Attributed to Einstein but considered a dubious association, the quote is something like “Insanity is doing the same thing over and over and expecting different results”.
Probably … hopefully… it’s a familiar story.
I can’t be the only one, I’ve seen it so many places.
Connecting applications raise incidents against your database saying their application is erroring repeatedly whilst running a query and receiving an ORA-01013: user requested cancel of current operation.
Investigation shows that yes, sql monitoring can show a bunch of sequential executions of a particular query erroring with ORA-01013 – a really nice feature of sql monitoring that it captures these more often than not.
Nearly all of these errors occur around the 20 minute mark.
Do you have a 20 minute timeout?
Yes.
Why?
I’m not sure whether the mindset is inherited from other areas of the architecture. In the past, I happened to see this much more with applications involved in distributed transactions involving queues etc and which bring with them a whole lot more complexity.
Of course it all depends on what “normal” is for any query but an awful lot of these tend to have “normal” as not a million miles away from timeout and it doesn’t take much to nudge them over the limit – a different data distribution, some other load on the database, etc, etc
But I never understand what value that timeout brings.
No, it’s not “stuck”.
No, it’s not “hanging”.
It’s executing. It’s just not executing as fast as your expectations, possibly baseless, thought.
Of course sometimes we get a plan “flip” and something which was relatively fast becomes relatively slow.
But even then I think the proper application-level alerting and notification of that whilst allowing it to complete the execution where possible is infinitely better than spending hours getting hold of a DBA, raising the right sort of ticket, often telling them what you want profiled/baslined/patched all potentially done under a high priority because the timeout won’t let the relevant process run to completion.
After a frustrating period of quite a few minutes but less than an hour of trying various combinations, it seems to be that rebuilding an index partition online with compression is actually a two stage affair although if it explicitly states this in the documentation then I missed it. And I did an online search for similar and neither was the answer stated obviously in the first couple of pages of results or in the documents I was directed to.
I am doing some tests with enabling index key compression – a great feature by the way if it somehow passed you by over the many years it’s been around – and am expecting significant space savings.
In my particular case, it’s about time I played with some “newer” features and whilst the ADVANCED COMPRESSION HIGH scares me off a bit – not least because of some of the associated bug notes on Oracle Support for versions < 23cai – what I’ve read about ADVANCED COMPRESSION LOW holds for me fewer reservations as it seems pretty much like old style key compression but with some intelligence (heuristic not artificial) around optimal # of keys and whether that is suitable for every block.
So the two stage syntax would appear to be:
alter index <index_name> modify partition <partition_name> compress advanced low;
alter index <index_name> rebuild partition <partition_name> online parallel n;
Now as I write this, I have a horrible sense of deja vu that I’ve already written about this in the past but I had a look and I can’t find that either.
Recently I’ve found myself working on a set of data which requires a fair amount of crunching and “de-duplication” of data sent from the system of record (SOR) to identify what was the last event sent for a particular set of keys that meets certain unindexed filter criteria.
Which means I’m wanting to write code which drives from a set of data, and then joins to an inline aggregation view for each row from the driving rowsource.
The aggregation dataset is too large and open-ended to aggregate independently and hash join to the driving data set.
So by instinct I find myself wanting this to operate with a nested loop from the driving rowsource and for each row to push the join predicates into aggregation view.
However, I have found that specifying the correct push predicate hint syntax is a bit awkward and I try to be mindful of those who inevitably will come to the code at a later date to make a change.
There’s a good chance they won’t have heard of predicate pushing unfortunately and I strongly dislike leaving behind a legacy of heavily hinted code, although I make some exceptions.
Having sought out a high level 2nd opinion on a comment on Jonathan Lewis’s blog, I am now having a tendency to write such code using a LATERAL join, or at least experiment with it as an alternative approach.
You might have come across the lateral join not least from when it first started appearing in internal transformations for some ANSI SQL. But here is some documentation on this as a fully fledged SQL feature from 12c onwards:
But it should never be unexpected that any new feature might experience some degree of bugs and/or wrong result issues.
However I have been encouraged by the fact that in my circumstances, under my conditions, the results are good and the query is executing how I imagined it to with no further direction from hints, etc
With respect to the join conditions, you can specify the lateral view syntax in a couple of ways – either in the standard ON () join conditions, or the slightly strange scope of referencing the outer join conditions within the view itself, example here of the latter.
Rather than giving an emp-dept example of the lateral view which you can find elsewhere, not least in the links already provided and elsewhere, I’m going to provide a cutdown, sanitised snippet of the code I’m using in the real world although I’m sure this doesn’t really help with knowing when you might use this feature because I can’t give you all the background, data, objects from the real world either.
The snippet is from a INSERT SELECT. I’ve cut a whole bunch of information from the plan not least so you don’t see the horrible numbers that the optimizer throws up in terms of estimated rows and cost! (Which might/should be a warning sign that you wish to dig deeper into)
Really I want to just leave an idea of the scenario where I’ve been considering and experimenting with lateral joins.
with subq_active_positions as
(select /*+ */
rrs.f_id
, rrs.f_source
, rrs.reg
, max(rrs.c_id) keep (dense_rank first order by rrs.eventtime desc) c_id
, max(rrs.c_source) keep (dense_rank first order by rrs.eventtime desc) c_source
from regreport_event_state rrs
where 1 = 1
and reg = p_regulation
--
... some additional date filters
--
group by
rrs.f_id
, rrs.f_source
, rrs.reg
having ... some early filtering...
