It asked me to solve a CAPTCHA when I wasn't adding a link. I don't know if that's expected.

It eventually posted a ==New section== for me, but it was slow. I got the red error message on a Reply.

In T279119#6967144, @Whatamidoing-WMF wrote:

It asked me to solve a CAPTCHA when I wasn't adding a link. I don't know if that's expected.

CAPTCHA is expected, tho I'm glad you mentioned it as I'd forgotten to include it in the task description.

So, I installed https://www.mediawiki.org/wiki/Excimer and followed the instructions there to produce a flame graph of the relevant API request, adding the following code in LocalSettings.php:

if ( isset( $_POST['action'] ) && $_POST['action'] === 'discussiontoolsedit' ) {
	$profiler = new ExcimerProfiler;
	$profiler->setPeriod( 0.0001 );
	$profiler->setEventType( EXCIMER_CPU );
	$profiler->start();
	register_shutdown_function( function () use ( $profiler ) {
		$profiler->stop();
		$pipe = popen( "/var/www/html/wikis/292134a3316743768728e0552c1e7cc3/w/FlameGraph/flamegraph.pl > /var/www/html/wikis/292134a3316743768728e0552c1e7cc3/w/profile.svg", "w" );
		fwrite( $pipe, $profiler->getLog()->formatCollapsed() );
	} );
}

Which produces the following output:

Two problems are apparent, I annotated them on the image:

1. [blue circles] We parse the wikitext of the page 5 times, which is definitely way too many (at most 2 would be reasonable, one for the current user viewing the page and one canonical parse using default settings for non-logged-in users).
   
   This probably affects production wikis as well, but caching and job queue setup might make it less bad. In particular, the last 2 parses happening inside DeferredUpdates probably are executed in the job queue in production.

2. [green circles] Within each parse, about half of the time is spent waiting on HTTP requests (using the Guzzle library), which are probably fetching image data from Commons.
   
   This is specific to the setup on Patchdemo, in production wikis this is done more efficiently with direct database queries.

I'm not sure what we can (or should) do about these, I'm stopping here for now and might look at this more next week.

[green circles] Within each parse, about half of the time is spent waiting on HTTP requests (using the Guzzle library), which are probably fetching image data from Commons.

I think this is T235551.