Re: OCTET_LENGTH is wrong

> Peter Eisentraut <peter_e@gmx.net> writes:
> > I noticed OCTET_LENGTH will return the size of the data after TOAST may
> > have compressed it.  While this could be useful information, this
> > behaviour has no basis in the SQL standard and it's not what is
> > documented.  Moreover, it eliminates the standard useful behaviour of
> > OCTET_LENGTH, which is to show the length in bytes of a multibyte string.
> 
> I wondered about that too, the first time I noticed it.  On the other
> hand, knowing the compressed length is kinda useful too, at least for
> hacking and DBA purposes.  (One might also like to know whether a value
> has been moved out of line, which is not currently determinable.)

It seems the behavior of OCTET_LENGTH varies acording to the
corresponding data type:

TEXT: returns the size of data AFTER TOAST
VARCHAR and CHAR: returns the size of data BEFORE TOAST

I think we should fix at least these inconsistencies but am not sure
if it's totally wrong that OCTET_LENGTH returns the length AFTER
TOAST. The SQL standard does not have any idea about TOAST of course.
Also, I tend to agree with Tom's point about hackers and DBAs.

> I don't want to force an initdb at this stage, at least not without
> compelling reason, so adding more functions right now is not feasible.
> Maybe a TODO item for next time.
>
> That leaves us with the question whether to change OCTET_LENGTH now
> or leave it for later.  Anyone?

My opinion is leaving it for 7.3, with the idea (adding new
functions).

> BTW, I noticed that textlength() is absolutely unreasonably slow when
> MULTIBYTE is enabled --- yesterday I was trying to profile TOAST
> overhead, and soon discovered that what I was looking at was nothing
> but pg_mblen() calls.  It really needs a short-circuit path for
> single-byte encodings.

It's easy to optimize that. However I cannot access CVS anymore after
the IP address change. Will post patches later...
--
Tatsuo Ishii