)
, subq_ap_with_eligible_events as
(select /*+ */ p.*, aud.*
from subq_active_positions p
inner join lateral
(select *
from
(select /*+ qb_name(sq1) */
max(rra.rowid) keep (dense_rank first order by rra.eventtime desc) rd
, max(rra.eventtype) keep (dense_rank first order by rra.eventtime desc) last_qualifying_event
, rra.f_id
, rra.f_source
, rra.c_id
, rra.c_source
, rra.reg
from regreport_event_report rra
where 1 = 1
--
and rra.f_id = p.f_id
and rra.f_source = p.f_source
and rra.c_id = p.c_id
and rra.c_source = p.c_source
and rra.reg = p.reg
--
group by
rra.f_id
, rra.f_source
, rra.c_id
, rra.c_source
, rra.reg
) x
where last_qualifying_event = 'Something_or_other') aud
on 1 = 1)
select *
from subq_ap_with_eligible_events;
===========================================================================================================
| Id | Operation | Name |
===========================================================================================================
| 0 | INSERT STATEMENT | |
| 1 | TEMP TABLE TRANSFORMATION | |
| 2 | PX COORDINATOR | |
| 3 | PX SEND QC (RANDOM) | :TQ10001 |
| 4 | LOAD AS SELECT (TEMP SEGMENT MERGE) | SYS_TEMP_0FD9E0321_A450058D |
| 5 | FILTER | |
| 6 | SORT GROUP BY | |
| 7 | PX RECEIVE | |
| 8 | PX SEND HASH | :TQ10000 |
| 9 | SORT GROUP BY | |
| 10 | PX BLOCK ITERATOR | |
| 11 | TABLE ACCESS STORAGE FULL | REGREPORT_EVENT_STATE |
| 12 | PX COORDINATOR | |
| 13 | PX SEND QC (RANDOM) | :TQ40001 |
| 14 | LOAD AS SELECT (HIGH WATER MARK BROKERED) | CANNED_REPORT_DATA |
| 15 | SORT AGGREGATE | |
| 16 | PARTITION HASH SINGLE | |
| 17 | TABLE ACCESS BY GLOBAL INDEX ROWID BATCHED | REGREPORT_EVENT_REPORT |
| 18 | INDEX RANGE SCAN | REGREPORT_EVENT_REPORT_PK |
| 19 | OPTIMIZER STATISTICS GATHERING | |
| 20 | BUFFER SORT | |
| 21 | PX RECEIVE | |
| 22 | PX SEND ROUND-ROBIN | :TQ40000 |
| 23 | SORT ORDER BY | |
| 24 | NESTED LOOPS | |
| 25 | VIEW | |
| 26 | NESTED LOOPS | |
| 27 | PX COORDINATOR | |
| 28 | PX SEND QC (RANDOM) | :TQ30000 |
| 29 | VIEW | |
| 30 | PX BLOCK ITERATOR | |
| 31 | TABLE ACCESS STORAGE FULL | SYS_TEMP_0FD9E0321_A450058D |
| 32 | PX COORDINATOR | |
| 33 | PX SEND QC (RANDOM) | :TQ20001 |
| 34 | VIEW | VW_LAT_75FFB6E0 |
| 35 | FILTER | |
| 36 | SORT GROUP BY | |
| 37 | PX RECEIVE | |
| 38 | PX SEND HASH | :TQ20000 |
| 39 | SORT GROUP BY | |
| 40 | FILTER | |
| 41 | PX PARTITION HASH ITERATOR | |
| 42 | TABLE ACCESS BY GLOBAL INDEX ROWID BATCHED | REGREPORT_EVENT_REPORT |
| 43 | INDEX RANGE SCAN | REGREPORT_EVENT_REPORT_PK |
| 44 | TABLE ACCESS BY USER ROWID | REGREPORT_EVENT_REPORT |
===========================================================================================================
Possibly this sounds a bit esoteric at beast and highly situation specific and yet without a useful standalone data setup to adequately explain the whys and wherefores. It is what it is.
The summary is if you want to join to an inline view and you think you want to push predicates, maybe do some testing around a lateral join to that inline view.
I’ve been meaning to blog about this for a while but it’s really hard these days to find interesting things to blog about which have the right balance of words vs technical content. Apart from anything else, it’s hard to get the real world evidence anonymised and out of the workplace and I just don’t have the time to model problems from scratch for blog purposes – and few people seem to read blogs these days…
So there’s some performance problem that has been reported and you start digging. Probably like me you have a few usual suspects you will immediately turn to.
But let’s start with what I definitely do NOT start with:
Lists of objects with stale statistics
Lists of indexes which some script off the internet has determined need rebuilding
AWR reports
We need EVIDENCE. Evidence of the problem. Appropriately scoped evidence of the problem. Maybe it’s far too early to look for the WHYs but certainly we’re on the lookout for WHATs. If something took some time, it was actively doing something or actively waiting on something. So let’s look for that/those something/s.
At least 8 times out of 10, we’re probably talking about a specific SQL statement eventually identifiable so hopefully you have a SQL ID (and f it’s not that well that could be the start of another set of blog posts)
Assuming that is what it is, what do you go to next?
For me, it’s probably DBA_HIST_SQLSTAT.
