DNF Counting

We use DNF Counting to get statistics about the number of Fedora installations.

Contact Information

Owner

Fedora Infrastructure Team

Contact

#fedora-admin, #fedora-noc, [email protected]

Servers

log01, proxy0*

Purpose

Give interested parties information about the number of Fedora installations.

Repositories

What it is

DNF Counting is a way for us to gather statistics about the number of Fedora installations, differentiated by version, spin, etc. On the infrastructure side this is implemented by a bunch of scripts and a Python package (mirrors-countme).

Scope

This SOP concerns itself with the infrastructure side of the equation. For any issues with the various frontends logging in to be counted (DNF, PackageKit, …), contact their respective maintainers or upstreams.

How it works

Clients (DNF, PackageKit, …) have been modified so they add a countme variable in their requests to mirrors.fedoraproject.org once a week. This ends up in our webserver log data which lets us generate usage statistics.

Cron jobs are set up on log01 which collect http log files from the various web proxies, combine them (accesses to different backend services including mirrors.fedoraproject.org are scattered across the proxy logs), and produce statistics from them. The various pieces live in a) the mirrors-countme project (Python package and related scripts to generate statistics from the log data) and b) shell scripts in the web-data-analysis role in Ansible:

  • sync-http-logs.py (Ansible) syncs individual log files from various hosts including proxies to log01.

  • combineHttpLogs.sh (Ansible) combines the logs for the different web sites which are scattered across the proxy hosts.

  • condense-mirrorlogs.sh & mirrorlist.py (Ansible) extracts hosts from the combined log data.

  • countme-update.sh (Ansible) drives countme-update-rawdb.sh & countme-update-totals.sh (mirrors-countme) which generates statistics.

  • countme-trim-raw (mirrors-countme) to trim the intermediary database file (raw.db).

Changes implemented in the Q2/2023 DNF mirrors-countme initiative

  • The “traditional“ statistics which were done before DNF learned about the countme variable were reimplemented: Count any individual IP sighted, no matter if with or without countme. This is necessary to count systems which don’t have that feature in their DNF or YUM, and – while giving different numbers – gives us an idea how things develop when compared to the same numbers for more modern OSes.

  • The countme-trim-raw tool was implemented, to trim the intermediary database raw.db which contains necessary information gleamed from parsing the merged log files. This database grows steadily and – with the brought back counting of any individual IP sighted – quickly, so once these data have been safely turned into the final statistics, we wanted a way to remove them so that the local volume were it is stored doesn’t fill up completely.

  • The project repository was cleaned up, i.e. large data files used in integration tests were removed because they made cloning the repository unnecessarily slow, for a couple hundred KB of code, the repo was more than 300 MB in size. In the context, the repository was moved from Pagure to GitHub.

  • Unused code was removed, the remaining code was refactored and condensed to remove redundancies and comprehensive unit tests were added so that the barrier to contributing is lower and changes are less risky.

Changes implemented in the Q3/2021 DNF Counting Initiative

During the Q3/2021 DNF Counting Initiative, a number of changes were implemented which improved the DNF Counting backend in the areas of monitoring & debugging, performance & robustness.

  • The involved scripts send messages about state changes and errors to the fedora-messaging bus. State changes are e.g. start and finish of a complete script or of its individual steps.

  • The shell script which syncs log files from various hosts to log01 (syncHttpLogs.sh) was reimplemented in Python (as sync-http-logs.py), with several improvements which reduced the time it takes for syncing from 6-7 hours to little more than 30 minutes per day:

    • All log files for one date of one host are synced in one call to rsync. This greatly reduces overhead.

      The reason to sync these files one-by-one previously was because rsync only allows differing file names when syncing single files, which we have: the log files on the hosts contain their date in the name, on log01 they don’t but are stored in directories for each date.

      To overcome this limitation, sync-http-logs.py maintains a shadow structure of hard links with dates in their names, and rsync operates on this structure instead, which are linked back to "date-less" file names afterwards for further processing.

    • Because syncing log files from some hosts is pretty slow, several hosts are synced in parallel.

  • Previously, syncHttpLogs.sh and combineHttpLogs.sh were run from individual cron jobs which were set to run a couple of hours apart. Sometimes, this caused problems because the former wasn’t finished when the latter started to run (i.e. a race condition). Now, sync-http-logs.py and combineHttpLogs.sh are run from one cron job to avoid this.

  • Previously, the scripts where scattered across the web-data-analysis, awstats and base roles. All of the deployment has been consolidated into the web-data-analysis role, awstats has been removed.

  • The mirrors-countme Python package and scripts are packaged as RPM packages in Fedora, previously they were deployed from a local clone of the upstream git repository.

Reboot me

Yes, just reboot. Or don’t. There are no continuously running services, everything is regularly run as cronjobs.

Logs

The sync-http-logs.py script sends relatively verbose output to syslog. Other than that, the closest anything comes to logs are mails sent if cronjobs produce (error) output and messages sent to the bus.

