Skip to content

Building 70 Projects ranging from beginner to advanced so anyone can β€” learn from, build upon, use as a reference, or even copy directly. Owned by my cybersecurity company πŸ‘‡

33.8 KB
Raw Blame

System Architecture

This document breaks down how the system is designed and why certain architectural decisions were made.

High Level Architecture

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                        FastAPI App                          β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                        β”‚
                        β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                 RateLimitMiddleware                         β”‚
β”‚            (ASGI request interceptor)                       β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                        β”‚
                        β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                    RateLimiter                              β”‚
β”‚  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”     β”‚
β”‚  β”‚  Algorithm   β”‚  β”‚Fingerprinter β”‚  β”‚   Storage    β”‚     β”‚
β”‚  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜     β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                        β”‚
                        β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                  LayeredDefense                             β”‚
β”‚  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”           β”‚
β”‚  β”‚Layer 1:  β”‚β†’ β”‚Layer 2:  β”‚β†’ β”‚Layer 3:        β”‚           β”‚
β”‚  β”‚Per-User  β”‚  β”‚Endpoint  β”‚  β”‚Circuit Breaker β”‚           β”‚
β”‚  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜           β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                        β”‚
                        β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚              Storage Backend (Redis/Memory)                 β”‚
β”‚         Atomic counters + Lua scripts                       β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Component Breakdown

RateLimitMiddleware (src/fastapi_420/middleware.py)

  • Purpose: Intercept every HTTP request before it reaches route handlers
  • Responsibilities: Apply default rate limits, exclude health check endpoints, add rate limit headers to responses
  • Interfaces: ASGI middleware protocol, receives Request and call_next, returns Response

RateLimiter (src/fastapi_420/limiter.py)

  • Purpose: Core orchestrator that coordinates all rate limiting logic
  • Responsibilities: Initialize storage and algorithms, extract fingerprints, check limits, handle errors with fail-open behavior
  • Interfaces: Can be used as decorator @limiter.limit(), dependency injection Depends(RateLimitDep), or called directly await limiter.check()

CompositeFingerprinter (src/fastapi_420/fingerprinting/composite.py)

  • Purpose: Identify clients reliably across multiple attributes
  • Responsibilities: Combine IP, User-Agent, auth tokens, TLS fingerprints into unique identifier
  • Interfaces: async extract(request) -> FingerprintData, configurable via FingerprintLevel presets

LayeredDefense (src/fastapi_420/defense/layers.py)

  • Purpose: Three-layer protection against different attack types
  • Responsibilities: Check per-user limits, per-endpoint limits, global circuit breaker in sequence
  • Interfaces: async check_all_layers() returns RateLimitResult or raises EnhanceYourCalm

Storage (Redis: src/fastapi_420/storage/redis_backend.py, Memory: src/fastapi_420/storage/memory.py)

  • Purpose: Atomic counter operations with window/bucket state management
  • Responsibilities: Increment counters atomically, get current state, handle expiration, health checks
  • Interfaces: Protocol defined in src/fastapi_420/types.py:371-429 with methods like increment(), consume_token(), get_window_state()

Data Flow

Primary Use Case Flow: Request Rate Limiting

Step by step walkthrough of what happens when a request hits the API:

1. Request arrives β†’ ASGI Middleware (middleware.py:65)
   FastAPI receives request, passes to RateLimitMiddleware
   Middleware checks if path is excluded (health endpoints)

2. Middleware β†’ RateLimiter.check() (limiter.py:181)
   Determines rate limit rules for this endpoint
   Calls limiter with request object and rules

3. RateLimiter β†’ Fingerprinter (limiter.py:198)
   Extracts client fingerprint from request
   Returns FingerprintData with IP, User-Agent, auth, etc.

4. RateLimiter β†’ LayeredDefense (if using defense system)
   OR directly to Algorithm.check() for simple cases
   Passes fingerprint and rules to defense layers

5. LayeredDefense β†’ Storage (layers.py:89-143)
   Layer 1: Check per-user limit (key: user:endpoint:window)
   Layer 2: Check per-endpoint limit (key: endpoint:global:window)
   Layer 3: Check circuit breaker (global counter)

6. Storage β†’ Redis/Memory Lua Script (lua/sliding_window.lua:1-32)
   Execute atomic increment-and-check operation
   Return {allowed, remaining, reset_after}

7. Result propagates back β†’ Middleware (middleware.py:91)
   If allowed: Add RateLimit-* headers, pass to route handler
   If denied: Return HTTP 420 with Retry-After header

Example with code references:

