Fraud assessment "undefined"

[bajtos]

When we evaluate measurements, the fraud assessment pipeline often assigns an undefined value to some.

[bajtos]

Data for round 528:

dry-run
  measurements_OK: 187,880
  measurements_INVALID_TASK: 9,446
  measurements_DUP_INET_GROUP: 165,958
total: 363,284

production:
  measurements_OK: 182,212
  measurements_INVALID_TASK: 9,446
  measurements_DUP_INET_GROUP: 153,823
  measurements_undefined: 17803
total: 363,284

[juliangruber]

Do you have a suspicion where _undefined is coming from? Shall we pair on this?

[bajtos]

My current thinking: There is a race condition - we receive new measurements and round-start events close together. The round evaluation starts running before we fetch & preprocess the last batch of measurements. However, these measurements are still added to our in-memory store. So later, when we collect stats, we iterate over these measurements too. I have an idea how to fix this. I hope to have a PR ready tomorrow. 🤞🏻

[bajtos]

We haven't seen "fraud assessment undefined" in the last 90 months. I believe the partial fixes we made as part of #233, most notably #224, solved the problem.