cursor_sharing='SIMILAR'将被废弃

A number of customers have seen an increase in the number of child cursors since migrating to Oracle Database 11g Release 2. This can lead to many problems including complete CPU saturation of a machine requiring a database instance bounce or general database performance issues in the form of waits on mutexes and ‘library cache lock’.
From Oracle versions 9.0 through 11.1, an oracle instance would limit the number child cursors in the shared pool associated with a single parent to 1024 before it would mark the parent OBSOLETE effectively invalidating it and all the children. Then a new parent with one child would be created and used going forward. But this would only limit the degradation of performance for some environments rather than fix something that could be addressed more effectively through improved application coding. (The attempt to address this from the database side also introduced other issues like bug 5177766). The child limit was removed by Oracle development because it was only masking an application problem at the expense of database performance for better designed applications. In addition, the obsolete code would not work in cases when SQL was wrapped within PL/SQL. The fundamental problem that obsolete code was masking is application code that was written incorrectly with regards to the ability to be shared. For example,  it is not written with user binds or the literal characteristics differ to a high degree.
Therefore, setting cursor_sharing = SIMILAR is highly discouraged in Oracle Database 11g Release 2 (and generally has not been recommended for most environments even in earlier versions) for several reasons:
1) This parameter is generally overly restrictive in what it actually allows to be shared. SIMILAR tells oracle to try and share cursors by replacing all literals with binds for legacy applications, but directs that sharing only be performed when all the replaced literal values were exactly the same (in the case of predicates referencing columns with histograms or using inequality operators such as BETWEEN, <, and !=)
2) This parameter seems to bypass a lot of the improvements made with Oracle Database 11g’s Adaptive Cursor Sharing feature and other abilities in the Cost Based Optimizer code to make better decisions on what execution plans should and should not be shared.
3) Having many child cursors all associated with 1 parent cursor could perform much worse than having many parent cursors that would be seen with having the default setting of cursor_sharing = EXACT (or FORCE). The scenario of many thousands of child cursors associated with 1 parent results in a potential bottleneck for searches for matching cursors (soft parsing) within the library cache.
The cursor_sharing parameter was introduced as a workaround for legacy applications that could not scale because they had not yet been redesigned to use bind variables. It has been presumed that most applications have been redesigned since then. If you are still using such an application, our recommendation is to set cursor_sharing = FORCE. This setting maximizes cursor sharing while leveraging the Adaptive Cursor Sharing framework to generate multiple execution plans based on different literal value ranges if necessary.

What do you need to do?

Change the cursor_sharing to either FORCE or EXACT, keeping in mind the effects of either.

this is due to the setting in the init.ora:
cursor_sharing=SIMILAR

Please set immediately
alter system set cursor_sharing=’FORCE’ scope=both;

to reduce the number of versions for a single sql statement whic is overloading your shared_pool.

377847.1 for SIMILA not FORCE are not the same

When bind variables are peeked. The parse engine makes a decision as to the ‘safety’ of these peeked values for creating plans based upon whether it is felt that different values could produce different plans.

The usual (but not the only) reason for such different plans is the use of CURSOR_SHARING=SIMILAR and the presence of histogram column statistics on the column that the bind is being compared with when using the Cost Based Optimizer (CBO). If there are histograms on the column, then the bind value may be deemed to be ‘unsafe’ because there is the potential chance that the different values could produce a different explain plan and the selection of a single plan for all values may not be ‘safe’ in terms of performance. If the bind is deemed ‘unsafe’ then multiple children are created for each set of different bound values so that different plans can be associated with them. This occurs in a few scenarios but the most common is with histogram stats on an equality predicate.

CURSOR_SHARING=SIMILAR

With CURSOR_SHARING=SIMILAR whenever the optimizer looks at a replaced bind value to make a decision then that bind is checked to see if it should be considered unsafe. The check made is :

Is the operator NEITHER of   ‘=’ or ‘!=’
OR
Are there Column Histograms present on the column.

If either of these are true then the bind is deemed to be unsafe and a new cursor will be created (So binds used in non equality predicates (eg >, <, >=, <=, LIKE) are unsafe). To check for whether a bind is considered unsafe see:

Note:261020.1 High Version Count with CURSOR_SHARING = SIMILAR or FORCE

With histogram stats on an equality predicate, this can cause severe problems (in terms of sharing) if there is, for example, a histogram on a main key column.eg:

select … from orders where orderid=’literal’;

If  there is a histogram on “orderid” then the bind will likely get marked unsafe and a new child will be produced for every single value of ‘literal’. The SQL would only be shared if the value of ‘literal’ matches exactly to a value used already.

Remember that if literals are converted to binds due to CURSOR_SHARING then they are subject to this checking, though unsafe binds are really only an issue if CURSOR_SHARING is SIMILAR.

In an OLTP type environment it would be sensible to only use histogram statistics on columns that need it (i.e. where there are only a few distinct values with heavy skew and where different plans are needed for different predicate values). This way most SQL is shared and the absence of histograms keeps predicates safe except where needed.

CURSOR_SHARING = FORCE

If CURSOR_SHARING = FORCE binds can still be “unsafe” (in terms of performance) if used by the optimizer in decisions but this should not then affect shareability of the SQL since CURSOR_SHARING=FORCE does not care about unsafe literals, but the cursor should stil lbe shared. e.g.: In the above example with orderid=’literal’, without histograms, the CBO does not need to look at ‘literal’ to determine the selectivity of the predicate and so the bind does not get marked unsafe.
If there is histograms, the predicate is marked as unsafe, but since FORCE uses the same plan whatever the circumstance, this does not matter. Only where non data literals for whom different values alter the actual meaning of the SQL (e.g. order by 1 versus order by 2) will an unsafe predicate have an affect on plans.

Note that, prior to 11g, unsafe literals are NOT covered by ‘bind mismatch’ in V$SQL_SHARED_CURSOR  as this is for user bind metadata mismatches. ie: different max bind lengths or bind type mismatches.
In 11g R2 (and 11.1.0.7 Patchset) a new column has been added to V$SQL_SHARED_CURSOR to check if literal replacement is used with CURSOR_SHARING=SIMILAR. The new column HASH_MATCH_FAILED  is set to “Y” if sharing fails due to a hash mismatch, such as the case with mismatched histogram data or a range predicate marked as unsafe by literal replacement. The unshared child cursors may have histogram data
on key columns used in equality predicates, or range predicates with literals which the optimizer has marked
as unsafe.

From the optimizer point of view, these additional plans for ‘unsafe’ bind variables explain why multiple good plans may occur for peeked bind variables, even though the standard behavior for binds is to use peeked binds for the initial parse and then use the resultant plan for all other iterations. With unsafe binds, different plans for different peeked values can occur.

Background Information

This issue has been coming more in to focus with dynamic sampling in 10.2 since the default was changed from 1 to 2. When optimizer_dynamic_sampling is greater than 1 then Dynamic sampling emulates statistics + histograms. If histograms are created, then binds may be marked as unsafe and produce different plans for different values. With optimizer_dynamic_sampling > 1 a predicate can create a new version of a sql statement for each different value, even if there are no histograms (or even statistics) on a table (since dynamic sampling may create these in the background).