A platform backend with queue-backed workers

Shape

One service with a module per domain: players, content, marketplace orders, leaderboards, referrals, uploads and agents. Each module follows controller → service → repository, and the API is described with OpenAPI.

Keeping requests fast

Anything that waits on someone else runs in the background. Scraping social posts from a third-party provider, recording activity on-chain, migrating stored media and evaluating referrals all go onto Redis-compatible queues with dedicated background workers, so request latency does not depend on an external API. Storage migrations have a dead-letter queue, and every worker listens for a shutdown signal so a deploy drains cleanly.

Abuse controls

• A referral reward is granted at most once per signup IP per hour, using SET NX with a TTL on a SHA-256 hash of the IP. A blocked reward fails silently instead of telling an abuser what tripped.
• Referral clicks are counted with a daily-salted IP hash, which allows daily unique counts without storing raw addresses.
• Per-client rate limiting sits in front of the router.

Sign-in

Players sign in with Sign-In with Ethereum (EIP-4361). The server issues a one-time nonce, the wallet signs a message containing it, and the backend recovers the signer from the signature, consumes the nonce so it cannot be replayed, and only then issues a JWT. Wallet addresses are stored exactly as the wallet presented them, because lowercasing a checksummed address breaks lookups and integrations.

How I'd build it in Spring Boot

• Modules: package-by-feature (player, marketplace, referral, …), each with a @RestController, a service and a Spring Data MongoDB repository; springdoc-openapi generates the API description.
• Sign-in: the nonce stored in MongoDB with a TTL index; the signer recovered with web3j's Sign.signedPrefixedMessageToKey; the nonce consumed with a single MongoTemplate.remove whose deleted count must be 1, so a replay fails; then a JWT validated by Spring Security's resource server on every later request.
• Background work: Kafka or RabbitMQ listeners with retries and a dead-letter destination, or Redis-backed queues drained by worker beans implementing SmartLifecycle; server.shutdown=graceful lets in-flight work finish on deploy.
• Referral abuse check: setIfAbsent("referral:signup-ip:" + hash, "1", Duration.ofHours(1)), with the IP hashed through MessageDigest (SHA-256) so raw addresses are never stored.
• Rate limiting: a Bucket4j servlet filter keyed by client.