Hi Martin,
that would not work. I am working for a service provider. The data is not ours.
We are by policy forbitten to run queries on prod other than against the
dictionary.
We must look for an other way. Maybe some instant clone would work.
Thanks
Lothar
----Ursprüngliche Nachricht----
Von : martin.a.berger@xxxxxxxxx
Datum : 15/06/2018 - 21:04 (CEST)
An : l.flatz@xxxxxxxxxx
Cc : jonathan@xxxxxxxxxxxxxxxxxx, oracle-l@xxxxxxxxxxxxx
Betreff : Re: Re: Explain Plan and Security
If I followed this thread right, there is nothing you can do than execute the
query. Everything else will generate different results.
Is there any chance you get permission to execute the query if you can
guarantee it only runs for "a very short time"?
E.g. a special (proxy) user with a very strict LOGICAL_READS_PER_SESSION comes
to my mind.
Or you add an additional filter with "where 1 =impossible_function" and your
"impossible_function" does an execute immediate "select 1/0 from dual".
More methods come to my mind, but I'm sure you get the idea.
The execution-trap can be tested in non-profit environment and so you might
convince your customer?
hth,
berx
l.flatz@xxxxxxxxxx <l.flatz@xxxxxxxxxx> schrieb am Fr., 15. Juni 2018, 10:20:
Thanks to all for discussing.
Well, actually my case is that we have a sql monitor of a query that went
wrong. By analyzing the sql monitor result we have a fair idea what kind of
plan we want.
From the monitor we also have the exact bind variable values and can generate a
run script.
We can then take some action to fix the issue, like recalculation stats,
rewriting the query somewhat.
Of course we want to check if our fix works.
We are not allowed to run the query in production. (The Therefore the next best
thing would be good execution plan.)
I tried explain plan, but the bind variables matter and so far I never got a
good prediction.
It seems to me the whole matter is a bit more complex than I originally
thought. Therefore I really wan´t to ping the usual suspects.
Regards
Lothar
----Ursprüngliche Nachricht----
Von : jonathan@xxxxxxxxxxxxxxxxxx
Datum : 14/06/2018 - 19:56 (GMT)
An : oracle-l@xxxxxxxxxxxxx
Betreff : Re: Explain Plan and Security
Dominic,
There's no question that if the query has executed and you can get there in
time then the plan you get from display_cursor() is the plan that actually
happened, but we're discussing the point that we can get execution plans into
memory (for display_cursor()) to report) that have never executed - which leafs
to the point that those are plans that might never actually happen with any
real user inputs.
Regards
Jonathan Lewis
________________________________________
From: Dominic Brooks <dombrooks@xxxxxxxxxxx>
Sent: 14 June 2018 18:52
To: Jonathan Lewis
Cc: oracle-l@xxxxxxxxxxxxx
Subject: Re: Explain Plan and Security
Well ... dbms_xplan.display_cursor gives you definitively the execution plan
you just got for your SQL/child. It might not be the plan you get every
execution under all circumstances but you can’t take away that you got that
once. You can’t say the same about explain plan.
If someone is executing a piece of sql, say they are testing / making a change,
and they want to document the execution plan that they got during their test
and show that, for one execution at least the plan and performance was ok, then
dbms_xplan.display_cursor (or getting the same info direct from v$sql_plan) is
the source for that. That’s what I expect developers to provide plus the
runtime execution stats.
Ditto for extracting the plan information for any particular child cursor that
is in memory, display_cursor tells you what it is/was. No doubts.
Sent from my iPhone
On 14 Jun 2018, at 18:40, Jonathan Lewis <jonathan@xxxxxxxxxxxxxxxxxx> wrote:--
Andy,
I don't think I'd even be that generous. I can't think of any detail where
dbms_sql/dbms_xplan.display_cursor() gives you safer information than explain
plan / dbms_xplan.display().
I suppose the extra complexity of using dbms_sql might make you a little more
careful as you set up a test, and that could be a benefit.
Regards
Jonathan Lewis
________________________________________
From: andyklock@xxxxxxxxx <andyklock@xxxxxxxxx> on behalf of Andy Klock
<andy@xxxxxxxxxxxxxxx>
Sent: 14 June 2018 17:15
To: Jonathan Lewis
Cc: oracle-l@xxxxxxxxxxxxx
Subject: Re: Explain Plan and Security
Ah! Indeed. Thanks for that Jonathan. So, the takeaway is that DBMS_SQL is
slightly better than EXPLAIN PLAN. But, only slightly?
Thanks!
Andy K
On Thu, Jun 14, 2018 at 12:08 PM, Jonathan Lewis
<jonathan@xxxxxxxxxxxxxxxxxx<mailto:jonathan@xxxxxxxxxxxxxxxxxx>> wrote:
Andy,
If you add a "describe columns" to your SQL PARSE TEST BIND you find that you
do get a plan before the execute.