I want to compare other executions, if there are any. Depending on how often it executes and how long it executes for, I might aggregate this by AWR snapshot or by day.
What am I looking for ?
9 times out of 10 I expect to find a change in execution plan
Is there a large variety of very similarly performing plans?
Does a particular plan standout as particularly “bad”?
Maybe the optimizer arithmetic makes the mathematical choice quite close but in reality it makes a big difference
Maybe there’s a set of parameters which suit one way and not another
Hopefully though you can see evidence in some of the metrics which show you, not necessarily why something happened, but what was the contribution to the performance compared to previous executions.
For example, you might see that logical IO went through the roof but there can be a myriad of reasons for that, some of which might be caused by other processes
Note that the SQL_ID might not be constant – when it is it makes this easier
But maybe your code uses literals so every combination of literals is different but at least we can use FORCE_MATCHING_SIGNATURE to identify other executions
Or maybe the SQL is completely dynamic which different subqueries, so maybe we’re mining DBA_HIST_SQLTEXT for certain patterns of features of a particular piece of code
Then?
For me, it’s probably the real time sql monitoring report (RTSM) for that SQL – DBMS_SQLTUNE.REPORT_SQL_MONITOR
What might jump out there?
How long has/was it running?
Was it parallel? How parallel? Was all the parallelism it wanted to run with granted?
Which operations took the longest?
Did the optimizer get some of the estimates badly wrong?
If it did get some badly wrong, would it have made a difference? e.g.
Might it have had an a NESTED LOOP vs HASH JOIN decision
Might it have caused undersized workareas which spilled to TEMP?
If you’re familiar with the application, even the code, does the execution plan make sense to you?
Would you have done it differently? Maybe driven from a different table?
Where did the broad categories of time go?
CPU vs IO vs Application Waits for example
How much data did you read both logically and physically?
From the activity detail, if available, are there any wait events which stand out as taking all the time?
Note that if there are executions in memory, it can be very useful to compare metrics across specific individual executions by looking at GV$SQL_MONITOR
Maybe nothing is still in memory and you have to resort to historical information – see DBA_HIST_REPORTS and DBMS_AUTO_REPORT.REPORT_REPOSITORY_DETAIL?
So we start to build up a picture of contributing factors, elements which might have been cause, might have been effect, but we start joining up the dots.
As mentioned, very often it’s a simple plan change and you can fix it through your tool of choice – sql profile? Sql plan baseline? Sql patch? Change the code?
Often I find that changing the code is the best long term choice particularly if you look at it and think “well I wouldn’t have written it like that.” So often I find a badly performing piece of SQL is not structured in a natural way.
Anyway, what happens when we find that the execution plan is stable?
We can see that one or some executions definitely took a lot longer.
Perhaps it’s time to start looking at some of the data?
Another great thing about the RTSM reports is it tells you about the binds which were used for those executions.
Maybe you’re using binds and you’re comparing apples to oranges with small dates ranges vs large date ranges, etc, etc
Because most of the time, the problem is already identified using one or both of the sources above, I’m usually pretty slow at starting to dig into the actual parameters passed into the procedures.
At this point, well it depends. I might start looking at the tables involved and start slicing and dicing data to see whether there is some ridiculous data skew.
Twice recently I’ve had very notable issues caused not by an issue with the plan, not by anything happening on the database, but by an abuse of data. Normally by other systems sending data in a way they are not meant to but not necessarily in a way that we can control. Maybe not in a way that can be protected by some contraint or other. But in a way that causes a serious execution issue.
This comes from an event sourcing system, the long term repository for a variety of events being sent between systems triggering workflows, processing and reporting. And used to pass downstream to other controls and consolidation systems.
See line 25, what we’ve got here is a join to an inline view which is trying to offer some protection against “duplicate” events*.
And we can see that in those metrics.
We do an index range scan on line 29 fed by 202K input rows. But the output of those 202K lookups is 273 million rows which then are deduplicated by the SORT GROUP BY on line 26 back to 202K.
Ouch. That is where the pain is.
If we delve into the data for the bind parameters in question, we start to find examples of 64k “duplicates”. The SQL deduplication protection mechanism is expecting a handful of “duplicates”* not many thousands.
*For technical reasons, these are not duplicates. The upstream SORs have transmitted a very similar event many times in a very short space of time (due to a bug probably). The reporting systems will have generated the reports many times because of those events and so we the long term event sourcing system need to keep them. But the downstream systems try and extract additional meaning from these events and therein lies the problem for this code.
create table domtest
(col1 number
,col2 varchar2(200))
partition by range (col1) interval (1)
(partition p0 values less than (1))
compress for oltp;
> Table created.
insert into domtest
with x as (select level lvl from dual connect by level <= 10)
select lvl, rpad(lvl,200,'X') from x;
> 10 row(s) inserted.
commit;
> Statement processed.
