I’ve been doing some HTTP Archive metrics research and I’ve come across a problem.
To begin with, I was checking if features (Blink use counters) were used at all, and looking for a correlation with the FCP metric. That gave results that looked reasonable. Then I thought to improve it by checking if the features were used before FCP to exclude things that couldn’t have influenced it.
This is where things start to look odd. I see a lot (~79k) of negative values in the data I use, illustrated by this query:
CAST(JSON_EXTRACT(payload, ‘$._blinkFeatureFirstUsed.Features.DocumentAll’) AS FLOAT64) AS time
GROUP BY url, time
HAVING time < 0
I see that with
httparchive.pages.2019_07_01_mobile there are far fewer, just 108. Was this a deliberate fix, does it make sense to compare time values starting with the most recent crawl? Or are the timelines of this and other metrics still different?