Architecture

Request flow

Browser
  → SvelteKit page (apps/web/src/routes)
    → +page.ts loader
      → fetch /api/v1/mmr?region=…&riotId=…
        → Hono route in +server.ts
          → @mmr-calculator/db   (D1 cache check)
          → @mmr-calculator/riot-client (rate-limited fetches on miss)
          → @mmr-calculator/core (tier1 + tier2 + combine)
        ← (mu, sigma, ci90, breakdown)
    ← rendered Svelte component

The frontend never talks to Riot directly. All Riot calls flow through the rate-limited client running in the same Worker.

Packages

packages/core

Pure functions, no I/O. Trivially testable.

• tier1.ts — win-behavior estimator
• tier2.ts — lobby-composition estimator
• combine.ts — inverse-variance combination + sigma floor
• conversion.ts — tier/division/lp → MMR and the inverse mapping
• confidence.ts — CI helpers (90% / 80% / 50%)

packages/riot-client

Typed wrapper over the Riot API.

• client.ts — endpoint methods (league-v4, match-v5, summoner-v4, account-v1)
• rate-limiter.ts — token-bucket layered over the Durable Object's global counters
• regions.ts — platform/region routing tables
• schemas.ts — Zod schemas for every response we touch

packages/db

Drizzle ORM bound to D1.

• schema.ts — tables for summoners, matches, ranked_entries, mmr_snapshots
• repository.ts — query helpers (cache reads, snapshot writes)

apps/web

SvelteKit app. Routes:

• / — landing
• /summoner/[region]/[slug] — result page (loader hits /api/v1/mmr)
• /compare — multi-account compare
• /methodology — public-facing methodology page (separate from these dev docs)
• /faq
• /api/v1/* — Hono-mounted endpoints

Caching layers

Layer	TTL	Purpose
KV CACHE	minutes	Hot summoner / match-id lookups, dedup repeated requests
D1 ranked_entries	24h	Per-summoner rank, hit by Tier 2's 9-participant lookups
D1 matches	indefinite	Match payloads — they never change once finalized
D1 mmr_snapshots	indefinite	Daily snapshots for the future LP-delta tier

D1 is the load-bearing one. Tier 2's worst case is ~180 API calls; with D1 hitting on already-seen participants, repeat requests against popular players collapse to a handful.

Rate limiting

A single RateLimiterDO Durable Object holds the global token bucket. Every Riot call asks the DO for a token before going out. Configured for dev-key limits (20/1s, 100/2m) by default; production-key limits are a config swap.

The DO is bundled separately by apps/web/scripts/inject-do.mjs and stitched into the SvelteKit-generated _worker.js post-build. This is necessary because SvelteKit's adapter doesn't know about Durable Object class exports.

Cron jobs

wrangler.jsonc declares two triggers:

• 30 0 * * * — daily LP snapshot. For every summoner we've seen recently, write a row into mmr_snapshots. Feeds the future LP-delta tier.
• 0 */6 * * * — cache compactor. Drops ranked_entries older than 24h and trims matches for accounts inactive >90 days.

Both handlers live in apps/web/src/lib/server/cron.ts and are wired into the worker by the same post-build script that handles the DO.