select table_name, def_compression, def_compress_for
from user_part_tables
where table_name = 'DOMTEST';
TABLE_NAME DEF_COMPRESSION DEF_COMPRESS_FOR
DOMTEST ENABLED ADVANCED
select table_name, partition_name, high_value, compression, compress_for
from user_tab_partitions
where table_name = 'DOMTEST';
TABLE_NAME PARTITION_NAME HIGH_VALUE COMPRESSION COMPRESS_FOR
DOMTEST P0 1 ENABLED ADVANCED
DOMTEST SYS_P487584 2 ENABLED ADVANCED
DOMTEST SYS_P487585 3 ENABLED ADVANCED
DOMTEST SYS_P487586 4 ENABLED ADVANCED
DOMTEST SYS_P487587 5 ENABLED ADVANCED
DOMTEST SYS_P487588 6 ENABLED ADVANCED
DOMTEST SYS_P487589 7 ENABLED ADVANCED
DOMTEST SYS_P487590 8 ENABLED ADVANCED
DOMTEST SYS_P487591 9 ENABLED ADVANCED
DOMTEST SYS_P487592 10 ENABLED ADVANCED
DOMTEST SYS_P487593 11 ENABLED ADVANCED
If I want to change the default compression at table level so that new partitions are no longer compressed, then those three words make a big difference.
alter table domtest nocompress;
>Table altered.
This does what we want:
select table_name, def_compression, def_compress_for
from user_part_tables
where table_name = 'DOMTEST';
TABLE_NAME DEF_COMPRESSION DEF_COMPRESS_FOR
DOMTEST DISABLED -
But it does a whole lot we don’t. It doesn’t move the existing data but it does change the partition attributes for new data.
We hit a serious problem about 5 months ago when we upgraded one of the development databases from 19.12 to 19.13. Going back to 19.12 solved the problem so definitely for us something introduced in 19.13
Nothing out there at the time on Oracle Support or from other reliable sources.
Our problem happened under row cache load which in particular effected our commit builds when there was heavy dropping and creating of objects across a number of schemas.
Best indication so far is that this is related to the cloning feature of row cache objects, present in 19.13, new feature with aims to automatically create clones of row cache objects if it detects that the object is accessed frequently.
Apparently many bugs related to this feature have already been addressed but suggestion is that disabling the feature might prevent the issue, via the snappily named parameter _bug33046179_kqr_hot_copy_sleep_limit = 0.
Will follow-up with comment on effectiveness.
Addendum: Good news – parameter seems effective so far but you need to restart the database. Although you can change the parameter without restarting the database, in our experience you do need to restart if the circumstances around the bug might already have taken effect before you changed the parameter.
I’m triple posting this on oracle-l, my blog and on Oracle community to get some more help. Apologies if you’ve tried to help before and it didn’t come to anything.
When the problem originally surfaced it was after an upgrade from 11.2.0.4 to 19.6. Now version is 19.12.
Now I have one example which gets executed four or five times a day The example has two sql ids involved in the process but both are almost identical and have identical sql plans.
Queries in question get executed after specific data loads. For each data load where it gets executed, it will be doing pretty much the same workload day-on-day. The query gets supplied a collection which tends to have one or a couple of rows row and each row in the collection provides the values to prune to a specific subpartition.
The issue is that about once a week, an execution will go slow. Exactly the same sql plan each execution. When it works, it runs in a couple of seconds max. When it goes wrong, the partition list iterator decides to read every partition/subpartition.
There was some speculation about possible causes in the OTN thread. The best guess was that there is some timeout or invalidation or event or contention which means Oracle decides that it cannot reliably prune. There might be other data loads ongoing. There could be a partition operation (add new or exchange) which coincides with the degraded iterator performance. There are > 70k partitions / subpartitions.
select count(*) from dba_tab_partitions where table_name = 'POSITION_BALANCE';
COUNT(*)
----------
60941
select count(*), sum(case when segment_created = 'YES' then 1 else 0 end) from dba_tab_subpartitions where table_name = 'POSITION_BALANCE';
COUNT(*) SUM(CASEWHENSEGMENT_CREATED='YES'THEN1ELSE0END)
---------- -----------------------------------------------
78154 33705
I’m reasonably comfortable that what I’ve described is roughly what’s happening. ASH confirms that it’s reading lots of subpartitions Just got no idea why and how to figure that out. The average-ish once-a-week occurence makes it tricker.
Even though when it runs in runs in a couple of seconds, I do have some of the ASH samples that show this.
