Researcher experience: friction points using comma2k19 / old rlogs

[mxj567-dotcom]

Describe the bug

(Posting under "PC bug report" as it's the closest template — this is researcher/dataset
UX feedback rather than a bug.)

Context: we're academic researchers studying driver monitoring and takeover behavior.

1. Old logs: deprecated fields & migration discoverability
   In comma2k19-era logs, the fields are hidden: controlsState's
   enabled/active/state live under the deprecated group, and driver monitoring is
   driverStateDEPRECATED.
2. Dataset cards: document which streams are populated
   In comma2k19 the driver monitoring stream exists but is all zeros (DM model wasn't
   running on those devices) — this isn't mentioned anywhere, so researchers interested
   in driver state only find out after decoding. A short "dataset card" per release
   (which messages are populated, at what rates, known gaps) would set expectations and
   make the datasets easier to cite in academic work.
3. Suggestion: a minimal research quickstart
   One documented example — rlog → pandas DataFrame with common channels (enabled,
   vEgo, steeringPressed, driver state probs, onroad events) — would open the datasets
   to a much wider research audience. Happy to contribute our decoding snippet as a
   starting point.

OS Version

Mac

openpilot version or commit

No response

Additional info

No response

[adeebshihadeh]

Excellent feedback, thanks!

For No. 1, we have a solution called migration.py. I think exposing it better will help, particularly making it the default for all log reading paths.

[mxj567-dotcom]

Excellent feedback, thanks!

For No. 1, we have a solution called migration.py. I think exposing it better will help, particularly making it the default for all log reading paths.

Thanks Adeeb — migration.py looks like exactly what we needed; we'll try it on the
comma2k19 rlogs.

For our use case (research), making it the default for all log reading paths would
fully solve this: we mostly consume logs as LogReader → DataFrame for statistical
analysis, so having old datasets read back with current field names out of the box
removes the last barrier. Happy to test against comma2k19 once it lands. 👍