Re: backtrace_on_internal_error

On 05.12.23 11:55, Peter Eisentraut wrote:
> I think the implementation would be very simple, something like
>
> diff --git a/src/backend/utils/error/elog.c
> b/src/backend/utils/error/elog.c
> index 6aeb855e49..45d40abe92 100644
> --- a/src/backend/utils/error/elog.c
> +++ b/src/backend/utils/error/elog.c
> @@ -498,9 +498,11 @@ errfinish(const char *filename, int lineno, const
> char *funcname)
>
>     /* Collect backtrace, if enabled and we didn't already */
>     if (!edata->backtrace &&
> -       edata->funcname &&
> -       backtrace_functions &&
> -       matches_backtrace_functions(edata->funcname))
> +       ((edata->funcname &&
> +         backtrace_functions &&
> +         matches_backtrace_functions(edata->funcname)) ||
> +        (edata->sqlerrcode == ERRCODE_INTERNAL_ERROR &&
> +         backtrace_on_internal_error)))
>         set_backtrace(edata, 2);
>
>     /*
>
> where backtrace_on_internal_error would be a GUC variable.

It looks like many people found this idea worthwhile.

Several people asked for a way to set the minimum log level for this
treatment.

Something else to note: I wrote the above code to check the error code;
it doesn't check whether the original code write elog() or ereport().
There are some internal errors that are written as ereport() now.
Others might be changed from time to time; until now there would have
been no external effect from this. I think it would be weird to
introduce a difference between these forms now.

But then, elog() only uses the error code ERRCODE_INTERNAL_ERROR if the
error level is >=ERROR. So this already excludes everything below.

Do people want a way to distinguish ERROR/FATAL/PANIC?

Or do people want a way to enable backtraces for elog(LOG)? This didn't
look too interesting to me. (Many current elog(LOG) calls are behind
additional guards like TRACE_SORT or LOCK_DEBUG.)

If neither of these two are very interesting, then the above code would
already appear to do what was asked.