BTW, on the principle of "where else did we make the same mistake",
I looked through the other aggregates using float8_regr_accum.
Most seem okay, but float8_regr_intercept does this:

        PG_RETURN_FLOAT8((Sy - Sx * Sxy / Sxx) / N);

Seems to me that expression is also prone to internal
overflow/underflow.  Underflow probably isn't a huge issue,
since the result will reduce to Sy/N which is likely to be good
enough.  But can we do anything about overflow?

One simple change that might make things better is to compute

        PG_RETURN_FLOAT8((Sy - Sx * (Sxy / Sxx)) / N);

on the theory that the sums of products are likely to both be large.

                        regards, tom lane


Reply via email to