RE: Improving query performance further
Date: Thu, 13 Oct 2022 08:59:23 -0400
Message-ID: <021801d8df03$a0907fb0$e1b17f10$_at_rsiz.com>
If I understand you correctly, the three sets of data are versions of the same rows such that one set of rows is in use at a time.
IF that is the case, you can eliminate the partitioning complications altogether by using three separate tables with alter synonym to identify the correct table to currently be in use. Before you fire off the alter synonym, the freshly loaded (or swapped in) table can be force read through the SGA so it has nothing like any delayed block cleanouts or other lingering references to UNDO and you can run all the best of whatever statistics (including histograms or not) best serve the CBO.
Removing the artificial PART_COL (which is duplicate information in each row replaced in content value by the NAME of the table) would also make the better of max_val, low_val or low_val, max_val the appropriate primary key (which is somewhat smaller, but going from once per row to once per page as shown by the compress results means only a tiny bit smaller) and the CBO would of course have a good chance to pick an optimal plan, which may still be better re-writing the query as Sayan described, but without the part_col at all.
This approach also avoids any chance that global statistics get in the way or that side effects on the use of the “actively being read” partition by the truncate/reload of another partition, because other than sometimes being referenced by the same synonym in turn, the tables no longer have any relationship with each other from the viewpoint of the RDBMS dictionary.
Since the distinct count (115K-ish) is nearly the row count (150K-ish), compressing the value parts can’t buy much, but it’s nearly free and low has fewer distinct than high. If that’s usual and nothing else mitigates otherwise, put low first and compress 1. (But that is slicing the onion pretty thin.)
IF any of the rows have high value less than low value I’m supposing those “illegal” rows are dealt with on load. IF NOT let us know.
Finally (hoping I didn’t miss this in the thread), is “column1” the only column queried in this way and if so, is column1 very long? IF column1 dominates the use of this table AND it is some reasonable size (and someone already mentioned this) it MAY be useful (an experiment would confirm or deny) adding it to the index. It can only help when you get a row, so that seems unlikely in your actual case.
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of yudhi s
Sent: Wednesday, October 12, 2022 3:35 PM
To: jlewisoracle_at_gmail.com
Cc: Oracle L
Subject: Re: Improving query performance further
Thank you so much for the details.
Regarding the table design, the partition column PART_COL is holding three distinct values with equal number of rows in each partition. Also this table is a truncate + load kind of master table in which data is loaded in one of the partitions by truncating that partition daily. And almost all the rows in each of the partitions are mostly the same with very minimal difference. So basically the partition key i.e part_col is not a business column but it's helping the application to cater the 24 by 7 availability of master data without being impacted by the truncate+load process. I understand it could have been done with just two partitions but not sure why we are maintaining three partitions here. No update/delete is performed in this table. The column part_col is helping to maintain the primary key along with max_val, low_val, so we can say basically the combination of max_val and min_valis unique only for a specific partition.
The oracle version is 19.15.
I tried on dev by copying some sample data, but i see by compressing the index compress 1(i.e. part_col) , its just reducing the consistent gets from ~82 to ~79 and same effect is observed , if i move the column part_col to the end of the primary key i.e. index on (min_val, max_val,part_col). Even just indexing (min_val,max_val). So it seems it will not help to a great extent in minimizing the IO or response time.
Also seeing the data pattern , i see the input bind value is varying a lot and for few of the incoming binds are closer to MAX_VAL column and in those cases the index on (MAX_VAL , MIN_VAL) will help , however many input binds are closer to MIN_VAL column value, so they will suffer with that index order. And for those index on reverse order i.e. (MIN_VAL, MAX_VAL) will be suitable. There seems no easy choice here.
Also for the majority of the values, the first 6 digits are the same for column max_val and min_val, but I do see few of the cases where it's not the same.
Below is the statistics of two executions I manually did for two sample bind variables. But as i mentioned the current execution yields a single row only 10% of the time and rest of the time it results in zero rows. So below two executions were giving zero rows only for selected input binds.
