apachebench does help with efficiency. It is
inefficient (not to
irresponsible) to wait to benchmark your code until you have a
users actually relying on it.
It isn't matter of benchmarking, it is matter of relative resource
costs, compared with other projects, and no way you'd get the answer
with benchmarking tool.
Do note, we have a cluster which serves thousands of other requests,
so we may have capacity to serve various exceptions, but we may as
well want to spend that capacity on more efficient things, like, um,
wikis without eyecandies (interface language changes for english
language wikis? way to go!)
Apachebench will allow you to set the header
needed to trigger the language content shift in mediawiki while
simultaneously hammering the server with such requests.
We already know the costs, we do profile. We have way more detailed
costs than anything 'ab' can provide.
We have way more capacity in our cluster than a single-threaded
benchmark tool can load properly.