Real-time ratings: technical architecture

Real-time rating is a bundle of services that turns millions of gaming events into updated positions in hundreds of milliseconds, while maintaining honesty, auditability and economic manageability. Below is a practical constructor from requirements to operation.

1) Requirements and targets

Latency to UI: ≤ 200-300 ms p95 from event arrival to position update.

Accuracy: correct offset of the "best N "/divisions/tie-breaks during load and failures.

Availability: ≥ 99. 9% for the tournament period; degradation without loss of integrity.

Idempotency: "exactly once" at the accrual level, "at least once" in transport.

Scale: horizontal growth by region/division/team (top-K on the ball).

Audit: unchangeable logs of rules, events, charges and issues.

2) System diagram (10,000 ft level)

Клиент → Event Ingest → Scoring Engine → Stream Aggregator (Top-K/Top-M) → Write-Optimized Store → Hot Cache/Edge → UI/Comms → Rewards/Claim

Фоново: Snapshotter, Anti-Fraud/RG, Analytics & A/B, Audit/Observability.

3) Event Ingest

Tasks: protection against duplicates, bursts, network losses.

Idempotency: each event has an 'event _ uuid'; "seen/accrued" check.

Delivery semantics: transport - at-least-once, points accounting - exactly-once.

Backpressure: queues with dynamic batches; at overload - prioritization of "active" tours.

Time windows: allow late events (for example, up to 90 seconds) - route to the "recalculation window."

Time: server UTC, on the client - only locale mapping.

4) Scoring Engine

Principles: versioned rules, anti-pay-to-win, easy rollback.

Versioning: 'rule _ set _ id' + feature-flags per-region/division.

Models: multiplier ('Score = round (100log2 (Win/Bet + 1))'), win-based with log-curve, wager-based (only as background meta).

Offset "best N ": keep the top-N structure per player/day/tour; summation/median by format.

Exceptions: bonus funds' coef _ bonus ≤ 0. 5 'or' 0'; cap per spin/day; quotas per provider.

Tie-break: 1) earlier achievement of the final speed; 2) greater diversity; 3) deterministic seed.

5) Stream aggregation and top K

The goal: to quickly recalculate positions without scanning the entire set.

Sharding: 'shard _ id = hash (division_id	region	tournament_id)`; fixed volume of top K in the shard.
Algorithms: Min-heap for top-K, Order-statistics tree or "counted B-tree"; for commands - Sum Top-M contributions.
Incrementality: update only changed positions; diff subscriptions (player-centric push).
Merging: local top-K on a shard → global top through k-way merge (regular).
Later events: "shadow" recalculation and safe update with state versioning.

6) Storage: write vs read

Write-optimized store: accrual log, best-N states, tie-breaks, metadata (record/updates only).

Read-optimized store (Query): denormalized "showcases" for UI: top 100, "my position," "top of the day/division," contribution to the team.

Snapshotter: hourly "snapshots" of the leader board; fast rollback/replay, reports and audits.

Consistency: eventual for broad lists, read-your-writes for "my position."

7) Cash and Edge

Hot cache: last positions and score of players on the shard; TTL 1-5 sec; disability by event.

Edge layer/CDN: static blocks, "cold" results pages, but not accrual logic.

Fallback: during degradation - "bronze mode": UI update every X seconds + heavy widget block.

8) Time, windows and timezones

Single "true" time: UTC on services; deadlines in UTC.

Localization - Time conversion on the client (user time zone).

Borders: "soft window" closing 1-3 min for network lags (in T&C).

We go word class: clock-skew monitoring, NTP alerts, drift measurement.

9) Dealing with late and duplicate events

Late events: keep the "reconciliation window"; recalculation locally + "calm" adjustment of positions from the notification feed.

Doubles: 'event _ uuid' + (player, game, bet, ts) in a narrow window; magazine "seen."

Accrual idempotence - Event → change transaction with applied/rolled_back status

10) Anti-fraud and RG (built into the pipeline)

Behavioral scoring: uniform intervals, super-speed, "no noise" - flags.

Connection graph: common devices/IP/payments → clusters; hold awards, manual verification.

