Extreme Nulls

This note is a variant of a note that I wrote a few months ago about the impact of nulls on column groups. The effect showed up recently on a client site with a little camouflage that confused the issue for a little while, so I thought it would be worth a repeat.  We’ll start with a script to generate some test data:

rem
rem     Script:         pt_hash_cbo_anomaly.sql
rem     Author:         Jonathan Lewis
rem     Dated:          Dec 2018
rem     Purpose:        
rem
rem     Last tested 
rem             12.1.0.2
rem

create table t1 (
        hash_col,
        rare_col,
        n1,
        padding
)
nologging
partition by hash (hash_col)
partitions 32
as
with generator as (
        select 
                rownum id
        from dual 
        connect by 
                level <= 1e4 -- > comment to avoid WordPress format issue
)
select
        mod(rownum,128),
        case when mod(rownum,1021) = 0 
                then rownum + trunc(dbms_random.value(-256, 256))
        end case,
        rownum,
        lpad('x',100,'x')               padding
from
        generator       v1,
        generator       v2
where
        rownum <= 1048576 -- > comment to avoid WordPress format issue
;

create index t1_i1 on t1(hash_col, rare_col) nologging
local compress 1
;

begin
        dbms_stats.gather_table_stats(
                ownname     => null,
                tabname     => 'T1',
                granularity => 'ALL',
                method_opt  => 'for all columns size 1'
        );
end;
/

I’ve got a hash-partitioned table with 32 partitions; the partitioning key is called hash_col, and there is another column called rare_col that is almost alway null – roughly 1 row in every 1,000 holds a value. I’ve added a local index on (hash_col, rare_col) compressing the leading column since hash_col is very repetitive, and gathered stats on the partitions and table. Here’s a view of the data for a single value of hash_col, and a summary report of the whole data set:

select  
        hash_col, rare_col, count(*)
from
        t1
where
        hash_col = 63
group by
        hash_col, rare_col
order by
        hash_col, rare_col
;

  HASH_COL   RARE_COL   COUNT(*)
---------- ---------- ----------
        63     109217          1
        63     240051          1
        63     370542          1
        63     501488          1
        63     631861          1
        63     762876          1
        63     893249          1
        63    1023869          1
        63                  8184

9 rows selected.

select
        count(*), ct
from    (
        select
                hash_col, rare_col, count(*) ct
        from
                t1
        group by
                hash_col, rare_col
        order by
                hash_col, rare_col
        )
group by ct
order by count(*)
;

  COUNT(*)         CT
---------- ----------
         3       8183
       125       8184
      1027          1

Given the way I’ve generated the data any one value for hash_col will have there are 8,184 (or 8,183) rows where the rare_col is null; but there are 1027 rows which have a value for both hash_col and rare_col with just one row for each combination.

Now we get to the problem. Whenever rare_col is non null the combination of hash_col and rare_col is unique (though this wasn’t quite the case at the client site) so when we query for a given hash_col and rare_col we would hope that the optimizer would be able to estimate a cardinality of one row; but this is what we see:

variable n1 number
variable n2 number

explain plan for
select /*+ index(t1) */
        n1
from
        t1
where
        hash_col = :n1
and     rare_col = :n2
;

select * from table(dbms_xplan.display);

========================================

--------------------------------------------------------------------------------------------------------------------
| Id  | Operation                                  | Name  | Rows  | Bytes | Cost (%CPU)| Time     | Pstart| Pstop |
--------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                           |       |   908 | 10896 |    76   (0)| 00:00:01 |       |       |
|   1 |  PARTITION HASH SINGLE                     |       |   908 | 10896 |    76   (0)| 00:00:01 |   KEY |   KEY |
|   2 |   TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| T1    |   908 | 10896 |    76   (0)| 00:00:01 |   KEY |   KEY |
|*  3 |    INDEX RANGE SCAN                        | T1_I1 |   908 |       |     2   (0)| 00:00:01 |   KEY |   KEY |
--------------------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------
   3 - access("HASH_COL"=TO_NUMBER(:N1) AND "RARE_COL"=TO_NUMBER(:N2))

The optimizer has predicted a massive 908 rows. A quick check of the object stats shows us that this is “number of rows in table” / “number of distinct keys in index” (1,048,576 / 1,155, rounded up).

Any row with rare_col set to null cannot match the predicate “rare_col = :n2”, but because the optimizer is looking at the statistics of complete index entries (and there are 1048576 of them, with 1155 distinct combinations, and none that are completely null) it has lost sight of the frequency of nulls for rare_col on its own. (The same problem appears with column groups – which is what I commented on in my previous post on this topic).