select /*+ parallel(4) */ instance_number, sql_id, sql_plan_hash_value, sql_full_plan_hash_value, sql_exec_id, sql_exec_start, min(sample_time), max(sample_time), max(sample_time) - min(sample_time) duration, count(*) cnt_samples
, sum(case when in_sql_execution = 'Y' then 1 else 0 end) cnt_in_sql_execution
, sum(case when in_parse = 'Y' then 1 else 0 end) cnt_in_parse
, sum(case when in_hard_parse = 'Y' then 1 else 0 end) cnt_in_hard_parse
from dba_hist_active_sess_history h
where 1=1
and h.user_id IN (select user_id from dba_users where username like '%TODS%')
and sql_id in ('fw985yayy1mfz','0gk5hj9wnp2vs')
--and sql_id = 'ft2prc1xx2hmd'
and sql_exec_id is not null
group by instance_number, sql_id, sql_plan_hash_value, sql_full_plan_hash_value, sql_exec_id, sql_exec_start
order by h.sql_exec_start;
INSTANCE_NUMBER SQL_ID SQL_PLAN_HASH_VALUE SQL_FULL_PLAN_HASH_VALUE SQL_EXEC_ID SQL_EXEC_START MIN(SAMPLE_TIME) MAX(SAMPLE_TIME) DURATION CNT_SAMPLES CNT_IN_SQL_EXECUTION CNT_IN_PARSE CNT_IN_HARD_PARSE
--------------- ------------- ------------------- ------------------------ ----------- -------------------- ------------------------------- ------------------------------- ------------------- ----------- -------------------- ------------ -----------------
4 0gk5hj9wnp2vs 2221895109 2867134094 67108864 23-MAY-2022 09:18:24 23-MAY-22 09.18.27.378000000 AM 23-MAY-22 09.18.27.378000000 AM +00 00:00:00.000000 1 1 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331648 24-MAY-2022 07:59:49 24-MAY-22 07.59.50.659000000 AM 24-MAY-22 07.59.50.659000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554432 25-MAY-2022 07:09:07 25-MAY-22 07.09.08.709000000 AM 25-MAY-22 07.09.08.709000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554434 25-MAY-2022 09:41:53 25-MAY-22 09.41.59.718000000 AM 25-MAY-22 12.41.49.605000000 PM +00 02:59:49.887000 1054 1054 0 0
2 fw985yayy1mfz 2221895109 2867134094 33554434 25-MAY-2022 10:57:41 25-MAY-22 10.57.49.221000000 AM 25-MAY-22 01.56.37.861000000 PM +00 02:58:48.640000 1048 1048 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331648 25-MAY-2022 11:41:04 25-MAY-22 11.41.05.539000000 AM 25-MAY-22 11.41.05.539000000 AM +00 00:00:00.000000 1 1 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331648 27-MAY-2022 08:01:28 27-MAY-22 08.01.30.371000000 AM 27-MAY-22 08.01.30.371000000 AM +00 00:00:00.000000 1 1 0 0
3 fw985yayy1mfz 2221895109 2867134094 50331648 27-MAY-2022 08:01:38 27-MAY-22 08.01.40.611000000 AM 27-MAY-22 08.01.40.611000000 AM +00 00:00:00.000000 1 1 0 0
1 0gk5hj9wnp2vs 2221895109 2867134094 16777218 01-JUN-2022 06:48:24 01-JUN-22 06.48.27.979000000 AM 01-JUN-22 09.48.17.547000000 AM +00 02:59:49.568000 1054 1054 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331649 01-JUN-2022 09:27:03 01-JUN-22 09.27.10.915000000 AM 01-JUN-22 09.27.10.915000000 AM +00 00:00:00.000000 1 1 0 0
3 fw985yayy1mfz 2221895109 2867134094 50331648 02-JUN-2022 07:11:55 02-JUN-22 07.11.57.315000000 AM 02-JUN-22 07.11.57.315000000 AM +00 00:00:00.000000 1 1 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331648 07-JUN-2022 07:04:13 07-JUN-22 07.04.17.155000000 AM 07-JUN-22 07.04.17.155000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554434 07-JUN-2022 08:17:56 07-JUN-22 08.17.58.693000000 AM 07-JUN-22 11.17.49.285000000 AM +00 02:59:50.592000 1054 1054 0 0
1 0gk5hj9wnp2vs 2221895109 2867134094 16777216 08-JUN-2022 07:09:22 08-JUN-22 07.09.24.427000000 AM 08-JUN-22 07.09.24.427000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554432 08-JUN-2022 08:14:25 08-JUN-22 08.14.26.278000000 AM 08-JUN-22 08.14.26.278000000 AM +00 00:00:00.000000 1 1 0 0
3 fw985yayy1mfz 2221895109 2867134094 50331649 09-JUN-2022 06:51:27 09-JUN-22 06.51.29.219000000 AM 09-JUN-22 06.51.29.219000000 AM +00 00:00:00.000000 1 1 0 0
3 fw985yayy1mfz 2221895109 2867134094 50331648 10-JUN-2022 11:10:13 10-JUN-22 11.10.14.595000000 AM 10-JUN-22 11.10.14.595000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554433 13-JUN-2022 08:45:43 13-JUN-22 08.45.45.509000000 AM 13-JUN-22 08.45.45.509000000 AM +00 00:00:00.000000 1 1 0 0
2 fw985yayy1mfz 2221895109 2867134094 33554433 13-JUN-2022 08:45:55 13-JUN-22 08.45.55.749000000 AM 13-JUN-22 08.45.55.749000000 AM +00 00:00:00.000000 1 1 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331648 15-JUN-2022 07:08:19 15-JUN-22 07.08.22.275000000 AM 15-JUN-22 07.08.22.275000000 AM +00 00:00:00.000000 1 1 0 0
3 0gk5hj9wnp2vs 2221895109 2867134094 50331649 15-JUN-2022 09:08:54 15-JUN-22 09.08.56.387000000 AM 15-JUN-22 12.08.46.083000000 PM +00 02:59:49.696000 1054 1054 0 0
1 fw985yayy1mfz 2221895109 2867134094 16777216 16-JUN-2022 06:54:04 16-JUN-22 06.54.05.259000000 AM 16-JUN-22 06.54.05.259000000 AM +00 00:00:00.000000 1 1 0 0
1 0gk5hj9wnp2vs 2221895109 2867134094 16777217 16-JUN-2022 08:54:13 16-JUN-22 08.54.18.891000000 AM 16-JUN-22 11.54.08.844000000 AM +00 02:59:49.953000 1054 1054 0 0
