Averages suck, why compare to average? about tmi

[addyosmani]

Thanks Ilya. I agree there have to be better metrics available to compare against - the median is a good first step in moving away from the average. I'm happy to dive into the quantile approach you suggested too.

On the technical side, it appears there are a few existing modules for parsing data from bigquery, so it should in theory be relatively trivial to get the median in place. I may have to ask some follow up questions on reasoning quantiles for individual user queries.

[igrigorik]

@addyosmani the numbers don't change that much, I think we can just import + version a CSV with results. Here's a quick query:

SELECT * FROM
  (SELECT 'desktop' type,
    round(avg(round(bytesImg/1024))) average,
    NTH(10, quantiles(round(bytesImg/1024))) p10,
    NTH(20, quantiles(round(bytesImg/1024))) p25,
    NTH(30, quantiles(round(bytesImg/1024))) p30,
    NTH(50, quantiles(round(bytesImg/1024))) p40,
    NTH(60, quantiles(round(bytesImg/1024))) p60,
    NTH(70, quantiles(round(bytesImg/1024))) p70,
    NTH(75, quantiles(round(bytesImg/1024))) p75,
    NTH(80, quantiles(round(bytesImg/1024))) p80,
    NTH(85, quantiles(round(bytesImg/1024))) p85,
    NTH(90, quantiles(round(bytesImg/1024))) p90,
    NTH(95, quantiles(round(bytesImg/1024))) p95
    FROM [httparchive:runs.2014_10_15_pages]),
  (SELECT 'mobile' type,
    round(avg(round(bytesImg/1024))) average,
    NTH(10, quantiles(round(bytesImg/1024))) p10,
    NTH(20, quantiles(round(bytesImg/1024))) p25,
    NTH(30, quantiles(round(bytesImg/1024))) p30,
    NTH(50, quantiles(round(bytesImg/1024))) p40,
    NTH(60, quantiles(round(bytesImg/1024))) p60,
    NTH(70, quantiles(round(bytesImg/1024))) p70,
    NTH(75, quantiles(round(bytesImg/1024))) p75,
    NTH(80, quantiles(round(bytesImg/1024))) p80,
    NTH(85, quantiles(round(bytesImg/1024))) p85,
    NTH(90, quantiles(round(bytesImg/1024))) p90,
    NTH(95, quantiles(round(bytesImg/1024))) p95
  FROM [httparchive:runs.2014_10_15_pages_mobile]);

type,average,p10,p25,p30,p40,p60,p70,p75,p80,p85,p90,p95
desktop,1207.0,47.0,144.0,269.0,614.0,853.0,1189.0,1411.0,1706.0,2112.0,2762.0,4073.0
mobile,660.0,17.0,59.0,114.0,302.0,439.0,639.0,816.0,994.0,1232.0,1639.0,2513.0

[addyosmani]

That approach works for me. I've signed up for BigQuery and am happy to check-in and update the CSV periodically. Will let you know once we've updated to support medians. Shouldn't take long.

[addyosmani]

Implemented in branch. How do you feel about this for the output?

[igrigorik]

I think you're still average in your calculation, and I just realized that I omitted p50 in above query - doh.

What's your thinking behind asking the site author to get to a median? Seems like we should, instead, ask them to make it as small as possible to meet their use case: no reason to stop at p50, and sometimes (due to business requirements, etc), it may not be possible either. I'd suggest something like...

• Your image weight: 3.31MB
  • Desktop: 90th percentile (median site: 700KB)
  • Mobile: 95th percentile (median site: 350KB)
• needs some call to action at the end...

[addyosmani]

It's my turn to doh! :) I completely misread the field heading. You're right - we're still reading the average here.

What's your thinking behind asking the site author to get to a median?

The thought here was that getting them to quantify their image weight compared to how well/badly the rest of the web is doing, you had some incentive to work on trimming down your overall size of said images. Percentiles make more sense to use than averages however.

I'd suggest something like...

I'll get a version done that works as suggested. The revised query should be:

SELECT * FROM
  (SELECT 'desktop' type,
    round(avg(round(bytesImg/1024))) average,
    NTH(10, quantiles(round(bytesImg/1024))) p10,
    NTH(20, quantiles(round(bytesImg/1024))) p25,
    NTH(30, quantiles(round(bytesImg/1024))) p30,
    NTH(40, quantiles(round(bytesImg/1024))) p40,
    NTH(50, quantiles(round(bytesImg/1024))) p50,
    NTH(60, quantiles(round(bytesImg/1024))) p60,
    NTH(70, quantiles(round(bytesImg/1024))) p70,
    NTH(75, quantiles(round(bytesImg/1024))) p75,
    NTH(80, quantiles(round(bytesImg/1024))) p80,
    NTH(85, quantiles(round(bytesImg/1024))) p85,
    NTH(90, quantiles(round(bytesImg/1024))) p90,
    NTH(95, quantiles(round(bytesImg/1024))) p95
    FROM [httparchive:runs.2014_10_15_pages]),
  (SELECT 'mobile' type,
    round(avg(round(bytesImg/1024))) average,
    NTH(10, quantiles(round(bytesImg/1024))) p10,
    NTH(20, quantiles(round(bytesImg/1024))) p25,
    NTH(30, quantiles(round(bytesImg/1024))) p30,
    NTH(40, quantiles(round(bytesImg/1024))) p40,
    NTH(50, quantiles(round(bytesImg/1024))) p50,
    NTH(60, quantiles(round(bytesImg/1024))) p60,
    NTH(70, quantiles(round(bytesImg/1024))) p70,
    NTH(75, quantiles(round(bytesImg/1024))) p75,
    NTH(80, quantiles(round(bytesImg/1024))) p80,
    NTH(85, quantiles(round(bytesImg/1024))) p85,
    NTH(90, quantiles(round(bytesImg/1024))) p90,
    NTH(95, quantiles(round(bytesImg/1024))) p95
  FROM [httparchive:runs.2014_10_15_pages_mobile]);

does that look right?

[igrigorik]

Re, query: yep. Let's drop the average column, not even sure why I included it -- only thing it does is promote the bad practice.

The thought here was that getting them to quantify their image weight compared to how well/badly the rest of the web is doing, you had some incentive to work on trimming down your overall size of said images. Percentiles make more sense to use than averages however.

For the call to action, how about...

• Your image weight: 3.31MB
  • Your site delivers more image bytes than 90% of desktop sites:
    • +2.3MB compared to a site in 75th percentile
    • +2.7MB compared to a site in 50th percentile
    • +3.1MB compared to a site in 25th percentile
  • Your site delivers more image bytes than 95% of mobile sites:
    • ...

[addyosmani]

Current progress:

Next: Your site delivers more image bytes than X% of mobile sites:

[igrigorik]

[addyosmani]

Support for this has landed in master and should be in the next release.