1. Request β†’ middleware.py:65-75
   async def dispatch(self, request, call_next):
       if not await self._should_limit(request):
           return await call_next(request)

2. Middleware β†’ limiter.py:181-199
   result = await self.limiter.check(
       request, limit, key_func=self.key_func, raise_on_limit=False
   )

3. Fingerprinter β†’ fingerprinting/composite.py:96-125
   fingerprint = await self._fingerprinter.extract(request)

4. Algorithm β†’ algorithms/sliding_window.py:31-43
   result = await storage.increment(
       key=key, window_seconds=rule.window_seconds, limit=rule.requests
   )

5. Storage β†’ storage/redis_backend.py:236-258
   result = await self._execute_script(
       "sliding_window", keys=[key], args=[window_seconds, limit, now]
   )

6. Response β†’ middleware.py:91-101
   if not result.allowed:
       exc = EnhanceYourCalm(result=result)
       return self._create_420_response(exc)

Circuit Breaker Flow

When the system detects abnormal load:

1. Every request β†’ CircuitBreaker.record_request() (circuit_breaker.py:57)
   Increment global counter: circuit:global:requests:{window}

2. CircuitBreaker.check() β†’ Get request count (circuit_breaker.py:45)
   Calculate total requests in current window
   Compare to threshold (default: 10,000/minute)

3. If threshold exceeded β†’ Trip circuit (circuit_breaker.py:76)
   Set is_open = True
   Record failure_time
   Log warning with request count

4. Subsequent requests β†’ Check bypass rules (layers.py:232-252)
   Mode: ADAPTIVE β†’ Allow authenticated users
   Mode: LOCKDOWN β†’ Block almost everything
   Mode: DISABLED β†’ No circuit breaker

5. After recovery_time (default: 30s) β†’ Half-open state (circuit_breaker.py:90)
   Allow limited traffic to test recovery
   If successful, close circuit
   If still overloaded, re-open

6. Circuit closes β†’ Normal operation resumes
   Reset failure_count, clear is_open flag

Design Patterns

Factory Pattern for Algorithms and Storage

What it is: Create objects without specifying exact class. Client code asks for "sliding window algorithm" and factory returns the correct instance.

Where we use it:

  • Algorithm factory: src/fastapi_420/algorithms/__init__.py:11-28
  • Storage factory: src/fastapi_420/storage/__init__.py:16-21

Why we chose it: Runtime configuration. Users set ALGORITHM=sliding_window in environment variables. The factory picks the right class at startup rather than compile time. Makes testing easier too, you can mock the factory to return test implementations.

Trade-offs:

  • Pros: Loose coupling, easy to add new algorithms without changing existing code
  • Cons: Extra indirection layer, can make stack traces harder to follow

Example implementation from src/fastapi_420/algorithms/__init__.py:11-28:

def create_algorithm(algorithm_type: Algorithm) -> BaseAlgorithm:
    algorithm_map: dict[Algorithm, type[BaseAlgorithm]] = {
        Algorithm.SLIDING_WINDOW: SlidingWindowAlgorithm,
        Algorithm.TOKEN_BUCKET: TokenBucketAlgorithm,
        Algorithm.FIXED_WINDOW: FixedWindowAlgorithm,
        Algorithm.LEAKY_BUCKET: SlidingWindowAlgorithm,  # Alias
    }
    
    algorithm_class = algorithm_map.get(
        algorithm_type, SlidingWindowAlgorithm  # Safe default
    )
    return algorithm_class()

Strategy Pattern for Algorithms

What it is: Define a family of interchangeable algorithms. All implement the same interface but with different behavior.

Where we use it: All algorithms inherit from BaseAlgorithm (src/fastapi_420/algorithms/base.py:17-49) and implement check() and get_current_usage().

Why we chose it: Allows swapping rate limiting algorithms at runtime without changing the RateLimiter code. Want to switch from sliding window to token bucket? Just change one config value.

Trade-offs:

  • Pros: Clean separation of concerns, algorithms are independently testable
  • Cons: Can't optimize for algorithm-specific features, must fit common interface

Dependency Injection for Storage and Settings

What it is: Instead of classes creating their dependencies (storage = RedisStorage()), they receive them as constructor arguments.

Where we use it: RateLimiter.__init__(settings, storage) at src/fastapi_420/limiter.py:74-80

Why we chose it: Testing and flexibility. In production, inject RedisStorage. In tests, inject MemoryStorage. In edge cases, inject a mock. The RateLimiter doesn't care what storage implementation it gets as long as it implements the protocol.

