessons from K Gopalakrishna, Ric Van Dyke and couple others, the LHWM is the boundary at which all blocks below it are gauranteed to be formatted and used at least once. Above the HHWM we are gauranteed that all blocks are unformatted and never used. Between the two we have the "grey area"; namely, a mixture of blocks that have and have not been used. I believe there are two bitmaps (or two levels of bitmaps?), one that maintains the freelists for below the LWHM, and another for above the HHWM. On Full Table Scans, Oracle will scan everything below the LHWM, and then use the L2 bitmap to find blocks that need to be scanned above the LHWM.
Oracle seems to have gone to great lenghts to make these two watermarks as confusing as possible. In the documentation, and when using packages like dbms_space, we are led to believe there is only one HWM, even for an ASSM tablespace.
Thankfully, folks have discovered "dump file". Norman Dunbar gives a couple quick examples. This is good stuff; there is a lot in the dump files that would warrant a closer look. But for right now, let's look at the watermark bits of information:
alter system dump datafile 5 block min 1 block max 100;
Extent Control Header
-----------------------------------------------------------------
Extent Header:: spare1: 0 spare2: 0 #extents: 1 #blocks: 8
last map 0x00000000 #maps: 0 offset: 2716
Highwater:: 0x01403030 ext#: 1 blk#: 7 ext size: 8
#blocks in seg. hdr's freelists: 0
#blocks below: 7809
mapblk 0x00000000 offset: 1
Disk Lock:: Locked by xid: 0x000a.013.000003a4
--------------------------------------------------------
Low HighWater Mark :
Highwater:: 0x01403030 ext#: 992 blk#: 7 ext size: 8
#blocks in seg. hdr's freelists: 0
#blocks below: 7809
mapblk 0x01402aa1 offset: 177
Level 1 BMB for High HWM block: 0x01403019
Level 1 BMB for Low HWM block: 0x01403019
This is not necessarily the world's best example, but heck, it is raw data. =) I have yet to decipher what all that means, but I am working on it. I believe that the LHWM "blocks below" will always be less than the HHWM "blocks below", but the difference would count the number of blocks in that grey area.
As an aside, I also found that you can dump all the file headers with:
alter session set events 'immediate trace name file_hdrs level 10';
The level is a bit confusing, as I tried various numbers and the resulting trace seem to be exactly the same.
On the RMAN side, we came across a peculiar curiosity. With no recovery catalog, if you recover a controlfile and have archive logs beyond the time of the controlfile, we found that on OSEE 10.2.0.1, RMAN is not smart enough to roll forward if you specify an "UNTIL TIME" in the future relative to the controlfile. Instead, you will see errors that resemble messages if the backup is corrupted; RMAN will attempt to go backwards in time to find a file that matches an appropiate SCN. Of course, it will never find one. This behavior is "fixed" in OSEE 10.2.0.4 (we did not try any of the intermediate patch levels), and RMAN will now roll the database forward beyond the controlfile time.
First off, Oracle only mentions them in passing. If you search for LHWM (low high water mark), you will get a couple hits in metalink for a bug and nothing in the documentation. So, what the heck is a low high watermark anyway? And a high high watermark?
Apparently, after a couple lessons from K Gopalakrishna, Ric Van Dyke and couple others, the LHWM is the boundary at which all blocks below it are gauranteed to be formatted and used at least once. Above the HHWM we are gauranteed that all blocks are unformatted and never used. Between the two we have the "grey area"; namely, a mixture of blocks that have and have not been used. I believe there are two bitmaps (or two levels of bitmaps?), one that maintains the freelists for below the LWHM, and another for above the HHWM. On Full Table Scans, Oracle will scan everything below the LHWM, and then use the L2 bitmap to find blocks that need to be scanned above the LHWM.
Oracle seems to have gone to great lenghts to make these two watermarks as confusing as possible. In the documentation, and when using packages like dbms_space, we are led to believe there is only one HWM, even for an ASSM tablespace.
Thankfully, folks have discovered "dump file". Norman Dunbar gives a couple quick examples. This is good stuff; there is a lot in the dump files that would warrant a closer look. But for right now, let's look at the watermark bits of information:
alter system dump datafile 5 block min 1 block max 100;
Extent Control Header
-----------------------------------------------------------------
Extent Header:: spare1: 0 spare2: 0 #extents: 1 #blocks: 8
last map 0x00000000 #maps: 0 offset: 2716
Highwater:: 0x01403030 ext#: 1 blk#: 7 ext size: 8
#blocks in seg. hdr's freelists: 0
#blocks below: 7809
mapblk 0x00000000 offset: 1
Disk Lock:: Locked by xid: 0x000a.013.000003a4
--------------------------------------------------------
Low HighWater Mark :
Highwater:: 0x01403030 ext#: 992 blk#: 7 ext size: 8
#blocks in seg. hdr's freelists: 0
#blocks below: 7809
mapblk 0x01402aa1 offset: 177
Level 1 BMB for High HWM block: 0x01403019
Level 1 BMB for Low HWM block: 0x01403019
This is not necessarily the world's best example, but heck, it is raw data. =) I have yet to decipher what all that means, but I am working on it. I believe that the LHWM "blocks below" will always be less than the HHWM "blocks below", but the difference would count the number of blocks in that grey area.
As an aside, I also found that you can dump all the file headers with:
alter session set events 'immediate trace name file_hdrs level 10';
The level is a bit confusing, as I tried various numbers and the resulting trace seem to be exactly the same.
On the RMAN side, we came across a peculiar curiosity. With no recovery catalog, if you recover a controlfile and have archive logs beyond the time of the controlfile, we found that on OSEE 10.2.0.1, RMAN is not smart enough to roll forward if you specify an "UNTIL TIME" in the future relative to the controlfile. Instead, you will see errors that resemble messages if the backup is corrupted; RMAN will attempt to go backwards in time to find a file that matches an appropiate SCN. Of course, it will never find one. This behavior is "fixed" in OSEE 10.2.0.4 (we did not try any of the intermediate patch levels), and RMAN will now roll the database forward beyond the controlfile time.
Subscribe to:
Posts (Atom)