Plan hash value: 692467662
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
| 0 | SELECT STATEMENT | | 1 | | 0 |00:00:00.01 | 169 |
| 1 | PARTITION RANGE SINGLE | | 1 | 20548 | 0 |00:00:00.01 | 169 |
| 2 | TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| TABLE1 | 1 | 20548 | 0 |00:00:00.01 | 169 |
|* 3 | INDEX RANGE SCAN | PK_TABLE1 | 1 | 20548 | 0 |00:00:00.01 | 169 |
Predicate Information (identified by operation id):
3 - access("PART_COL"=:B2 AND "MAX_VAL">=:B1 AND "MIN_VAL"<=:B1) filter("MAX_VAL">=:B1)
Plan hash value: 692467662
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers | Reads |
| 0 | SELECT STATEMENT | | 1 | | 0 |00:00:01.99 | 37 | 37 |
| 1 | PARTITION RANGE SINGLE | | 1 | 20548 | 0 |00:00:01.99 | 37 | 37 |
| 2 | TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| TABLE1 | 1 | 20548 | 0 |00:00:01.99 | 37 | 37 |
|* 3 | INDEX RANGE SCAN | PK_TABLE1 | 1 | 20548 | 0 |00:00:01.99 | 37 | 37 |
Predicate Information (identified by operation id):
3 - access("PART_COL"=:B2 AND "MAX_VAL">=:B1 AND "MIN_VAL"<=:B1) filter("MAX_VAL">=:B1)
On Wed, Oct 12, 2022 at 6:18 PM Jonathan Lewis <jlewisoracle_at_gmail.com> wrote:
The table has three partitions, so it would be better to run the queries for each partition in turn. Because of the bind variables the optimizer might still use the global stats to estimate cost and/or cardinality, but there might be something about the partition-level stats that affect the choice of strategy (e.g the max(max_val - min_val) might be very large only in one partition out of three, and that partition may have only a very small number of rows).
You appear to be storing a 19 digit number as a varchar2(4000) - it's a bad idea (generally) to store numbers in character columns, and it's a bad idea to define character columns with lengths that are vastly longer than needed. The latter may not matter very much in this case with a tiny table, but ask Sayan to point you to his blog notes on the effects of row definitions being larger than the blocksize.
As Sayan said - can you give us some idea of how the data is treated - for example, does a row appear in one partition and then get from one partition to the next over time; is there something about the data that makes the first 6 digits of the values significant; are rows constantly updated on these column, or is there a constant trickle of inserts with a balancing number of deletes etc. The right strategy is often dictated by the data patterns and activity.
Which version of Oracle, by the way ?
Regards
Jonathan Lewis
On Wed, 12 Oct 2022 at 05:30, yudhi s <learnerdatabase99_at_gmail.com> wrote:
Thank you so much.
There are only three different distinct values for the partition key column PART_COL and thus a total of three partitions exists on this table. And each partition holding an equal amount of rows says ~150K each. And on an avg, this query returns zero rows almost 9 times out of 10 executions.
I will try to see the behaviour of IOT VS Keep Pool VS Domain indexes on one of lower environment. I have heard about it to be helpful on text column searches for like operator etc, but cant think of right away how this domain indexes will help the range/between predicate here.
_at_Sayan, Below is the output of the query :-
M0 M2 M1 M3 M7 M5 M819 9999999999 109999999999999 2226470000000000000 5599860000000000000 2226470099999999999 5599869999999999999
The statistics on this table is not up-to date , but below is how it looks like. So pattern wise its as below.
COLUMN_NAME NUM_DISTINCT LOW_VALUE HIGH_VALUE DENSITY NUM_NULLS PART_COL 3 80 C103 0.0000014697539337964 0 LOW_VAL 114992 323232363437303030303030303030303030303 5353939383630303030303030303030303030 0.00000869625713093085 0 HIGH_VALUE 115792 32323236343730303939393939393939393939 35353939383639393939393939393939393939 0.00000863617521072268 0
NUM_ROWS BLOCKS AVG_ROW_LEN
248942 3776 96
On Wed, Oct 12, 2022 at 6:41 AM Sayan Malakshinov <xt.and.r_at_gmail.com> wrote:
Hi Andy
Oh, yes, I also had a couple of examples of own odci indexes for such cases, but it would be great to see your variant!
Best regards,
Sayan Malakshinov
Oracle performance tuning expert
Oracle Database Developer Choice Award winner
Oracle ACE
http://orasql.org
On Wed, 12 Oct 2022, 02:05 Andy Sayer, <andysayer_at_gmail.com> wrote:
I have a demo for a domain index which allows this sort of query to be very fast. Essentially it creates an index entry for each value (with your desired granularity) between your min and max columns and does look ups based on that. The domain index knows what the granularity is so as well as :bind between x and y it queries new_domain_column between :bind - granularity and :bind + granularity.
Unfortunately, the demo is on the pc that I left in the UK when I moved to Canada. This might be enough information for you to get going with, otherwise I was planning on publishing my work when I visit the UK in December.
The other, easier but less guaranteed, way of combining two range predicates on different columns is to partition the index by one of the columns. Eg your index would be on some_key, start_date and your partitioning would be on end_date. Oracle would then need to do a range scan using your some_key and start_date filters as access predicates once per partition which matches your end_date filter. You’d have to decide which column is the partition key and how wide to make the partitions, and you might not gain a lot.