DECLARE
cursor_name INTEGER;
rows_processed INTEGER;
m_desc_table dbms_sql.desc_tab;
m_colcount number;
BEGIN
cursor_name := dbms_sql.open_cursor;
dbms_sql.parse(cursor_name, 'select /* SQL_PARSE TEST BIND */ * from
parse_test where n = :n', dbms_sql.NATIVE);
dbms_sql.bind_variable(cursor_name, ':n', 2);
dbms_sql.describe_columns( cursor_name, m_colcount, m_desc_table );
END;
/
SQL> select sql_id, plan_hash_value, sql_text from V$sql where sql_text like
'%SQL_PARSE TEST BINDwq%'
2 /
SQL_ID PLAN_HASH_VALUE
------------- ---------------
SQL_TEXT
------------------------------------------------------------------------------------------------------------------------------------
911jt1m3dxrba 903671040
select sql_id, plan_hash_value, sql_text from V$sql where sql_text like
'%SQL_PARSE TEST BINDwq%'
Regards
Jonathan Lewis
________________________________________
From: andyklock@xxxxxxxxx<mailto:andyklock@xxxxxxxxx>
<andyklock@xxxxxxxxx<mailto:andyklock@xxxxxxxxx>> on behalf of Andy Klock
<andy@xxxxxxxxxxxxxxx<mailto:andy@xxxxxxxxxxxxxxx>>
Sent: 14 June 2018 16:56
To: mauro.pagano@xxxxxxxxx<mailto:mauro.pagano@xxxxxxxxx>
Cc: Jonathan Lewis; oracle-l@xxxxxxxxxxxxx<mailto:oracle-l@xxxxxxxxxxxxx>;
l.flatz@xxxxxxxxxx<mailto:l.flatz@xxxxxxxxxx>
Subject: Re: Explain Plan and Security
Agreed. It's kind of a cool idea, however, BINDs are checked after the
DBMS_SQL.PARSE call and is only evaluated after the call to DBMS_SQL.EXECUTE
https://nam01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdocs.oracle.com%2Fdatabase%2F121%2FARPLS%2Fd_sql.htm%23i996870&data=02%7C01%7C%7C7f3f2e59790c49647afd08d5d21ddcb7%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636645948067016044&sdata=4kyFobAwG5FlvMc2tloVVZpwaeRzJ%2BpMu7JvwWzxJqo%3D&reserved=0
I ran a quick test to see what shows up in the cursor cache after setting
bind_variable, but not calling EXECUTE and as expected you don't get a plan
at all.
SQL> DECLARE
cursor_name INTEGER;
rows_processed INTEGER;
BEGIN
cursor_name := dbms_sql.open_cursor;
dbms_sql.parse(cursor_name, 'select /* SQL_PARSE TEST */ * from parse_test
where n = 1', dbms_sql.NATIVE);
END;
/
PL/SQL procedure successfully completed.
SQL> DECLARE
cursor_name INTEGER;
rows_processed INTEGER;
BEGIN
cursor_name := dbms_sql.open_cursor;
dbms_sql.parse(cursor_name, 'select /* SQL_PARSE TEST BIND */ * from
parse_test where n = :n', dbms_sql.NATIVE);
dbms_sql.bind_variable(cursor_name, ':n', 2);
END;
/
PL/SQL procedure successfully completed.
SQL> DECLARE
cursor_name INTEGER;
rows_processed INTEGER;
BEGIN
cursor_name := dbms_sql.open_cursor;
dbms_sql.parse(cursor_name, 'select /* SQL_PARSE TEST BIND 2 */ * from
parse_test where n = :n', dbms_sql.NATIVE);
dbms_sql.bind_variable(cursor_name, ':n', 2);
rows_processed := dbms_sql.execute(cursor_name);
dbms_sql.close_cursor(cursor_name);
END;
/
PL/SQL procedure successfully completed.
SQL> @findsq SQL_PARSE
SQL_ID PLAN_HASH_VALUE LENGTH(SQL_FULLTEXT) SUB_TEXT EXECUTIONS
AVG_ELAPSED
------------- --------------- --------------------
-------------------------------------------------- ---------- -----------
4wbhzrjq0k7fd 464636435 57 select /* SQL_PARSE TEST */ * from parse_test
wher 0 .002671
8xbgz1hbjk0vz 0 63 select /* SQL_PARSE TEST BIND */ * from parse_test
0 .000467
0bavj3vaszvw2 464636435 65 select /* SQL_PARSE TEST BIND 2 */ * from
parse_te 1 .00503
Final note, if you don't actually execute the SQL then you don't get all that
other Oracle runtime stuff like cardinality feedback or dynamic sampling, etc
which adds to even more headaches.