3 fw985yayy1mfz 2221895109 2867134094 50331649 16-JUN-2022 11:16:29 16-JUN-22 11.16.31.491000000 AM 16-JUN-22 11.16.31.491000000 AM +00 00:00:00.000000 1 1 0 0
4 0gk5hj9wnp2vs 2221895109 2867134094 67108865 17-JUN-2022 07:33:10 17-JUN-22 07.33.11.282000000 AM 17-JUN-22 07.33.11.282000000 AM +00 00:00:00.000000 1 1 0 0
4 fw985yayy1mfz 2221895109 2867134094 67108865 17-JUN-2022 07:33:20 17-JUN-22 07.33.21.522000000 AM 17-JUN-22 07.33.21.522000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554432 21-JUN-2022 06:59:24 21-JUN-22 06.59.25.728000000 AM 21-JUN-22 06.59.25.728000000 AM +00 00:00:00.000000 1 1 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554433 21-JUN-2022 09:14:05 21-JUN-22 09.14.10.464000000 AM 21-JUN-22 12.14.00.096000000 PM +00 02:59:49.632000 1054 1054 0 0
2 0gk5hj9wnp2vs 2221895109 2867134094 33554434 21-JUN-2022 11:07:36 21-JUN-22 11.07.44.480000000 AM 21-JUN-22 02.07.23.680000000 PM +00 02:59:39.200000 1053 1053 0 0
4 fw985yayy1mfz 2221895109 2867134094 67108865 21-JUN-2022 11:43:04 21-JUN-22 11.43.13.010000000 AM 21-JUN-22 02.42.33.651000000 PM +00 02:59:20.641000 1051 1051 0 0
4 fw985yayy1mfz 2221895109 2867134094 67108866 21-JUN-2022 12:06:49 21-JUN-22 12.06.57.586000000 PM 21-JUN-22 03.06.38.514000000 PM +00 02:59:40.928000 1053 1053 0 0
3 fw985yayy1mfz 2221895109 2867134094 50331649 23-JUN-2022 10:05:02 23-JUN-22 10.05.05.667000000 AM 23-JUN-22 01.04.45.251000000 PM +00 02:59:39.584000 1053 1053 0 0
1 0gk5hj9wnp2vs 2221895109 2867134094 16777216 23-JUN-2022 10:57:04 23-JUN-22 10.57.04.857000000 AM 23-JUN-22 10.57.04.857000000 AM +00 00:00:00.000000 1 1 0 0
This is a section of a RTSM report when it was ok:
Potentially I could set system level trace for this sql id and wait for it to reoccur but what events would I be setting to try to capture what causes this runtime decision? I presume it wouldn’t be an optimizer trace?
I have a couple of performance issues at the moment which link back to queries against large tables choosing to do a FULL segment scan rather than using a non-prefixed local index.
The problems right now are not currently reproducible in or outside of production.
But when I look at a specific point in time at the reproduction of a problem, the system state is not necessarily as it was when the problem was occurring.
My main theory is that somewhere on these large partitioned tables there is intermittently an unusable partition of the index perhaps due some aspect of a data load, data movement or other maintenance operation on a different subpartition (where each partition represents a different day of a specific feed and each subpartition represents a version of that daily load).
However I have no evidence currently to support that theory.
If this theory was in the right ballpark, what would it mean?
That a query using bind variables (tick) would not use the index by default as the optimizer would not be able to guarantee that that shareable plan would not be accessing the unusable.
A query using literals would not be subject to the same restrictions as the optimizer would know, if an unusable subpartition was relevant, that the unusable subpartition was not relevant to this specific query (and any DDL which might make a relevant partition unusable would invalidate the shared cursor).
Whilst I was musing over this yesterday, I had a vague sense of familiarity and I found these previous posts to job my memory:
create table t1
(pkey varchar2(24) not null
,spkey number not null
,id varchar2(24) not null
,version number not null
,status varchar2(24) not null
,filler varchar2(255))
partition by list (pkey)
subpartition by list (spkey)
subpartition template
(subpartition sp_1 values (1))
(partition p_0 values ('X'));
alter table t1 add partition p_abc values ('ABC');
alter table t1 add partition p_def values ('DEF');
alter table t1 add partition p_ghi values ('GHI');
create index i1 on t1 (id) local;
create sequence s1 start with 10000000;
create type o1 is object
(pkey varchar2(24)
,spkey number
,id varchar2(24));
/
create type c1 is table of o1;
/
select object_name, subobject_name, object_type from user_objects where object_name = 'T1' order by object_name, subobject_name nulls first;
OBJECT_NAM SUBOBJECT_ OBJECT_TYPE
---------- ---------- -----------------------
T1 TABLE
T1 P_0 TABLE PARTITION
T1 P_0_SP_1 TABLE SUBPARTITION
T1 P_ABC TABLE PARTITION
T1 P_ABC_SP_1 TABLE SUBPARTITION
T1 P_DEF TABLE PARTITION
T1 P_DEF_SP_1 TABLE SUBPARTITION
T1 P_GHI TABLE PARTITION
T1 P_GHI_SP_1 TABLE SUBPARTITION
select object_name, subobject_name, object_type from user_objects where object_name = 'I1' order by object_name, subobject_name nulls first;
OBJECT_NAM SUBOBJECT_ OBJECT_TYPE
---------- ---------- -----------------------
I1 INDEX
I1 P_0 INDEX PARTITION
I1 P_0_SP_1 INDEX SUBPARTITION
I1 P_ABC INDEX PARTITION
I1 P_ABC_SP_1 INDEX SUBPARTITION
I1 P_DEF INDEX PARTITION
I1 P_DEF_SP_1 INDEX SUBPARTITION
I1 P_GHI INDEX PARTITION
I1 P_GHI_SP_1 INDEX SUBPARTITION
And I’m going to seed some data simply:
declare
p sys.odcivarchar2list := sys.odcivarchar2list('ABC','DEF','GHI');