I’ve often said in the past that you shouldn’t create histograms on data unless your code is going to use them. In this case I need to stop the optimizer from looking at the index.distinct_keys and one way to do that is to create a histogram on one of the columns that defines the index; and I’ve chosen to do this with a fairly arbitrary size of 10 buckets:

execute dbms_stats.gather_table_stats(user,'t1',method_opt=>'for columns rare_col size 10')

explain plan for
select /*+ index(t1) */
        n1
from
        t1
where
        hash_col = :n1
and     rare_col = :n2
;

select * from table(dbms_xplan.display);

========================================

--------------------------------------------------------------------------------------------------------------------
| Id  | Operation                                  | Name  | Rows  | Bytes | Cost (%CPU)| Time     | Pstart| Pstop |
--------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                           |       |     1 |    12 |     2   (0)| 00:00:01 |       |       |
|   1 |  PARTITION HASH SINGLE                     |       |     1 |    12 |     2   (0)| 00:00:01 |   KEY |   KEY |
|   2 |   TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| T1    |     1 |    12 |     2   (0)| 00:00:01 |   KEY |   KEY |
|*  3 |    INDEX RANGE SCAN                        | T1_I1 |     1 |       |     1   (0)| 00:00:01 |   KEY |   KEY |
--------------------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------
   3 - access("HASH_COL"=TO_NUMBER(:N1) AND "RARE_COL"=TO_NUMBER(:N2))

Bonus observation

This problem came to my attention (and I’ve used a partitioned table in my demonstration) because I had noticed an obvious optimizer error in the client’s execution plan for exactly this simple a query. I can demonstrate the effect the client saw by running the test again without creating the histogram but declaring hash_col to be not null. Immediately after creating the index I’m going to add the line:

alter table t1 modify hash_col not null;

(The client’s system didn’t declare the column not null, but their equivalent of hash_col was part of the primary key of the table which meant it was implicitly declared not null). Here’s what my execution plan looked like with this constraint in place:

--------------------------------------------------------------------------------------------------------------------
| Id  | Operation                                  | Name  | Rows  | Bytes | Cost (%CPU)| Time     | Pstart| Pstop |
--------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                           |       |   908 | 10896 |    76   (0)| 00:00:01 |       |       |
|   1 |  PARTITION HASH SINGLE                     |       |   908 | 10896 |    76   (0)| 00:00:01 |   KEY |   KEY |
|   2 |   TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| T1    |   908 | 10896 |    76   (0)| 00:00:01 |   KEY |   KEY |
|*  3 |    INDEX RANGE SCAN                        | T1_I1 |    28 |       |     2   (0)| 00:00:01 |   KEY |   KEY |
--------------------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------
   3 - access("HASH_COL"=TO_NUMBER(:N1) AND "RARE_COL"=TO_NUMBER(:N2))

Spot the difference.

The estimate of index rowids is far smaller than the estimate of the rows that will be fetched using those rowids. This is clearly an error.

If you’re wondering how Oracle got this number divide 908 by 32 (the number of partitions in the table) – the answer is 28.375.

Fortunately it’s (probably) an error that doesn’t matter despite looking worryingly wrong. Critically the division hasn’t changed the estimate of the number of table rows (we’ll ignore the fact that the estimate is wrong anyway thanks to a different error), and the cost of the index range scan and table access have not changed. The error is purely cosmetic in effect.

Interestingly if you modify the query to be index-only (i.e. you restrict the select list to columns in the index) this extra division disappears.

Summary

1) If you have a B-tree index where one (or more) of the columns is null for a large fraction of the entries then the optimizer may over-estimate the cardinality of a predicate of the form: “(list of all index columns) = (list of values)” as it will be using the index.distinct_keys in its calculations and ignore the effects of nulls in the individual columns. If you need to work around this issue then creating a histogram on one of the index columns will be sufficient to switch Oracle back to the strategy of multiplying the individual column selectivities.

2) There are cases of plans for accessing partitioned tables where Oracle starts by using table-level statistics to get a suitable set of estimates but then displays a plan with the estimate of rows for an index range scan scaled down by the number of partitions in the table. This results in a visible inconsistency between the index estimate and the table estimate, but it doesn’t affect the cardinality estimate for the table access or either of the associated costs – so it probably doesn’t have a destabilising effect on the plan.