Re: New "raw" COPY format

On Wed, Oct 30, 2024, at 09:14, Joel Jacobson wrote:
> $ psql -f bench_result.sql

Ops, I realized I benchmarked a debug build,
reran the benchmark with `meson setup build --buildtype=release`,
and also added benchmarking of HEAD:

[plot image showing time distribution by format and version]

For those unable to see the image, here is the same data as a text table:

select format, version, min(time_ms), round(avg(time_ms),2) avg, max(time_ms), round(stddev(time_ms)) as stddev from bench group by 1,2 order by 1,2;

format | version | min | avg | max | stddev
--------+---------+----------+---------+----------+--------
csv | HEAD | 2862.537 | 2895.91 | 2968.420 | 27
csv | v16 | 2834.663 | 2869.02 | 2926.489 | 28
csv | v17 | 2958.570 | 3002.71 | 3056.949 | 31
raw | v16 | 1697.899 | 1723.48 | 1798.020 | 22
raw | v17 | 1720.251 | 1753.56 | 1811.399 | 28
text | HEAD | 2442.845 | 2476.60 | 2543.767 | 33
text | v16 | 2358.062 | 2389.06 | 2472.382 | 32
text | v17 | 2517.778 | 2552.65 | 2621.161 | 30
(8 rows)

I think the improvement for the 'text' format between HEAD and v16 could be just noise,
possibly due to different binary layout when compiling.

v17 (single function) seems slower than v16 (separate functions), also in release build.

/Joel