>So you mean to say: Using your SP you’ve manually modified the density
>while preserving the histogram – and the CBO uses then the manually
>modified density from the dictionary for calculation but not the NewDensity any longer?

Yes if you use the default and only supported value (null) for the “flags” parameter of dbms_stats.set_column_stats – as far as I understand, in this case the statistics get labeled as “user generated stats” (dba_tab_cols.user_stats = ‘YES’). If you set “flags” to 2, you are pretending to be dbms_stats or a system routine (dba_tab_cols.user_stats = ‘NO’) and hence NewDensity gets used.

It makes sense – if the user has provided her own value for “density”, use it instead of NewDensity.

NB: I have exchanged in the past many, many emails with Wolfgang and I’ve noticed that He commonly uses flags=2 in his scripts … that should explain everything nicely :)

thanks for the clarification.

So you mean to say: Using your SP you’ve manually modified the density while preserving the histogram – and the CBO uses then the manually modified density from the dictionary for calculation but not the NewDensity any longer?

I thought that Wolfgang meant to say that he has generated a histogram manually along with a density, but the CBO sill applied the NewDensity algorithm to his made-up histogram.

At least two things come into my mind that might matter:
- I don’t know the version he has used to test
- He seems to have setup a made-up histogram rather than preserving an existing one, however I don’t see where this exactly should make any difference

Anyway, thanks for the feedback. If I find some time I’ll test it myself resp. ask Wolfgang for clarification what he did and what version he used to test.

Randolf

thank you very much for your positive feedback, I really appreciate it – as you “may” know, it takes a lot of time to write about this kind of things trying to be clear and exhaustive …

I use the following tiny SP to set the density while preserving the histogram:

-- a utility procedure to change density, preserving the histogram
create or replace procedure set_density (
  p_table_name  varchar2,
  p_column_name varchar2,
  p_new_density number,
  p_flags       number default null
)
is
  l_distcnt     number;
  l_old_density number;
  l_nullcnt     number;
  l_srec        dbms_stats.statrec;
  l_avgclen     number;
  l_new_density number;
begin
  -- get the current column statistics
  dbms_stats.get_column_stats (
    ownname => user,
    tabname => p_table_name,
    colname => p_column_name,
    distcnt => l_distcnt,
    density => l_old_density,
    nullcnt => l_nullcnt,
    srec    => l_srec,
    avgclen => l_avgclen
  );

  -- reset them, overwriting "density"
  if p_new_density is not null then
    l_new_density := p_new_density;
  else
    l_new_density := l_old_density;
  end if;

  dbms_stats.set_column_stats (
    ownname => user,
    tabname => p_table_name,
    colname => p_column_name,
    distcnt => l_distcnt,
    density => l_new_density,
    nullcnt => l_nullcnt,
    srec    => l_srec,
    avgclen => l_avgclen,
    no_invalidate => false,
    flags         => p_flags
  );

  dbms_output.put_line ('density of '||p_table_name||'.'||p_column_name||' changed from '||l_old_density||' to '|| l_new_density);

end set_density;
/
show errors;

I’ve tested it right now again, albeit on 11.1.0.7 only, and it actually preserves the histogram. It is a script I’ve used for some years now.

Notice how it works : it fetches the histogram from the data dictionary, and rewrites it back verbatim with a new density, possibly changing the “flags” as well …

regarding your comment “Playing with density”: I’ve suggested in a recent discussion to refer to your excellent blog series on Oracle-l (and try to set the density manually to get the “old” density behaviour back) – yesterday Wolfgang Breitling mentioned when we met at the UKOUG that manually setting the density actually removes the histogram, and generating a histogram manually will still use the NewDensity but not the density provided as part of the user-generated statistics.

I haven’t had time yet to test this personally, but now that Wolfgang has mentioned it I’m quite sure I remember that he’s right regarding setting the density manually – it removes the complete histogram immediately.

Can you comment on what you meant by above: “If you set your own column stats using dbms_stats.set_column_stats…”?

Have you tried to modify the density only or generating your own histogram?

Randolf