MLOps: operating models

1) The role of exploitation in iGaming

In iGaming, models affect real money and regulation: RG interventions, anti-fraud, payments, KYC, limits, offers and recommendations. Operation is a reliable presentation of predictions with guaranteed SLO, traceability and safety.

• Predictable releases and rollbacks without downtime.
• Data consistency and offline/online features.
• Observability: quality, drift, honesty, privacy.
• TCO reduction: performance, cache, GPU/CPU mixes.
• Compliance (audit/DSAR/Legal Hold/ethics).

2) Serving architectures

Batch (offline): night/hourly scoring (limits, segments). Pros: cheaper, more stable. Cons: No instant reaction.
Stream (near-real-time): event processing (bets, anomalies) with 1-5 min windows.
Online (sync API): <100-300 ms p95 for UX/risk solutions, caching and degradation.
Hybrid: "baseline from batch + online refinement" (example: 7-day RG risk + online session triggers).

• Ensemble/Stacking with a light "gate model" on a critical path.
• Fallback heuristics in case of model failure/feature.
• Circuit Breaker and rate limiting on peaks or when providers degrade.

3) Model Registry and Version Management

Model Registry: versions, owners, release date, metrics (AUC/PR, calibration), dataset_version, feature_set_version, usage restrictions.
Model Card: task, data/features, fairness/privacy section, risk zones, review frequency.
Release policy: 'MAJOR. MINOR. PATCH '+ mandatory rollback plan.
Champion-Challenger: parallel challenger run with reports; automatic promotion when criteria are met.

4) Online features and consistency

Feature Store: offline (training) and online (inference) showcases with strict contracts.
Time travel and point-in-time join in training.
Idempotent updates feature and protection against target leakage.
Consistency: read-your-writes or SLA delivery guarantees (for example, ≤ 60 seconds).
Feature policy: allow/deny lists, masking, tokenization, proxy PII prohibition.

5) Release strategies

Shadow: all load → champion; challenger receives a copy of the requests, answers do not affect the business.
Canary: 1-10% of traffic → new version; comparison of KPI/metrics, auto-rollback by thresholds.
Blue-Green: two server/endpoint pools; DNS/route switching.
Flags: fine tuning by markets/tenants/channels.

6) Observability and alerting

• Reliability: error rate, timeouts, p50/p95/p99 latency, QPS, saturation.
• Data/features: freshness, completeness, distributions, anomalies, omissions, schema drift.
• Quality: calibration, post-fact metrics (AUC/PR, uplift), intervention response.
• Drift: at inputs (PSI/KS) and at outputs (score drift).
• Ethics/fairness: EO/EOp-deltas, disparate impact.
• Privacy: Attack-AUC (membership/inversion) ≈ 0. 5, ε -usage (if DP).
• Business: chargeback, RG interventions, conversion of offers - segmented.

• p95 latency ≤ 200 ms (online scoring RG/anti-fraud).
• Error rate ≤ 0. 1% 5-min. mean.
• Drift PSI ≤ 0. 2 in key features; EOp-delta ≤ 3 p.p.
• Freshness feature ≤ 60 sec; gaps ≤ 0. 5%.
• Calibration of ACE ≤ 0. 02.

7) Incidents and playbooks

Sev levels: P1 (payout blocking/RG error), P2 (error growth> threshold), P3 (quality degradation).
Auto-mitigation: switching to champion, lowering the frequency of requests, enabling fallback rules, isolating "toxic" features.

Runbooks: checklists for "feature are outdated," "drift has grown," "feed typing has changed," "GPU is exhausted."

Post mortem: RCA, fix plan, update tests/thresholds/contracts.

8) Experimentation and change control

A/B and multi-armed bandit - only stratified by key groups (country/channel/device).
Ethical stop rules: with a sharp increase in RG risk/complaints.
Dual-run showcase features and models before switching.
Versioning of KPIs and definitions (BI contract) for stable interpretation of results.

9) Security and privacy in sales

mTLS/TLS 1. 3, request signature, anti-replay (nonce/idempotency).
Secrets from Secrets Manager, JIT issuance, audit.
Tokenization of inputs/logs; PII inhibition in trails.
TEE/Confidential VIP Payment Inference/AML (if required).
Access policies (RBAC/ABAC/JIT) to features and endpoints.
DSAR/Legal Hold: A trace of solutions for explainability and deletion by token.

10) Performance and cost

Cache (feature/score) with TTL, especially for stable signals.
Quantization/distillation for acceleration (INT8/FP16).
Autoscaling: horizontal by QPS/latency, vertical by batch-size.
CPU/GPU hybrid: latency-critical on GPU, "mass" on CPU.
Cold start tracing, model heating.
Model pool and "sticky routing" by market/tenant for cache locality.

