| name | mobile-backend-infra |
| description | Baseline infrastructure for a mobile backend -- edge caching, feature flags, and multi-region strategy. Use when planning or auditing the deployment shape of a mobile backend. |
Mobile Backend Infra
Instructions
Mobile traffic is globally distributed, latency-sensitive, and correlated with user behavior (morning/evening peaks). The infra needs a CDN in front, feature flags everywhere, and a sober region strategy.
1. Edge Caching
Put a CDN in front of every public endpoint, even JSON APIs.
- Static/media: long
Cache-Control: public, max-age=31536000, immutable with content-hashed URLs.
- Cacheable reads (feeds, listings): short TTLs (10–60 s) with
Cache-Control: public, max-age=30, stale-while-revalidate=60 -- reduces thundering herd when a release hits.
- Authenticated but cacheable per-user:
Cache-Control: private, max-age=30; keep out of the shared CDN tier.
Key the CDN cache on Accept, Accept-Language, and the major version (/v1). Avoid caching by Authorization (cardinality explosion).
Purge by URL or by tag (Fastly surrogate keys, Cloudflare cache tags) on mutation events:
await cdn.purgeByTag(`article:${id}`);
2. Feature Flags
Every user-visible feature ships behind a flag. Flags are evaluated at two layers:
- Server-side (authoritative): mobile clients receive the computed flag value in
GET /v1/config or in session tokens.
- Client-side (optional): the flag SDK re-evaluates locally for offline usage; the server value wins on reconnect.
GET /v1/config
{
"flags": {
"home_v2": true,
"checkout_stripe_intent": true,
"ai_suggestions": { "enabled": true, "model": "v3" }
},
"refresh_after": 1713522660
}
Principles:
- Kill switch per critical path. A bad flag must roll back in < 60 s globally.
- Ramp gradually: 1% -> 10% -> 50% -> 100%, with automatic guardrails (error rate, p95).
- Flag values are cached with short TTL on device (e.g., 5 min) so a mid-session switch does not break UX.
Vendors: LaunchDarkly, Statsig, ConfigCat; or self-host with a simple config service + pubsub.
3. Region Strategy
Three common postures:
A. Single region. Cheapest and simplest. Acceptable for early-stage apps with latency-tolerant use cases (< 300 ms global TTFB is achievable with edge caching + HTTP/3).
B. Multi-region read, single-region write. Reads served from the nearest region, writes forwarded to the primary. Fits most apps:
[us-east-1] (primary writes) <---replication--- [eu-west-1] [ap-south-1]
^ ^ ^
+----- mobile write ----+ | |
| v v
BFF/edge -----------> nearest read region
Replication: Postgres logical replication, Aurora Global, Cloud Spanner, DynamoDB global tables -- pick the one that fits your datastore.
C. Active-active multi-region. Complex. Needed for regulatory locality (EU data stays in EU) or low-latency writes globally. Budget for conflict resolution, global id generation (ULID / Snowflake), and a year of operational learning.
4. Pod and Cluster Sizing
- API pods: CPU-bound usually. Start at 2 vCPU / 2 GiB; scale out horizontally. HPA on CPU + request-per-second.
- Realtime nodes: memory-bound. See
realtime-scaling.
- Background workers: separate cluster / nodepool; noisy neighbors are common.
- Use a PodDisruptionBudget to keep at least (N-1) pods up during rollouts.
5. Networking
- TLS 1.2+ everywhere; TLS 1.3 where possible.
- HTTP/2 at the LB; HTTP/3 where the CDN supports it -- big gains on lossy cellular networks.
- Timeouts set at every layer: LB > ingress > pod > downstream; each smaller than the one above so upstream can retry intelligently.
6. Observability Stack
- Tracing: OpenTelemetry; propagate
traceparent from the mobile SDK through the stack.
- Logs: structured JSON with
request_id, user_id (hashed), endpoint, latency_ms, status, error.code.
- Metrics: RED (rate, errors, duration) per endpoint; saturation for pools and queues.
- SLOs: e.g., reads p95 < 200 ms from mobile edge, error rate < 0.5%. Page on SLO burn rate, not on every error.
7. Secrets and Config
- Secrets in a manager (AWS Secrets Manager, GCP Secret Manager, Vault). No secrets in env vars baked into images.
- Config separated from code; feature flags above; environment-specific values in a config service.
8. Cost Controls
Mobile backends easily spend > 50% of budget on egress and object storage. Watch:
- Egress by endpoint / region; cache miss storms on deploy day.
- Object storage lifecycle rules: transition originals to cheaper tiers after 90 days.
- Realtime fanout growth: linear in connections, quadratic if you misuse channels.
9. Incident Readiness
- Runbooks for the top 20 alerts.
- Feature-flag kill switches for every risky path.
- Regional failover drills quarterly.
- A rollback-first culture: the mobile app pins a minimum supported backend version; breaking a contract means roll forward a fix, not "force update".
10. Security Baseline
- WAF at the edge (managed rule set + your own for auth endpoints).
- Bot detection for public endpoints (login, signup, forgot-password).
- Attestation (App Attest, Play Integrity) bumps trust for critical flows.
- Regular dependency and image scans; patch cadence documented.
Checklist
