countme-totals: DNF user counting data

This repo contains aggregate countme data about DNF clients gathered from https://mirrors.fedoraproject.org/.

For more information about how this data is gathered, see the documentation for the mirrors-countme scripts: https://pagure.io/mirrors-countme/

What's here?

• totals.db: total counts, in a SQLite database
• totals.csv: same, in CSV format (with a header line)
• mirrorsdata-2020.csv: Counts of unique IP addresses, from https://data-analysis.fedoraproject.org/csv-reports/mirrors/mirrorsdata-2020.csv
• jupyter/: Jupyter notebooks showing how to use this data
  • jupyter/dnf-countme-pandas-demo.ipynb: how to summarize, slice, and graph data from totals.csv in Python, using pandas

What's being counted?

This data counts all valid countme requests (grouped per-week) since the feature was first enabled in Fedora 32, on 2020-02-11.

It's broken down by:

• The client OS name, version, variant, and arch,
• the "age" of the system (on a scale from 1-4), and
• which repo (and arch) it requested.

The actual requests look (roughly) like this:

User-Agent: libdnf (Fedora 32; workstation; Linux.x86_64)
GET /metalink?repo=fedora-modular-32&arch=x86_64&countme=1 HTTP/2.0

We split this up into 7 pieces of info:

User-Agent: libdnf ({os_name} {os_version}; {os_variant}; Linux.{os_arch})
GET /metalink?repo={repo_tag}&arch={repo_arch}&countme={sys_age} HTTP/2.0

And so each item in totals.db and totals.csv is a count of how many times we saw that particular combination of:

(os_name, os_version, os_variant, os_arch, sys_age, repo_tag, repo_arch)

in a given week.

totals.csv

The CSV version of this data contains a header line, and it looks like this:

week_start,week_end,hits,os_name,os_version,os_variant,os_arch,sys_age,repo_tag,repo_arch
2020-02-10,2020-02-16,2,Fedora,31,generic,x86_64,3,fedora-modular-31,x86_64
2020-02-10,2020-02-16,2,Fedora,31,generic,x86_64,3,updates-released-f31,x86_64
2020-02-10,2020-02-16,1,Fedora,31,workstation,x86_64,3,updates-released-modular-f31,x86_64
2020-02-10,2020-02-16,2,Fedora,31,workstation,x86_64,3,fedora-31,x86_64
...
2020-07-13,2020-07-19,2,Fedora,32,workstation,x86_64,1,rawhide-modular,x86_64
2020-07-13,2020-07-19,1,Fedora,32,,x86_64,1,fedora-32,x86_64
2020-07-13,2020-07-19,1,Fedora,32,silverblue,x86_64,1,fedora-rawhide,x86_64
2020-07-13,2020-07-19,1,Fedora,32,workstation,x86_64,3,updates-released-f33,x86_64
2020-07-13,2020-07-19,1,Fedora,32,workstation,x86_64,3,fedora-33,x86_64

The columns are:

1. week_start : ISO-formatted date for the first day (Monday) of the week

2. week_end : ISO-formatted date for the last day (Sunday) of the week

3. hits : number of requests during this week

4. os_name : NAME from /etc/os-release

5. os_version : VERSION_ID from /etc/os-release

6. os_variant : VARIANT_ID from /etc/os-release

7. os_arch : dnf $basearch

8. sys_age : System age, 1-4

9. repo_tag : the requested URL's repo= value

10. repo_arch : the requested URL's arch= value.

totals.db

The SQLite version of this data is in a table named countme_totals and looks like this:

hits	weeknum	os_name	os_version	os_variant	os_arch	sys_age	repo_arch	repo_tag
71482	2636	Fedora	32	workstation	x86_64	3	x86_64	updates-released-f32
70546	2636	Fedora	32	workstation	x86_64	3	x86_64	updates-released-modular-f32
67035	2636	Fedora	32	workstation	x86_64	3	x86_64	fedora-modular-32
67001	2636	Fedora	32	workstation	x86_64	3	x86_64	fedora-32
28674	2636	Fedora	32	workstation	x86_64	1	x86_64	updates-released-f32
...	...	...	...	...	...	...	...	...

Instead of week_start and week_end, we use weeknum, where weeks start on Monday at 00:00:00 UTC, and week 0 started on the first Monday of the UNIX Epoch (Mon Jan 5 1970).

To convert a weeknum to a POSIX timestamp:

((weeknum*7)+4)*86400

You can thus convert weeknum to a SQLite date (or datetime):

date(((weeknum*7)+4)*86400,'unixepoch')

But it's actually easier to use the Julian day number:

date(julianday('1970-01-05')+weeknum*7)
date(2440591.5+weeknum*7)

Counting hits vs. hosts

One important point about this data: it counts hits, not hosts. We can approximate the number of hosts we saw by grouping the counts by all the host identifiers (os_name, os_version, os_variant, os_arch, sys_age) - which gives us hits for that host type across all repos - and taking the max() of that set. Like so:

select date(2440591.5+weeknum*7) as week_start,
       os_name, os_version, os_variant, os_arch, sys_age,
       sum(hits) as totalhits, max(hits) as hosts
from countme_totals
where os_name == 'Fedora' and os_version=='31' and os_arch=='x86_64' and weeknum==2636
group by os_name, os_version, os_variant, os_arch, sys_age;