Thanks,
Andy
On Tue, Oct 11, 2022 at 5:38 PM, Sayan Malakshinov <xt.and.r_at_gmail.com> wrote:
I tuned a lot of similar queries: tried different methods (even spatial RTree indexes) and made a lot of tests, so I'll try to describe and summarize them all later, when I have more time, but for now just in short:
Let''s simplify it just to:
- table T(A,B)
- index (A,B)
- select * from T where :X between A and B
- ":X between A and B" is very selective
The main problem in this case is that you need to check all index entries A<:X, for example you have:
A | B
1 1 1 2 1 3
.....
1 100
2 107
2 111
....
2 215
3 204
3 206
....
3 299
...
998 99799
999 99801
...
999 99900
where each A has about 100 different B,
and your :X = 700, so, though there are just about 100 rows satisfying both conditions, INDEX RANGE SCAN with access predicate A<:X will scan 700*100 index entries, and filter out 699*100 of them by filter predicate (B>:X).
So we have just a few different situations and key points:
0. There is a limitation of max(B-A) and it's quite small, ie max(B[x]-A[x]) << max(B) - min(A) In this case, you can simply add a couple of additional predicates: select *
from T
where A between :X-:MAX_DIFF and :X
and B between :X-:MAX_DIFF and :X
;
So you''ll get a short search range for A (ACCESS predicates for the first column of the index). //For example, it may be short intervals or start/end dates of short time events, with check constraints like (b-a<N)
- There is a perfect positive correlation between A and B, ie if a2>a1, must be b2>b1
1.1 There is no intersection of any 2 intervals (Ax - Bx), (Ay - By)
This means that for any :X you need to find just 1 row, so you can stop to scan the index after the first found row. We can rewrite our query for that like this:
select *
from
(select * from T where A<=:X order by A desc fetch first 1 row only) -- to force IRS DESCENDING (IRS - index range scan) and stop after 1 row
where
B>:X -- our second predicate is one level upper, so we check just 1 row maximum
- IIRC there were some problems with IRS DESC in case of "fetch first", so I'll rewrite it to make it more reliable:
select *
from
(select *
from
(select * from T where A<=:X order by A desc) -- to force IRS DESCENDING
where rownum=1) -- and stop after 1 row
where
B>:X -- our second predicate
1.2 There may be intersections:
This means that you can't stop after the first found row satisfying our conditions. We need to find all of them. Unfortunately, there is no a documented method to make a dynamic SQL condition like "please stop, I don't need other rows", but we can simply use PL/SQL for that:
declare
X number := 34;
cursor c_1(p number) is
select * from T where A<=p order by A desc;
R c_1%rowtype;
begin
open c_1(X);
<<LL>>
loop
fetch c_1 into R;
if R.B>=X then
dbms_output.put_line(R.A ||' '||R.B); -- or pipe row or collect ROWIDS and return them as a collection, etc
else
exit LL;
end if;
end loop;
end;
/
As you can see, we stop fetching cursor C_1 when R.B becomes lower than X. Of course to make it more convenient you can wrap it into a pipelined pl/sql table function or even use it with INLINE pl/sql functions.
2. There is no correlation between A and B, ie if a2>a1, it doesn't mean that b2>b1 :
This one is the most difficult problem, because in case of unknown X we need to check all rows that have A<=X (or all rows with B>=X). Obviously, it would be great if we know some common patterns and distribution of A/B/X, for example:
- A and B are dates -- like table DT (start_date date, end_date date...)
- min(A) and min(B) ~ 10-20 years ago
- max(A) and max(B) ~ sysdate
- and X usually is in the range of [sysdate-10; sysdate]
- number of rows satisfying our conditions are pretty small In this case we could create index (B,A), just because we have less rows satisfying B>=X than A<=X.
Or sometimes it maybe useful to create 2 indexes (A) and (B) and use bitmap_and operation like this:
select/*+ index_combine(t) */ * from t where :X between a and b;
| Id | Operation | Name |
| 0 | SELECT STATEMENT | |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED| T |
| 2 | BITMAP CONVERSION TO ROWIDS | |
| 3 | BITMAP AND | |
| 4 | BITMAP CONVERSION FROM ROWIDS | |
| 5 | SORT ORDER BY | |
|* 6 | INDEX RANGE SCAN | IX_B |
| 7 | BITMAP CONVERSION FROM ROWIDS | |
| 8 | SORT ORDER BY | |
|* 9 | INDEX RANGE SCAN | IX_A | ------------------------------------------------------------
Predicate Information (identified by operation id):
6 - access("B">=TO_NUMBER(:X))
filter("B">=TO_NUMBER(:X))
9 - access("A"<=TO_NUMBER(:X))
filter("A"<=TO_NUMBER(:X))
===================================================================================
But in general, there is no silver bullet for such cases...