First steps to debug

The scripts send messages with a topic prefix of logging.stats to the bus, in various stages of their operation. If anything doesn’t work as it should, review if every step started is also finished, compare run times between days.

If anything crashes, cron should have sent mails to the recipients configured (at least [email protected]), which could also contain valuable information.

Ephemeral data

Generated CSV reports and images are in /var/www/html/csv-reports which are exposed on https://data-analysis.fedoraproject.org/ – but they get regenerated with every cycle of the scripts that is run.

Persistent data

All combined http log data is kept on the /fedora_stats NFS share. Log files from the proxy hosts are synced to /var/log/hosts/<hostname> locally, but these are just copies of what exists elsewhere already.

Other operational considerations

The scripts only process data from the previous three days (roughly). If they don’t run for a longer time, there might be gaps in the generated statistics which can be plugged by temporarily adjusting the respective settings in the scripts and re-running them.

Is there data that needs to be backed up?

Yes, but it’s on the /fedora_stats file share, so it’s assumed to get backed up regularly already.

Upgrading

mirrors-countme

The mirrors-countme shell and Python scripts create statistics from the already combined log data.

Making upstream changes available

Prerequisites: A change (bug fix or feature) is available in the main branch of mirrors-countme.

  1. Publish an upstream release

    From a clone of the upstream repository:

    1. In pyproject.toml, bump tool.poetry.version (e.g. to 0.1.2) and commit the change, e.g.:

      git commit -s -m "Version 0.1.2" -- pyproject.toml
    2. Tag the previous change with a GPG-signed tag:

      git tag -s 0.1.2
    3. Push both the change and the tag:

      git push origin main 0.1.2
    4. Create a source tarball (this will be created as e.g. dist/mirrors_countme-0.1.2.tar.gz):

      poetry build

      From the list of tags, select “Create release” in the menu for the respective tag, and attach the created tarball and wheel files to the created release.

  2. Update and Build the python-mirrors-countme Fedora Package

    From a clone of the Fedora package repository, in the rawhide branch:

    1. Bump the version in python-mirrors-countme.spec. No other changes are necessary, the packages uses automatic release fields and changelog.

    2. Download the source tarball, either manually or one of:

      spectool -g python-mirrors-countme.spec
      rpmspectool get python-mirrors-countme.spec
    3. Upload the source tarball to the lookaside cache:

      fedpkg new-sources mirrors_countme-0.1.2.tar.gz
    4. Commit the changes to the repository, e.g.:

      git commit -s -m "Version 0.1.2" -- python-mirrors-countme.spec
    5. Push the changes and build:

      git push && fedpkg build
    6. For any other active Fedora and EPEL branch, fast forward them to the state of the rawhide branch, push and build, e.g.:

      git checkout epel8 \
      && git merge --ff-only rawhide \
      && git push \
      && fedpkg build
  3. Submit Fedora/EPEL Package Updates

    Either submit the update via the Bodhi web interface, or from the command line in the respective checked out Fedora or EPEL branch, e.g.:

    fedpkg update --type bugfix --notes 'Put in some notes!'
  4. Tag with Infra-Tags in Koji

    1. Tag the build into the respective infra candidate tag in Koji, e.g.:

      koji tag-build epel8-infra-candidate
    2. Check that the build was picked up and signed (this should take no more than a few minutes), e.g.:

      koji buildinfo python-mirrors-countme-0.1.2-1.el8

      The build must be tagged with the corresponding *-infra-stg tag.

    3. Tag the build into the respective infra production tag in Koji, e.g.:

      koji tag-build epel8-infra

When the respective infra tag repository is updated, the new version should be ready to be installed/updated in our infrastructure.

Other scripts

Scripts other than what is contained in mirrors-countme live in the web-data-analysis role in Ansible. Simply "upgrade" them in place.

Deployment of updates

To deploy updated scripts, etc. from the Ansible repository, simply run the groups/logging.yaml playbook.

To update mirrors-countme, run the manual/update-packages.yml playbook with --extra-vars="package='mirrors-countme'" set.

The scripts send out status messages over fedora-messaging with a topic prefix of logging.stats.

How is it deployed?

All of this runs on log01.iad2.fedoraproject.org and is deployed through the web-data-analysis role and the groups/logserver.yml playbook, respectively.

The mirrors-countme upstream project publishes source tarballs to their corresponding releases in the repository on GitHub:

These are packaged in Fedora as the python-mirrors-countme (SRPM) and python3-mirrors-countme (RPM) packages.

Other scripts are located directly in the Fedora Infrastructure Ansible repository, in the web-data-analysis role.

Does it have any special requirements?

No.

Are there any security requirements?

The same as anything else that deals with log data.

Bug reports

Report bugs with mirrors-countme at its upstream project:

Anything concerning the cron jobs or other scripts should probably go into our Infrastructure tracker:

The same as anything else that deals with log data.