Troubleshooting: symptom -> cause -> fix¶
This page explains Troubleshooting: symptom -> cause -> fix and how it fits into the RecSys suite.
Who this is for¶
Anyone stuck. Use this as a quick lookup.
Report is empty or missing sections¶
Cause:
- wrong mode
- output path permission issue
Fix:
- verify --mode and config
- write to a writable path
"unknown schema" in validate¶
Cause:
- wrong schema name
Fix:
- use exposure.v1, outcome.v1, assignment.v1
Metrics are all zero¶
Cause:
- outcomes not joined to exposures
- event types not matching expectations
Fix:
- check request_id alignment
- inspect a few joined records
Everything looks like a win¶
Cause:
- you compared the same dataset against itself
- you sliced too much and found random wins
Fix:
- run AA-check or use a known baseline
- reduce slices and focus on primary metrics
Interleaving says A wins but A/B says B wins¶
Cause:
- interleaving measures relative ranker preference on the same traffic
- A/B includes broader effects and guardrails
Fix:
- use interleaving to choose between rankers
- use A/B to decide shipping
Read next¶
- Interpretation cheat sheet: Interpretation cheat sheet (recsys-eval)
- Runbooks: Runbooks: operating recsys-eval
- Data contracts: Data contracts: what inputs look like
- Integration: Integration: how to produce the inputs
- Interpreting results: Interpreting results: how to go from report to decision