Andy K
On Thu, Jun 14, 2018 at 11:45 AM, Mauro Pagano
<mauro.pagano@xxxxxxxxx<mailto:mauro.pagano@xxxxxxxxx><mailto:mauro.pagano@xxxxxxxxx<mailto:mauro.pagano@xxxxxxxxx>>>
wrote:
Lothar,
To add on Jonathan's "odd note", because of 9630032 (disabled by default) you
might see an even odder behavior (difference between describe vs exec).
Just saying that DBMS_SQL might translate in some headaches :-(
Cheers,
Mauro
On Thu, Jun 14, 2018 at 7:13 AM, Jonathan Lewis
<jonathan@xxxxxxxxxxxxxxxxxx<mailto:jonathan@xxxxxxxxxxxxxxxxxx><mailto:jonathan@xxxxxxxxxxxxxxxxxx<mailto:jonathan@xxxxxxxxxxxxxxxxxx>>>
wrote:
As far as I know explain plan will produce a misleading plan only if:
a) the query uses bind variables - which can't be peeked and are assumed to
be character
or
b) the literals used in the explain plan are a bad choice compared to what
happens in production
(which includes wrong type, wrong character set, wrong implicit date format
etc.)
Using dbms_sql won't (necessarily) be any better. If you supply a statement
with a bind variable in the text the call to dbms_parse will assume that it's
an unknown varchar - just as explain plan will. This is why you sometimes
see systems with lots of statements parsed twice per execute - the first time
was a parse call the that used guesses for bind types, the second was with
information about the actual bind types.
(I have an odd note from 16 years ago that you don't get the plan on the call
to dbms_parse, but have to call dbms_describe_colums as well).
Regards
Jonathan Lewis
________________________________________
From:
oracle-l-bounce@xxxxxxxxxxxxx<mailto:oracle-l-bounce@xxxxxxxxxxxxx><mailto:oracle-l-bounce@xxxxxxxxxxxxx<mailto:oracle-l-bounce@xxxxxxxxxxxxx>>
<oracle-l-bounce@xxxxxxxxxxxxx<mailto:oracle-l-bounce@xxxxxxxxxxxxx><mailto:oracle-l-bounce@xxxxxxxxxxxxx<mailto:oracle-l-bounce@xxxxxxxxxxxxx>>>
on behalf of
l.flatz@xxxxxxxxxx<mailto:l.flatz@xxxxxxxxxx><mailto:l.flatz@xxxxxxxxxx<mailto:l.flatz@xxxxxxxxxx>>
<l.flatz@xxxxxxxxxx<mailto:l.flatz@xxxxxxxxxx><mailto:l.flatz@xxxxxxxxxx<mailto:l.flatz@xxxxxxxxxx>>>
Sent: 14 June 2018 13:36:46
To:
oracle-l@xxxxxxxxxxxxx<mailto:oracle-l@xxxxxxxxxxxxx><mailto:oracle-l@xxxxxxxxxxxxx<mailto:oracle-l@xxxxxxxxxxxxx>>
Subject: Explain Plan and Security
Hi,
you might know Kerry´s classic blog:
https://nam01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fkerryosborne.oracle-guy.com%2F2008%2F10%2Fexplain-plan-lies%2F&data=02%7C01%7C%7C7f3f2e59790c49647afd08d5d21ddcb7%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636645948067016044&sdata=0MfIBWr%2FbqDjJeSA9AMeySqqeQnBo5odXPC8wxYqthY%3D&reserved=0.
Normally my work around for explain plan issues is to run the query and use
dbms_xplan.display_cursor.
Now I am working in an environment where I must not run a query, but I can do
explain plan.
But still I think I can not tolerate explain plan weaknesses.
I think it should be possble to use DBMS_SQL to parse a statement and receive
a proper plan without actually running the statement.
Then use dbms_xplan.display_cursor.
Before I spent time, has anybody done it already?
Regards
Lothar
--
https://nam01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.freelists.org%2Fwebpage%2Foracle-l&data=02%7C01%7C%7C7f3f2e59790c49647afd08d5d21ddcb7%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636645948067016044&sdata=KWBDk34WKqsWc5xaLQTbAk12iDSDkjI56zoLvoYEppM%3D&reserved=0
--
https://nam01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.freelists.org%2Fwebpage%2Foracle-l&data=02%7C01%7C%7C7f3f2e59790c49647afd08d5d21ddcb7%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636645948067016044&sdata=KWBDk34WKqsWc5xaLQTbAk12iDSDkjI56zoLvoYEppM%3D&reserved=0
--
https://nam01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.freelists.org%2Fwebpage%2Foracle-l&data=02%7C01%7C%7C7f3f2e59790c49647afd08d5d21ddcb7%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636645948067016044&sdata=KWBDk34WKqsWc5xaLQTbAk12iDSDkjI56zoLvoYEppM%3D&reserved=0