Quotas and caps: N best attempts/day, quotas for one slot/provider, cap points per spin/day.

Bonus funds: odds/elimination.

RG: points are only awarded within voluntary limits (time/deposits/losses); break reminders.

11) Multi-region and fault tolerance

Asset-to-asset by region: regional shards + global vault; proximity routing.

RTO/RPO: RTO <15 min (folback on snapshots), RPO = 0 for accruals (idempotent log replay).

Partition-tolerance: in case of a network gap - local leader board, subsequent consolidation and soft position correction.

FX and currencies: if prizes are in different currencies, we fix the moment of conversion, FX buffer 1-3%.

12) Security and privacy

Data: minimization (pseudonyms, without PII in the aggregator), encryption "in flight" and "at rest."

Access: principle of least privileges, secrets in KMS, access audit.

T & C/jurisdictions: rule/version logs, geo-blocks, age restrictions.

13) Observability and SRE practices

SLO: UI lag ≤ 300 ms p95, Claim time ≤ 10 min p95, uptime ≥ 99. 9%.

Metrics: events/sec, proportion of takes/rejects, late-rate, top-K heap-growth, stream lag, accrual errors, claim-rate.

Tracing: ingest → scoring → cache → UI.

Alerts: stop button (disable heavy experiments/reduce N best), auto degradation.

Post-mortem: RCA for 24-48 hours, action items list, error-budget policy.

14) Rewards & Claim

Auto award: in case of a close event - calculation of positions, issuance, receipts.

Claim-window: ≤ 72 hours, retray delivery, deterministic order of payments.

Transparency: history of awards in profile: rule, time, amount, source.

Holds: for rare/valuable prizes before KYC/AML.

15) A/B experiments and circuit evolution

What to test: N best (10/20), cap per spin (on/off), diversity bonus (0/50/100), aggregation type (amount/median).

Randomization: by region/channel/cluster - avoid overflows.

Guardrails: RTP, complaints/10k, RG signals - not degraded.

Migrations: storage scheme compatibility, dual-write for the duration of the migration.

16) UX-patterns real-time ratings

Rating center: rule of victory, timer, your score, remaining attempts, next reward.

Live feedback: "+ 40 points for x3," "2 attempts out of 20 remained," "you climbed 3 places."

Showcases: overall top 100, "my position," "top of the day/division," contribution to the team.

A11y: contrast, large numbers, focus states, voicing progress.

Localization: deadlines/timers in the local TZ, multilingualism.

17) Frequent failures and degradation scenarios

Failure	Symptom	Action
Peak load on Ingest	Queue growth	Temporary reduction of N best, prioritization of active tours
Stream processor drop	Stagnating top K	Feylover of workers, replay of the log, showing a snapshot
Inflated late-rate	Mass adjustments	Enlarge the "soft window," explicitly highlight replay in UI
Event doubles	"Double glasses"	Stricter deduplication, block of suspicious sources
Cache lag	"Jumping" positions	Disability by event, limiting heavy requests

18) Checklist launch real-time ratings

Rules/Economics

Formula, top N, tiebreaker, divisions
RewardRate, payout width, wagers, claim ≤ 72 h

Equipment

Idempotent Ingest, Late-window, Sharding
Stream-Aggregator (топ-K/Top-M), Snapshotter
Hot Cache/Edge, Health/Status API, stop button

Safety/RG

Behavioral scoring, linkage graph, mouthguards/quotas
RG limits, reminders, KYC/AML for big prizes

Data/A/B

Dashboards: lag, late-rate, doubles, fairness metrics
Experimental design and guardrails, dual-write during migrations

UX

Rating center, micro-feedback, A11y, localization
Short rules in 1 screen + full T&C

A reliable real-time rating is not one service, but a disciplined ecosystem: idempotent reception, versioned scoring, top-K stream aggregation, fast cache, status snapshots and built-in anti-fraud/RG mechanisms. Add strict SLO/SLA, comprehensible UX and change verification via A/B - and you get leaderboards that players trust, and the product gets predictable growth of Retention, Stickiness and LTV.