Trade-offs:

  • Pros: Testability, flexibility, explicit dependencies
  • Cons: More verbose initialization, dependency management complexity

Example from src/fastapi_420/limiter.py:74-80:

def __init__(
    self,
    settings: RateLimiterSettings | None = None,
    storage: Storage | None = None,
) -> None:
    self._settings = settings or get_settings()
    self._storage = storage  # Injected, not created

Layer Separation

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚    Layer 1: Application (FastAPI Routes)              β”‚
β”‚    - Defines endpoints and business logic             β”‚
β”‚    - Doesn't know about rate limiting internals      β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                         ↓
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚    Layer 2: Rate Limiting Logic                       β”‚
β”‚    - RateLimiter, LayeredDefense, Algorithms          β”‚
β”‚    - Doesn't know about HTTP details                  β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                         ↓
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚    Layer 3: Storage Abstraction                       β”‚
β”‚    - Storage protocol, Redis/Memory implementations   β”‚
β”‚    - Doesn't know about rate limiting concepts        β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Why Layers?

Separation allows independent evolution. You can:

  • Swap storage backends without touching rate limit logic
  • Change algorithms without modifying HTTP handling
  • Test each layer in isolation

In 2019, GitHub migrated from MySQL to a custom storage system for rate limiting. Because their rate limit logic was separate from storage, the migration took weeks instead of months.

What Lives Where

Layer 1 (Application):

  • Files: examples/app.py, user route handlers
  • Imports: FastAPI, depends on Layer 2 via middleware or dependencies
  • Forbidden: Direct storage access, algorithm selection

Layer 2 (Rate Limiting Logic):

  • Files: limiter.py, algorithms/, defense/, fingerprinting/
  • Imports: Storage protocol, Pydantic models, async utilities
  • Forbidden: HTTP-specific code (Request/Response), storage implementation details

Layer 3 (Storage):

  • Files: storage/memory.py, storage/redis_backend.py, storage/lua/*.lua
  • Imports: Redis client, asyncio, dataclasses
  • Forbidden: Rate limiting concepts (what a "limit" or "window" means), HTTP details

Data Models

RateLimitRule

From src/fastapi_420/types.py:96-155:

@dataclass(frozen=True, slots=True)
class RateLimitRule:
    requests: int
    window_seconds: int

Fields explained:

  • requests: Maximum number of requests allowed in the window. Must be positive integer. Setting this too low (like 1/minute) makes APIs unusable. Too high defeats the purpose.
  • window_seconds: Time window in seconds. Common values: 1 (per second), 60 (per minute), 3600 (per hour), 86400 (per day). Must be positive.

Relationships:

  • Multiple rules can apply to one endpoint (example: "100/minute AND 1000/hour")
  • Rules are parsed from strings like "100/minute" via RateLimitRule.parse() at line 119-149
  • Algorithms use the window_seconds to calculate which time bucket to check

FingerprintData

From src/fastapi_420/types.py:158-202:

@dataclass(slots=True)
class FingerprintData:
    ip: str
    ip_normalized: str
    user_agent: str | None = None
    accept_language: str | None = None
    accept_encoding: str | None = None
    headers_hash: str | None = None
    auth_identifier: str | None = None
    tls_fingerprint: str | None = None
    geo_asn: str | None = None

Fields explained:

  • ip: Raw IP address from request, as seen by the server (might be proxy IP)
  • ip_normalized: Processed IP for rate limiting. For IPv6, this is the /64 network prefix. For IPv4, usually same as raw IP.
  • user_agent: Browser identification string. Used for fingerprinting but not alone (easily spoofed).
  • auth_identifier: User ID from JWT token, API key, or session cookie. Most reliable identifier when present, hashed by default for privacy.
  • headers_hash: SHA256 hash of header ordering and values. Browsers send headers in consistent order, bots often don't. 16-character hex string.
  • tls_fingerprint: JA3 hash of TLS handshake parameters. Requires proxy to populate X-JA3-Fingerprint header.

Relationships:

  • Produced by CompositeFingerprinter.extract() at src/fastapi_420/fingerprinting/composite.py:96-163
  • Converted to rate limit key via to_composite_key() at lines 186-202
  • Different FingerprintLevels include different fields in the key

RateLimitResult

From src/fastapi_420/types.py:66-92:

@dataclass(frozen=True, slots=True)
class RateLimitResult:
    allowed: bool
    limit: int
    remaining: int
    reset_after: float
    retry_after: float | None = None

Why frozen and slots: Results are immutable once created (frozen=True prevents modification). Slots reduce memory overhead by avoiding __dict__ attribute, important when handling thousands of requests per second.

