Current Stack Architecture

FTL runs on Go microservices hosted on Azure Container Apps, backed by PostgreSQL Flex and Azure Cache for Redis. The stack is sized for 50K concurrent users across a 39-day tournament (June 11 – July 19, 2026).

---

Language and HTTP Framework

Go 1.22 + Fiber v2

Fiber wraps fasthttp and avoids the net/http allocations that add latency at high concurrency. Every service (api-server, ws-server, flusher) is a standalone Go binary.

Decision: Go over Node.js or Python — statically compiled binary, goroutine-per-connection model handles 50K WebSocket connections with predictable memory; no GIL, no runtime startup overhead, no npm supply chain.

Entry points: - ftl-backend/cmd/api-server/main.go — HTTP API - ftl-backend/cmd/ws-server/main.go — WebSocket server - ftl-backend/cmd/flusher/main.go — background flusher

The Fiber app is configured with ProxyHeader: "X-Forwarded-Proto" so HSTS activates correctly behind Container Apps' TLS-terminating proxy.

---

Database

Azure Database for PostgreSQL Flexible Server 16

Setting	Value
SKU	D2ds_v5 (2 vCPU, 8 GB RAM)
Storage	256 GiB Premium SSD
Pooler	PgBouncer, transaction mode, port 6432
Extensions	pg_stat_statements, pg_cron, pg_partman, pgcrypto
HA	Disabled
Backup	LRS, 7-day retention

Decision: PostgreSQL Flex over Cosmos DB — standard SQL, cheaper for write-heavy workloads, no per-RU pricing, familiar tooling, pg_cron for periodic tasks.

The api-server and flusher connect through pgxpool (Go's idiomatic pgx pool). The api-server pool is capped at 20 connections; PgBouncer multiplexes those to the DB server (default pool size 50, max client connections 2000).

---

Cache and Trade Execution

Azure Cache for Redis — Standard C1 (1 GB)

Setting	Value
Ops/sec capacity	40K ops/sec
Working set	~180 MB (18% of 1 GB)
Persistence	Disabled — Postgres is the source of truth
maxmemory-policy	allkeys-lru

Decision: Redis Lua scripts over Azure Service Bus — single-threaded atomic execution eliminates distributed locks. One Lua script call replaces what would require a Service Bus enqueue + worker dequeue + optimistic retry cycle. Latency drops from ~50ms (Service Bus round trip) to <1ms (in-process Lua).

The four Lua scripts live at: - ftl-backend/internal/redis/lua/trade_execute.lua — legacy BUY/SELL - ftl-backend/internal/redis/lua/position_open.lua — CFD open (ADR-0004) - ftl-backend/internal/redis/lua/position_close.lua — CFD close (ADR-0004/ADR-0005/ADR-0006) - ftl-backend/internal/redis/lua/ping.lua — health check

Each script runs in Redis's single-threaded interpreter. No two scripts execute concurrently for the same Redis instance, giving atomic read-modify-write without any application-level locking.

---

Container Hosting

Azure Container Apps — Consumption Plan, Central India

Service	Replicas (peak)	vCPU	GiB	Role
ws-server	3 (min 2)	2	4	WebSocket connections + fan-out
api-server	2 (min 2)	1	2	HTTP API, Sportmonks poller
flusher	1 (never scales)	0.5	1	Redis → Postgres sync
Peak total	6	8.5	17

maxReplicas hard limits: ws-server = 5, api-server = 5, flusher = 1. Without these limits, auto-scale could reach $5,000+/mo.

ws-server requires two non-default ingress settings:

ingress:
  transport: auto        # NOT http2 — WebSocket uses HTTP/1.1 upgrade
  stickySessions:
    affinity: sticky     # pins a client to the same replica after WS upgrade

Decision: custom Go WebSocket server on Container Apps over Azure Web PubSub — Web PubSub costs ~$1,200/mo for 50K connections. Three Container Apps replicas at $30/mo each total ~$90/mo: roughly 15x cheaper.

ws-server must run in Single revision mode for sticky sessions to work. api-server and flusher use standard HTTP and can run in multi-revision mode for blue/green deploys.

Warning

Never run scale-up.sh or scale-down.sh against a Container App in Multiple revision mode. That configuration caused a ₹1,234/day overrun in staging (May 2026) by accumulating 9 warm revisions. Always verify activeRevisionsMode == "Single" first.

