Arrival

An ETA prediction service, built on a feature store that cannot drift.

One feature declaration, retrieved two ways: point-in-time for training, by entity for serving, on top of Dispatch's trip telemetry. A test replays a day and asserts the two paths never disagree.

19,312 trips
107 tests
2.830 min honest lift
0.924 min leak, still measured

The finding

A leaked feature window buys 0.924 minutes of MAE that does not exist. The honest model is worth 2.830. On a quarter of the data, the same leak was worth 3.746 while the honest model was worth only 3.065. A leak flatters hardest exactly when a project is small, new, and least likely to be checked.

See the numbers →

Architecture

Where the feature store sits between an offline lane and an online lane, and why it is the only node both lanes touch.

Results

Every model's score, the outage that stays invisible for 35 minutes, and whether an A/B test could ever have seen a 2% improvement.

Handoff

Dispatch measures how wrong a flat delivery promise is. Arrival predicts what it should have been. fct_trip is the entire contract between them, diagrammed.

What each piece is doing

Named by the job it does, so the stack list is not the first thing you have to decode.

features.py One set of declarations, two retrieval paths. Training asks "as of this trip's assigned_at"; serving asks "as of now".
train.py Ridge regression in pure stdlib, chronological split, and a deliberately leaky variant scored beside the honest one.
registry.py Model artifacts versioned by content hash, so the same run is the same version and cannot be silently overwritten.
serve.py Prediction service reporting p50/p99 latency, never a mean, and flagging every fallback rather than hiding it.
monitor.py Feature and prediction drift, plus realised accuracy that only scores a prediction once its label has actually arrived.
experiment.py Hash-based A/B assignment and a minimum-detectable-effect calculation, so a measured win can be told apart from noise.

The trips are synthetic where Dispatch's own warehouse is absent, so the absolute minutes are a demonstration rather than a measurement. What is not circular is the ordering: a leak flatters a model, the flattery does not survive serving, and an accuracy dashboard cannot see an outage until its labels arrive.