Security Architecture

Threat Model

What we're protecting against:

  1. Brute force authentication attacks - Attacker tries millions of passwords on /login. Rate limit of 3-5/minute blocks this completely.
  2. API scraping and data harvesting - Competitor tries to download your entire product catalog. Rate limiting prevents bulk extraction.
  3. Resource exhaustion DoS - Attacker floods expensive endpoints (ML inference, report generation) to consume CPU/memory.

What we're NOT protecting against (out of scope):

  • Network layer DDoS - Use Cloudflare or AWS Shield for volumetric attacks (100+ Gbps). Application-layer rate limiting can't handle this volume.
  • Sophisticated bot farms - If attackers control thousands of residential IPs with real browsers, rate limiting alone won't stop them. Need CAPTCHA or behavioral analysis.
  • Internal threats - Authenticated users with valid credentials who abuse APIs. Requires different monitoring and response.

Defense Layers

From src/fastapi_420/defense/layers.py:47-84:

Layer 1: Per-User Per-Endpoint (most specific)
    ↓
  Checks: user_abc:POST:/api/upload:60s
  Purpose: Stop individual user abuse
  Example: User makes 100 requests/sec to upload endpoint
    ↓
Layer 2: Per-Endpoint Global (endpoint protection)
    ↓
  Checks: global:POST:/api/upload:60s
  Purpose: Prevent endpoint overload from distributed sources
  Example: 1000 different users each making 10 requests/sec
    ↓
Layer 3: Circuit Breaker (DDoS protection)
    ↓
  Checks: circuit:global:requests:60s
  Purpose: Protect entire API when under massive attack
  Example: 10 million requests/minute from botnet

Why multiple layers? Each layer addresses different attack patterns. In 2018, Fortnite's login servers went down despite rate limiting because they only had per-user limits. When 10 million players tried to log in simultaneously (legitimate traffic spike), the aggregate exceeded capacity. Per-endpoint global limits would have throttled the traffic to sustainable levels.

Storage Strategy

Memory Storage (Single Instance)

What we store:

  • Window counters: Dict[str, WindowEntry] where key is "ratelimit:v1:user:endpoint:identifier:window_id"
  • Token bucket states: Dict[str, TokenBucketState] keyed by identifier
  • Stored in OrderedDict for LRU eviction when max_keys exceeded

Why this storage: Simplicity and speed. In-memory access is microseconds vs Redis milliseconds. Perfect for development, testing, or single-instance APIs that don't need distributed state.

Schema design from src/fastapi_420/storage/memory.py:18-25:

@dataclass
class WindowEntry:
    count: int = 0                    # Requests in this window
    window_start: int = 0             # Unix timestamp / window_seconds
    expires_at: float = 0.0           # When to delete this entry

Limitations:

  • Lost on restart (no persistence)
  • Doesn't scale horizontally (can't share state between servers)
  • Max memory usage: max_keys * ~100 bytes = 10MB for 100k keys

Redis Storage (Distributed)

What we store:

  • Sliding window: Two keys per client per window: key:current_window and key:previous_window
  • Token bucket: Hash with fields {tokens, last_refill, capacity, refill_rate}
  • Lua scripts loaded once at startup, executed via EVALSHA

Why this storage: Production deployments run multiple API servers behind a load balancer. They need shared rate limit state. If Server A allows 50 requests and Server B allows 50, that's 100 total when the limit is 100. Redis provides the single source of truth.

Schema design from src/fastapi_420/storage/lua/sliding_window.lua:13-20:

local current_key = key .. ":" .. current_window
local previous_key = key .. ":" .. previous_window

local current_count = redis.call('GET', current_key) or 0
local previous_count = redis.call('GET', previous_key) or 0

-- Keys auto-expire after 2 windows to prevent memory leaks
redis.call('EXPIRE', current_key, window_seconds * 2)

Performance characteristics:

  • Latency: 1-5ms for local Redis, 10-50ms for remote
  • Throughput: 100k+ operations/sec per Redis instance
  • Memory: ~100 bytes per active key, 1M keys = 100MB

Configuration

Environment Variables

From src/fastapi_420/config.py:120-164:

RATELIMIT_ENABLED=true                  # Master switch, disable for testing
RATELIMIT_ALGORITHM=sliding_window      # sliding_window|token_bucket|fixed_window
RATELIMIT_DEFAULT_LIMIT=100/minute      # Fallback when no specific limit set
RATELIMIT_FAIL_OPEN=true                # Allow requests if storage fails
RATELIMIT_KEY_PREFIX=ratelimit          # Namespace for Redis keys
RATELIMIT_INCLUDE_HEADERS=true          # Add RateLimit-* headers to responses
RATELIMIT_LOG_VIOLATIONS=true           # Log when limits exceeded