---

Authentication

Raw Google OAuth 2.0

The api-server validates Google ID tokens directly against Google's JWKS endpoint. auth.NewGoogleVerifier fetches keys on startup and refreshes them in the background via googleVerifier.StartRefresh(ctx).

Post-validation, the server issues its own JWT signed with a local RSA key pair (JWTPrivateKey / JWTPublicKey from Key Vault), using RS256. Three token types exist: access, refresh, and a short-lived ws-token (60-second TTL) used for the WebSocket upgrade handshake.

Decision: raw Google OAuth over Auth0 or Firebase Auth — $0 MAU cost at any scale. Firebase Auth charges per MAU beyond 10K users. At 50K tournament users, Firebase would add ~$150/mo; Auth0 would add ~$800/mo.

Token generation: ftl-backend/scripts/gen-dev-keys.sh.

---

Cost Breakdown

Verified via Azure Retail Prices API, April 2026, Central India region.

Service	Monthly (USD)	Monthly (INR)
Container Apps (6 replicas, active billing)	$455	₹37,993
Bandwidth (3,100 GB)	$357	₹29,810
PostgreSQL Flex D2ds_v5 + 256 GiB	$217	₹18,120
Sportmonks API	$75	₹6,263
Redis Standard C1	$50	₹4,175
Azure Communication Services (email)	$25	₹2,088
Static Web Apps Standard	$9	₹752
Container Registry Basic	$5	₹418
Key Vault Standard	$2	₹167
Application Insights (1% sampling)	$0	₹0
Cloudflare Free	$0	₹0
Total	~$1,200	~₹1,00,200

Tournament total (39 days): ~$1,560 / ₹1,30,260.

---

Deployment Topology

graph TB
    subgraph Clients
        Browser["React SPA\n(Static Web Apps)"]
    end

    subgraph Edge
        CF["Cloudflare Free\nCDN + DDoS + WAF"]
    end

    subgraph "Azure Container Apps Environment"
        WS["ws-server\n3 replicas × 2vCPU/4GiB"]
        API["api-server\n2 replicas × 1vCPU/2GiB"]
        FL["flusher\n1 replica × 0.5vCPU/1GiB"]
    end

    subgraph "Data Layer"
        PG["PostgreSQL Flex\nD2ds_v5 / 256 GiB"]
        RD["Redis Standard C1\n1 GB / Lua + pub/sub"]
    end

    subgraph "External"
        SM["Sportmonks API\n€69/mo"]
        GO["Google OAuth 2.0\n$0"]
        EM["Azure Communication Services\n~$25/mo"]
    end

    subgraph "Operations"
        KV["Key Vault\n10 secrets"]
        CR["Container Registry\nBasic"]
        AI["Application Insights\n1% sampling"]
    end

    Browser -->|HTTPS| CF
    CF -->|api.ftl2026.com| API
    CF -->|wss://api.ftl2026.com sticky| WS

    API -->|pgxpool 20 conns| PG
    API -->|go-redis 20 conns| RD
    API -->|poll every 2s| SM
    API -->|validate token| GO
    API -->|queue email| EM

    WS -->|go-redis subscribe| RD

    FL -->|batch writes| PG
    FL -->|read dirty sets| RD

    KV -.->|secrets at startup| API
    KV -.->|secrets at startup| WS
    KV -.->|secrets at startup| FL
    CR -.->|image pull| API
    CR -.->|image pull| WS
    CR -.->|image pull| FL
    API -->|telemetry| AI
    WS -->|telemetry| AI

---

Other Services

Service	Role
Cloudflare Free	CDN for static assets (98%+ cache hit), DDoS, WAF, TLS termination. Enforces 100s WebSocket idle timeout — clients must send a heartbeat every 30s.
Azure Communication Services	Registration confirmation and milestone notification emails (~$25/mo). Falls back to NoopSender when COMMUNICATION_SERVICE_CONNECTION_STRING is unset so dev environments boot without secrets.
Sportmonks API	FIFA player data polled every 2 seconds during live matches. A circuit breaker opens after 5 consecutive failures; during the open window, the api-server uses the last-good base price with a DEGRADED flag.
Application Insights	Telemetry at 1% trace sampling to stay under the 5 GB free tier.

---