4. Deployment and Governance

Data Products and Monitoring

Specify interfaces and monitor the whole path from source data to decision outcome

Data Products and Monitoring

Treat the model as one component

HarborMart's score is usable only if five interfaces agree:

Producer givesConsumer assumesContract test
checkout service: order_time in UTCfeature job: event time, not load timereject future or unparsable timestamps
capacity service: available picker-minutesmodel: value known at checkoutcompare event timestamp with score timestamp
model: probability in [0,1][0,1]policy: calibrated late-delivery riskrange and calibration test
policy: accept, reroute or re-promiseoperations: action before confirmationlatency and action-log test
delivery system: actual arrivalevaluator: final labellabel completeness after a fixed delay

Version the schema, feature definitions, model, policy and cost assumptions together. A score from model v4 interpreted by policy v2 is not reproducible merely because both files exist.

Five monitors, five responses

MonitorExample triggerFirst response
servicep95 latency exceeds 300 ms for 10 minfall back to the tested rule
datascheduled picker-minutes missing above 1%stop scoring affected stores
modelfour-week Brier score exceeds the baselineinspect calibration and segment error
decisionoverride rate doublesinspect workflow, incentives and incidents
outcomelate rate rises while rejection risespause rollout; reassess effect and objective

Choose thresholds from operational tolerance and historical variation. “Alert when a line looks unusual” cannot be tested.

Drift is a diagnosis, not a verdict

  • Covariate drift: P(X)P(X) changes—for example, baskets become larger before a holiday.
  • Prior shift: P(Y)P(Y) changes—for example, storms raise the late-delivery base rate.
  • Concept drift: P(YX)P(Y\mid X) changes—for example, a routing redesign changes how distance predicts lateness.
  • Policy drift: the same score produces different actions because capacity or staff behaviour changes.

A large feature shift may be harmless; a small shift in a high-cost subgroup may matter. Compare distributions, calibration and decision value, then investigate the process that generated the change.

Worked incident

On Monday, HarborMart deploys a faster routing process. By Friday:

SignalBeforeAfter
predicted late risk8.1%8.0%
eventual late rate8.0%5.6%
actions labelled after 48 hours99.4%71.0%

Two explanations fit the early dashboard: the intervention helped, or late labels are incomplete. Do not recalibrate until label delay and missingness are resolved. First freeze the evaluation cohort by decision date and wait for its defined maturity window.

Selective labels and feedback

If high-risk orders are rejected, their delivery outcomes are never observed. The deployed policy changes its own future training data. Similarly, extra staff assigned to high-risk slots can make the model appear pessimistic because the intervention prevents the predicted event.

Log at least:

  1. the information and score available at decision time;
  2. the proposed and executed action;
  3. override actor and reason;
  4. exposure or treatment actually received;
  5. outcome and its observation time.

Without this decision log, one cannot separate model error from successful intervention.

Release ladder

StageWhat changesExit evidence
replaynothing operationalhistorical pipeline reproduces known results
shadowscores logged, no action changedlatency, coverage and calibration acceptable
limited pilota small eligible unit receives the policysafety and outcome stop rules remain clear
staged rolloutexposure expands by store or timestable effects and no material subgroup harm
routine operationpolicy is defaultrecurring review, challenger and rollback tested

Keep a simple fallback that does not depend on the failing component. Test rollback before it is needed.

Current organisational example

Airbnb describes Chronon as an end-to-end feature platform supporting temporally accurate backfills, online/offline consistency and automated monitoring for data quality, training–serving skew and feature drift. The transferable lesson is architectural: a feature is a time-indexed data product with producers, consumers and controls—not a column copied into a notebook.

This is an organisation's engineering account, useful for system design but not independent evidence of business impact.

Quick check

The score distribution and AUROC are stable, but staff now override 45% of recommendations instead of 8%. Is the system healthy?

Answer
Not enough evidence says so. The model monitor is stable, but the decision interface has changed. Audit which cases are overridden, why, whether the action was feasible, and how outcomes differ. The cause may be workflow friction, a new constraint, poor trust, gaming or a model blind spot.

Next: Governance, Fairness and Privacy

Copyright © 2026