Thread

Re: Problems with avg on interval data type

Thomas Lockhart <lockhart@alumni.caltech.edu> — 2001-05-18T16:47:10Z

> We have recently upgraded from 7.0.3 to 7.1 and a query which used
> to work is no longer working.
> The query does an avg on an interval column and now gets the error:
> ERROR:  Bad interval external representation '0'

OK, there is one case of interval constant which is not handled
correctly in the 7.1.x release -- the simplest interval specification
having only an unadorned integer. That is a bug, for which I have a
patch (or patches) available.

Before I post the patch (which should go into the 7.1.2 release as a bug
fix) I need feedback on a conventions dilemma, which led to the code
modifications which introduced the bug. Here it is:

Intervals usually indicate a time span, and can be specified with either
"# units" strings (e.g. '5 hours') or (as of 7.1) as "hh:mm:ss" (e.g.
'05:00').

A new construct, "a_expr AT TIME ZONE c_expr" is supported in 7.1, per
SQL99 spec. One of the possible arguments is

  a_expr AT TIME ZONE 'PST'

and

  a_expr AT TIME ZONE INTERVAL '-08:00'

It is this last style which leads to the problem of how to interpret
signed or unsigned integers as interval types. For example, in this
context

  INTERVAL '-8'

must be interpreted as having units of "hours", while our historical
usage has

  INTERVAL '8'

being interpreted as "seconds" (even with signed values). Currently, we
interpret various forms as follows:

  Value	Units
  +8	hours
  -8	hours
  8.0	seconds
  8	?? seconds ??

I would propose that the last example should be interpreted in units of
seconds, but that could be perilously close to the conventions required
for the signed examples. Comments?

                     - Thomas