begin
for i in 1 .. p.count
loop
for j in 1 .. 10000
loop
insert into t1
values (p(i), 1, 'PB:'||s1.nextval, 1, 'LATEST', rpad('X',255,'X'));
end loop;
end loop;
commit;
end;
/
select pkey, spkey, count(*), min(id), max(id) from t1 group by pkey, spkey;
PKEY SPKEY COUNT(*) MIN(ID) MAX(ID)
---- ---------- ---------- ------------------------ ------------------------
ABC 1 10000 PB:10000000 PB:10009999
DEF 1 10000 PB:10010000 PB:10019999
GHI 1 10000 PB:10020000 PB:10029999
And then just to keep it representative to my real world problem, I’m going to run a bulk update to set the status of some of the rows to SUPERSEDED (and which is then in the real world followed by the INSERT of some LATEST versions of those rows supplied by the client but no need to do that here):
declare
v1 c1 := c1(o1('DEF',1,'PB:10010001'),
o1('DEF',1,'PB:10010002'),
o1('DEF',1,'PB:10010003'),
o1('DEF',1,'PB:10010004'),
o1('DEF',1,'PB:10010005'));
begin
forall i in 1 .. v1.count
update /*+ find_me_dom */
t1
set status = 'SUPERSEDED'
where pkey = v1(i).pkey
and spkey = v1(i).spkey
and id = v1(i).id;
end;
/
And I’m going to lookup my sql id from v$sql and plug it into DBMS_XPLAN to see how my UPDATE performed:
select * from table(dbms_xplan.display_cursor('bjddz8c4jrk3y',0));
-------------------------------------------------------------
| Id | Operation | Name |
-------------------------------------------------------------
| 0 | UPDATE STATEMENT | |
| 1 | UPDATE | T1 |
| 2 | PARTITION LIST SINGLE | |
| 3 | PARTITION LIST SINGLE | |
| 4 | TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| T1 |
| 5 | INDEX RANGE SCAN | I1 |
-------------------------------------------------------------
Range Scan of I1 as desired
Now I will mark one subpartition of the index unusable (different subpartition from one affected by the update) and repeat the update, the getting of the SQL and the fetching of the plan (different child number):
alter index i1 modify subpartition P_ABC_SP_1 unusable;
--Repeat dml
--lookup sql
select * from table(dbms_xplan.display_cursor('bjddz8c4jrk3y',1));
-- ^ different child cursor number because marking the index subpartition as unusable
-- has invalidated the previous child cursor 0
---------------------------------------------
| Id | Operation | Name |
---------------------------------------------
| 0 | UPDATE STATEMENT | |
| 1 | UPDATE | T1 |
| 2 | PARTITION LIST SINGLE | |
| 3 | PARTITION LIST SINGLE | |
| 4 | TABLE ACCESS STORAGE FULL| T1 |
---------------------------------------------
The access method is no longer an index range scan.
As I covered initially, as part of the parse process the execution plan has to be good for all possible ranges of supplied values in the event that that SQL child cursor is shared for executions with different binds than those initially parsed with. Because an update (or select etc, etc) against rows for table subpartition (ABC,1) cannot use the index because that index subpartition is unusable, the optimizer has to discount it.
We can force the index:
declare
v1 c1 := c1(o1('DEF',1,'PB:10010001'),
o1('DEF',1,'PB:10010002'),
o1('DEF',1,'PB:10010003'),
o1('DEF',1,'PB:10010004'),
o1('DEF',1,'PB:10010005'));
begin
forall i in 1 .. v1.count
update /*+ find_me_dom index(t1 (id))*/
t1
set status = 'SUPERSEDED'
where pkey = v1(i).pkey
and spkey = v1(i).spkey
and id = v1(i).id;
end;
/
select * from table(dbms_xplan.display_cursor('3m2xyxgruxkpr',0));
-------------------------------------------------------------
| Id | Operation | Name |
-------------------------------------------------------------
| 0 | UPDATE STATEMENT | |
| 1 | UPDATE | T1 |
| 2 | PARTITION LIST SINGLE | |
| 3 | PARTITION LIST SINGLE | |
| 4 | TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| T1 |
| 5 | INDEX RANGE SCAN | I1 |
-------------------------------------------------------------
But note, as covered in my earlier link, if we force the index in a situation where it can’t be used, it will error in current versions:
alter index i1 modify subpartition P_DEF_SP_1 unusable;
declare
v1 c1 := c1(o1('DEF',1,'PB:10010001'),
o1('DEF',1,'PB:10010002'),
o1('DEF',1,'PB:10010003'),
o1('DEF',1,'PB:10010004'),
o1('DEF',1,'PB:10010005'));
begin
forall i in 1 .. v1.count
update /*+ find_me_dom index(t1 (id))*/
t1
set status = 'SUPERSEDED'
where pkey = v1(i).pkey
and spkey = v1(i).spkey
and id = v1(i).id;
end;
/
ORA-01502: index 'DOM.I1' or partition of such index is in unusable state
ORA-06512: at line 8
01502. 00000 - "index '%s.%s' or partition of such index is in unusable state"
*Cause: An attempt has been made to access an index or index partition
that has been marked unusable by a direct load or by a DDL
operation
*Action: DROP the specified index, or REBUILD the specified index, or
REBUILD the unusable index partition
If I rebuild that second partition again, then I can look at my second option – using literals not binds, at least for the partition keys, but really we then lose the option of FORALL.
