Background

Prior to T392880 and T351117, Varnish applied hostname normalization. This was (presumably) effective and was (presumably) observable in varnishkafka output:

• Kafkacat: webrequest_frontend_text
• Hadoop: wmf.webrequest and wmf.pageview_actor datasets
• Turnilo: webrequest_sampled_live

Problem

This isn't the case today. This came up during T405429 where I found some numbers didn't add up.

Turnilo: https://w.wiki/Fkf7

SELECT COUNT(*) _count, uri_host
FROM wmf.pageview_actor
WHERE year=2025 AND month=10 AND day=19
AND uri_host!=LOWER(uri_host)
GROUP BY uri_host ORDER BY _count DESC;

Commons.wikimedia.org
EN.m.wikipedia.org
commons.Wikimedia.org
EN.Wiktionary.org
En.Wiktionary.org
…