On Fri, Sep 09, 2022 at 10:45:37AM -0300, Ranier Vilela wrote:
> Based on work in [1].
> According to https://cplusplus.com/reference/cstdio/fprintf/
> The use of fprintf is related to the need to generate a string based on a
> format, which should be different from "%s".
> Since fprintf has overhead when parsing the "format" parameter, plus all
> the trouble of checking the va_arg parameters.
> I think this is one of the low fruits available and easy to reap.
> By replacing fprintf with its equivalents, fputs and fputc,
> we avoid overhead and increase security [2] and [3].
>
> The downside is a huge big churm, which unfortunately will occur.
> But, IHMO, I think the advantages are worth it.
> Note that behavior remains the same, since fputs and fputc do not change
> the expected behavior of fprintf.
>
> A small performance gain is expected, mainly for the client, since there
> are several occurrences in some critical places, such as
> (usr/src/fe_utils/print.c).
I agree with David [0]. But if you can demonstrate a performance gain,
perhaps it's worth considering a subset of these changes in hot paths.
Simple benchmark with David sort example.
1. make data
create table t (a bigint not null, b bigint not null, c bigint not
null, d bigint not null, e bigint not null, f bigint not null);
insert into t select x,x,x,x,x,x from generate_Series(1,140247142) x; -- 10GB!
vacuum freeze t;
2. client run
\timing on
\pset pager off
select * from t limit 1000000;
head:
Time: 418,210 ms
Time: 419,588 ms
Time: 424,713 ms
fprintf patch:
Time: 416,919 ms
Time: 416,246 ms
Time: 416,237 ms
regards,
Ranier Vilela