Re: Re: force index range scan vs index full scan through hints?

Hi,
i think you can hardly get one index range scan, whatever you do. Consider what that means:
It would mean the optimizer has to
 find the minimum start value out of all or clauses to start the scan. Further it has to find the maximum end value out of all clauses to end the scan. That could be non trivial as the conditions could contain bind variables and complex conditions. The ranges could be overlaping etc. (I was told the MS SQLserver has some hardcoded logic that will do what you want if it recognizes a certain query pattern. This pattern is used for ISAM emulation and it resembles to some extend the or .. between conditions you are describing.) What you can get is a number of range scans, one for each or clause. In order to achive it you need "OR Expansion" to happen. (https://blogs.oracle.com/optimizer/optimizer-transformations:-or-expansion)  You can trigger or expansion by the use_concat (oldfashioned) hint or the OR_EXPAND hint (newer and more precise). You can find a discussion here: https://jonathanlewis.wordpress.com/category/oracle/hints/. One issue that occured here was thaz your index hint can not be honored if or expansion did not happend first. However the index hint is an access path hint, which is checked after the transformation phase. The transformation phase is the phase where the optimizer will check if he does an or expansion or not. Once the optimizer has decided against an or expansion, the index range scan is not possible. Thus, seemingly the optimizer did not honor your hint. Regards
Lothar
----Ursprüngliche Nachricht----
Von : cstephens16_at_gmail.com
Datum : 13/08/2018 - 13:33 (GMT)
An : christopherdtaylor1994_at_gmail.com Cc : niall.litchfield_at_gmail.com, oracle-l_at_freelists.org Betreff : Re: force index range scan vs index full scan through hints? thanks chris/niall! that's exactly the solution we are pushing. they drop and recreate the entire schema so locking won't work but we've created a script to run after tables are recreated that sets stats and gets us the plans we want. for some reason that's seen as too complicated. SQL_ID changes constantly as number of "OR"'d conditions is dynamic. thanks for all the suggestions!
On Sat, Aug 11, 2018 at 4:19 PM Chris Taylor <christopherdtaylor1994_at_gmail.com> wrote: I do like Niall's suggestion about setting the stats. You could also lock those stats and as long as you're not using histograms they should be good to stay locked. Chris
On Sat, Aug 11, 2018, 4:34 PM <niall.litchfield_at_gmail.com> wrote: I suggest that you modify the code that creates the table on the fly to. 1. Create the table as currently
2..SET statistics on the table that tell the optimizer it's got lots of rows in it (and is quite large and .. and) same for the index. Then schedule a gather later. On Fri, 10 Aug 2018, 14:51 Chris Stephens, <cstephens16_at_gmail.com> wrote: Is there a way to tell Oracle to use an index range scan instead of an index full scan? We have a table that gets created on the fly and quickly accumulates 100's of millions of rows before statistics are collected. Never mind why the table has no statistics (it's a (longish) story). As the table accumulates data, another process hits it with a piece of SQL that has a variable number of predicates of the from "id between :id1 and :id2 OR id between :id2 and id3 OR...". We have an index to support that SQL. Oracle initially chooses to full scan that index instead of range scan it until statistics are collected. Full scan performance is unacceptable. Is there a way to force the index range scan version? We can't use baselines (another story). We've tried INDEX_RS and INDEX_RS_ASC after observing INDEX hint doesn't get us what we want since index is already being used. INDEX_RS and INDEX_RS_ASC aren't in official documentation so I'm not even sure they are valid hints. As I write this, it occurs to me that dynamic sampling might help here and will try that but that obviously comes with some overhead. Any other options? Thanks for any insight.
chris

--
http://www.freelists.org/webpage/oracle-l