11) iGaming cases (references)

RG scoring: online scoring at entry and in sessions; strict overrides (self-exclusion), the target metric is EOp + calibration.
Anti-fraud/payments: pre-authorization solutions <150 ms; FPR EO control, robust signal aggregators.
KYC/AML: thin-file support; PSI/MPC with partner; DSAR compatibility.
Personalization: uplift models and frequency limits; exclusion of high-risk from aggressive offers.

12) Metrics and SLO of operation (example)

13) Artifact patterns

13. 1 Release Notes

Model: 'rg _ risk @ 2. 1. 0` (MINOR)

Changes: added feature 'loss _ streak _ 7d'; calibration updated

Validation: shadow 14 days; delta KPI ≤ 0. 3%; EOp delta normal

Rollout: canary 10% EU → 50% → 100%

Rollback: flag'rg. use_v1=true`

Owner/Date/Ticket

13. 2 Model card (fragment)

Task: anti-fraud payments

Data: 'payments _ gold v3. 2 ', feature set' payout _ signals v1. 7`

Metrics: AUC = 0. 89, ACE=0. 015, FPR @ operas. threshold = 1. 2%

Fairness: EO TPR/FPR Δ ≤ 2 п.п. по «country/method»

Restrictions: VIP clients - with human-review only

Privacy: TEE-inference; logging without PII

Review: Once every 90 days

13. 3 Endpoint SLO policy (snippet)

yaml endpoint: /v1/score/rg slo:
latency_p95_ms: 200 success_rate: 0. 995 max_error_burst_per_5m: 50 data:
feature_freshness_s: 60 allowed_missing_pct: 0. 5 ethics:
eop_delta_pp: 3 privacy:
attack_auc_max: 0. 55

13. 4 Runbook "Features Out of Date"

1. Check the lag in the Feature Store and the source of the feed.
2. Switch to spare channel/cache.
3. Reduce traffic/enable fallback rules.
4. Communication in # ml-status; incident P2/P1 by SLA.
5. RCA and contract/retray edits.

14) Pre-release testing processes

Contracts feature: schema/enum/nullable, SLA freshness.
Data: DQ tests, point-in-time, target leak.
Model: unit/integration, calibration, stress/load.
Security: secrets, mTLS, Zero-PII in the logs.
Ethics/privacy: fairness-check, attack-suite.
Observability: dashboards/alerts, SLO configs.
Documentation: Release Notes + rollback-plan.

15) RACI (example)

ML Lead (A/R): quality, releases, metrics.
Data Platform (R): Feature Store, register, orchestration, observability.
Domain Owners (R): source contracts/feature.
Security/DPO (A/R): access, privacy, tokenization, TEE.
SRE/SecOps (R): Incidents, SLO, Autoscale, SOAR.
Analytics/Finance (C): impact on KPIs and reports.
Support/RG/Risk (C): human-in-the-loop and explainability.

16) Implementation Roadmap

0-30 days (MVP)

1. Model Registry + cards for high-impact models (RG/payout/anti-fraud).
2. Basic monitoring: latency, errors, freshness, drift inputs.
3. Shadow runs of new versions, canary contours.
4. Contracts feature and Zero-PII in the logs.
5. Runbooks and # ml-status channel.

30-90 days

1. Champion-Challenger and auto-promotion by criteria.
2. Fairness/privacy gates in CI/CD, attack-suite.
3. Caching, quantization, autoscale; SLO/cost budget.
4. BI/ML coordination of KPIs and online metrics; dashboards SLO.

3-6 months

1. Regular post-mortems, quarterly model reviews.
2. Geo/tenant isolation of endpoints, keys and features.
3. TEE/MPC for private payout inference/AML.
4. Full automation of Release Notes from lineage and diff.
5. External audit of processes (where required by license).

17) Anti-patterns

Release without shadow/canary and rollback plan.
Inconsistent offline/online features → degradation.
Logs with PII, absence of token-policy.
"Eternal" thresholds without revision; ignoring drift and calibration.
Lack of human-in-the-loop for high-risk solutions.
Experiments without stratification and ethical stop rules.

18) Related Sections

DataOps Practices, Access Control, Data Tokenization, Security and Encryption, Auditing and Versioning, Bias Mitigation, Confidential ML, Federated Learning, Data Retention Policies, Data Origin and Path, Data Ethics.

Total

Model exploitation is an engineering discipline at the production service level: clear contracts and versions, predictable releases, observability 24/7, manageable ethics/privacy risks, and transparent business impact. This makes ML a reliable product, not "the best script in a laptop."