alter index i1 rebuild subpartition P_DEF_SP_1;
declare
v1 c1 := c1(o1('DEF',1,'PB:10010001'),
o1('DEF',1,'PB:10010002'),
o1('DEF',1,'PB:10010003'),
o1('DEF',1,'PB:10010004'),
o1('DEF',1,'PB:10010005'));
begin
for i in 1 .. v1.count
loop
execute immediate
'update /*+ find_me_dom */
t1
set status = ''SUPERSEDED''
where pkey = '''||v1(i).pkey||''''||'
and spkey = '||v1(i).spkey||'
and id = :1' using v1(i).id;
end loop;
end;
/
select * from table(dbms_xplan.display_cursor('5a3vcac58x32q',0));
----------------------------------------
| Id | Operation | Name |
----------------------------------------
| 0 | UPDATE STATEMENT | |
| 1 | UPDATE | T1 |
| 2 | PARTITION LIST SINGLE | |
| 3 | PARTITION LIST SINGLE| |
| 4 | INDEX RANGE SCAN | I1 |
----------------------------------------
At least we have options whilst playing a waiting game to see if we can observe a problem state which might cause such an issue…. or wait for other possibilities to make themselves known…
I’m currently an Oracle Ace (for a few more weeks at least!), I love writing code, digging out significant performance gains in existing processes and troubleshooting problems. Simples.
If you’re interested, please drop me a mail, add a comment, send me a linkedin message, anything.
This is the distillation of a performance problem in some regulatory reporting where a mistake in the optimizer cardinality estimates causes all sorts of knock-on performance issues. This post does not look at those knock-on effects but examines what is happening with the optimizer in this case. Most of the digging in the issue belongs to Valerii Satybaldyev and thanks to Sayan Malakshinov, notably for the contributions on the oracle-l thread.
Script to reproduce (currently on 19.6):
drop table t1;
create table t1
(id number not null
,version number not null
,create_ts timestamp not null
,modify_ts timestamp
,status varchar2(24) generated always as (NVL2("MODIFY_TS",'SUPERSEDED','LATEST'))
,id2 number not null
,yn varchar2(1) not null
,business_date date not null);
insert into t1
(id, version, create_ts, id2, yn, business_date)
select rownum
, 1
, systimestamp
, rownum
, case when mod(rownum,2) = 1 then 'Y' else 'N' end
, trunc(sysdate,'MON') + mod(rownum,10)
from dual
connect by level <= 1000;
exec dbms_stats.gather_table_stats(USER,'T1');
explain plan for
with x as
(select * from t1
union all
select * from t1)
select *
from x
where yn = 'Y';
select * from table(dbms_xplan.display);
The problem comes with the estimate of 1 row in the outermost SELECT, particularly for subsequent joins this cardinality estimate causes significant performance problems.
If we look at the 10053 trace, we can see the root cause of the problem. Compare the BASE STATISTICAL INFORMATION snippets for T1 and for the inline view X. Note in particular the numbers for column YN which is the predicate as well as for ID2 and BUSINESS_DATE. Note that these are columns which appear after the virtual column STATUS. (in the original trace file, the entry under T1 for column #8 comes before column #7, I have reordered below just to make the correlation between T1 and X more obvious)
For inline view X, column STATUS inherits the stats for ID2, ID2 gets those of YN, YN those of BUSINESS_DATE and BUSINESS_DATE gets NO STATISTICS. So for our predicate on YN we get completely the wrong base statistics. Then in addition, because the supplied value is out of range from the wrong statistics, estimates get further pro-rated.
Potential workaround is not risk-free as it moves the implicit order of the selected columns which could affect certain code.
alter table t1 modify status invisible;
alter table t1 modify status visible;
explain plan for
with x as
(select * from t1
union all
select * from t1)
select *
from x
where yn = 'Y';
select * from table(dbms_xplan.display);
Plan hash value: 3505968351
------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1000 | 90000 | 548 (0)| 00:00:01 |
| 1 | VIEW | | 1000 | 90000 | 548 (0)| 00:00:01 |
| 2 | UNION-ALL | | | | | |
|* 3 | TABLE ACCESS STORAGE FULL| T1 | 500 | 20000 | 274 (0)| 00:00:01 |
|* 4 | TABLE ACCESS STORAGE FULL| T1 | 500 | 20000 | 274 (0)| 00:00:01 |
------------------------------------------------------------------------------------
Reason it works is that the virtual column is now logically at the end of the column definitions.
If we compare the trace once “fixed”, then we see the following in the X section:
Table: X Alias: X (NOT ANALYZED)
...
Column (#8): STATUS(VARCHAR2) NO STATISTICS (using defaults)
The views expressed are my own and not those of my employer or my clients. The views and opinions expressed by visitors to this blog are theirs and do not necessarily reflect mine.
OraStory 
Recent Comments