On Tue, Oct 11, 2022 at 10:58 PM Tim Gorman <tim.evdbt_at_gmail.com> wrote:
Have you considered making TABLE1 an IOT? No, not an "internet of things" but an "index-organized table"...
If the primary method of access is the primary key, and TABLE1 is has "narrow" rows (i.e. AVG_ROW_LEN less than 20-25 bytes or so), then an IOT could save on the number of logical reads. There's a lot of "if"s, but the best approach is not to think about it, but just go ahead and test it, side-by-side with the existing table. After all, it's only about ~100MB in size, right?
But, at the very least, it shouldn't be difficult to just put the table and the PK index together into the KEEP pool of the Buffer Cache? After all, although the ideal is to size the KEEP pool to accommodate the entire objects assigned to it, you certainly aren't required to size it that way. You just want to size it so that buffers flush out far more slowly than they do in the DEFAULT pool.
<rant>Too many DBAs think that the SGA should only use about 50% of the available physical memory on a host, which is nonsense. The Linux/UNIX operating systems only need a few GB of memory, and AWR can tell you unequivocally how much space is needed for PGA, so the SGA should be sized closer to the Oracle-imposed maximum of 90% of host physical memory. It's there. It's not being used. Use it. If I had a nickel for every unused GB of RAM on Oracle database servers, I could buy my own Hawaiian island.</rant>
Hope this helps!
Enjoy!
On 10/11/2022 2:04 PM, yudhi s wrote:
Hello Listers, We have a customer database on Oracle version 19C. We have a simple query as below. and as per current design is executing ~200 to 300 times per second and it's part of a bigger process and thus is one of the top consumers in that. Now as we are working to change the design to make the number of execution of this query lesser to help the process. But that needs much more impact analysis, so we were thinking of any possible easy way to make the individual execution of this query faster? Or say any structural change(new index etc.) which can further drop the IO/CPU requirement for individual execution of this query?
Currently this query is accessing table TABLE1 through a primary key which is on three columns (PART_COL,MIN_VALUE,MAX_VAL). The table is partitioned on column PART_COL. This table contains ~400K rows and is ~100MB in size. It's a master data kind of table.
SELECT column1 FROM TABLE1 WHERE PART_COL = :B2 AND :B1 BETWEEN MIN_VAL AND MAX_VALUE
Global Information
Status : DONE (ALL ROWS) Instance ID : 1 SQL Execution ID : 16777216 Execution Started : 10/11/2022 09:36:48 First Refresh Time : 10/11/2022 09:36:48 Last Refresh Time : 10/11/2022 09:36:48 Duration : .06173s Module/Action : SQL*Plus/- Program : sqlplus.exe Fetch Calls : 1
Binds
| Name | Position | Type | Value |
| :B2 | 1 | NUMBER | 2 |
| :B1 | 2 | VARCHAR2(4000) | XXXXXXXXXXX |
Global Stats
| Elapsed | Cpu | IO | Concurrency | Cluster | Fetch | Buffer | Read | Read |
| Time(s) | Time(s) | Waits(s) | Waits(s) | Waits(s) | Calls | Gets | Reqs | Bytes |
| 0.06 | 0.04 | 0.02 | 0.00 | 0.00 | 1 | 911 | 778 | 6MB |
SQL Plan Monitoring Details (Plan Hash Value=692467662)
| Id | Operation | Name | Rows | Cost | Time | Start | Execs | Rows | Read | Read | Activity | Activity Detail |
| | | | (Estim) | | Active(s) | Active | | (Actual) | Reqs | Bytes | (%) | (# samples) |
| 0 | SELECT STATEMENT | | | | | | 1 | | | | | |
| 1 | PARTITION RANGE SINGLE | | 10610 | 928 | | | 1 | | | | | |
| 2 | TABLE ACCESS BY LOCAL INDEX ROWID BATCHED | TABLE1 | 10610 | 928 | | | 1 | | | | | |
| 3 | INDEX RANGE SCAN | PK_TABLE1 | 10610 | 771 | | | 1 | | 770 | 6MB | | |
Predicate Information (identified by operation id):
3 - access("PART_COL"=TO_NUMBER(:B2) AND "MAX_VALUE">=:B1 AND "MIN_VAL"<=:B1) filter("MAX_VALUE">=:B1)
Statistics
37 recursive calls
0 db block gets
911 consistent gets
778 physical reads
41076 redo size
260 bytes sent via SQL*Net to client 489 bytes received via SQL*Net from client
1 SQL*Net roundtrips to/from client
28 sorts (memory)
0 sorts (disk)
0 rows processed
-- Best regards, Sayan Malakshinov Oracle performance tuning engineer Oracle ACE http://orasql.org -- http://www.freelists.org/webpage/oracle-lReceived on Thu Oct 13 2022 - 14:59:23 CEST