# Storage
RATELIMIT_REDIS_URL=redis://localhost:6379/0   # If set, use Redis; else memory
RATELIMIT_REDIS_MAX_CONNECTIONS=100            # Connection pool size
RATELIMIT_MEMORY_MAX_KEYS=100000               # LRU eviction threshold

# Fingerprinting
RATELIMIT_FP_LEVEL=normal                      # strict|normal|relaxed|custom
RATELIMIT_FP_USE_IP=true
RATELIMIT_FP_USE_USER_AGENT=true
RATELIMIT_FP_TRUST_X_FORWARDED_FOR=false       # Enable behind proxies

# Defense
RATELIMIT_DEFENSE_MODE=adaptive                # adaptive|lockdown|disabled
RATELIMIT_DEFENSE_GLOBAL_LIMIT=50000/minute
RATELIMIT_DEFENSE_CIRCUIT_THRESHOLD=10000      # Requests/minute to trip circuit
RATELIMIT_DEFENSE_CIRCUIT_RECOVERY_TIME=30     # Seconds before retry

Configuration Strategy

Development: Use defaults with memory storage. Override via .env file in project root. Settings loaded at src/fastapi_420/config.py:183-188:

@lru_cache
def get_settings() -> RateLimiterSettings:
    return RateLimiterSettings()  # Auto-loads from .env

Production: Set environment variables in container orchestration (Kubernetes ConfigMap, Docker Compose, systemd). The @lru_cache decorator ensures settings load only once per process.

Validation: Pydantic validates settings at startup. Invalid values cause immediate failure with clear error messages. See src/fastapi_420/config.py:165-179:

@model_validator(mode="after")
def validate_limits(self) -> RateLimiterSettings:
    RateLimitRule.parse(self.DEFAULT_LIMIT)  # Raises if invalid
    for limit in self.DEFAULT_LIMITS:
        RateLimitRule.parse(limit)
    return self

Performance Considerations

Bottlenecks

Where this system gets slow under load:

  1. Redis network latency - Every rate limit check requires at least one Redis call. At 5ms latency, max throughput is 200 requests/sec per connection. Solution: Connection pooling (default 100 connections = 20k requests/sec).

  2. Fingerprint computation - Extracting and hashing headers takes ~100 microseconds. Under 10k requests/sec, this is 1 second of CPU time. Solution: Cache fingerprints per request in middleware context (not implemented in base project, shown in challenges).

  3. Lua script compilation - First execution of a Lua script requires compilation. Subsequent calls use EVALSHA with cached script hash. See src/fastapi_420/storage/redis_backend.py:106-125 for script loading.

Optimizations

What we did to make it faster:

  • Pre-loaded Lua scripts: Scripts load once at startup (src/fastapi_420/storage/redis_backend.py:106-125), not on every request. EVALSHA is 10x faster than EVAL.
  • Atomic operations: Single Redis call per rate limit check instead of get-increment-set sequence. Eliminates race conditions and reduces network round trips.
  • Connection pooling: Redis connection pool reuses connections (src/fastapi_420/storage/redis_backend.py:78-90). Creating new connections costs ~10ms each.

Benchmark results (from internal testing):

  • Memory storage: 50,000 checks/sec on single core
  • Redis storage (local): 15,000 checks/sec with default pool
  • Redis storage (remote): 2,000 checks/sec with 50ms latency

Scalability

Vertical scaling: Add more CPU and memory to API servers. Rate limiter is CPU-bound for memory storage, network-bound for Redis. Vertical scaling helps memory storage but not Redis (limited by single Redis instance throughput).

Horizontal scaling: Add more API servers behind load balancer. Memory storage DOES NOT scale horizontally (each server has independent state). Redis storage scales perfectly (shared state).

For >100k requests/sec:

  • Use Redis Cluster (sharding across multiple Redis instances)
  • Consider Memcached or custom storage backend
  • Cache fingerprints to reduce computation

Design Decisions

Decision 1: Sliding Window as Default Algorithm

What we chose: Sliding window counter with weighted interpolation between fixed windows.

Alternatives considered:

  • True sliding window (store all timestamps): Rejected because O(n) memory per client where n = limit. A 1000/hour limit requires storing 1000 timestamps. Current approach uses O(1) memory (two counters).
  • Fixed window: Rejected because of boundary burst problem. Attackers can make 2x limit by timing requests at window edges.
  • Token bucket: Considered but sliding window is easier to explain. "100 per minute" is clearer than "100 tokens with 1.67/second refill rate."

Trade-offs:

  • Gained: 99.997% accuracy with constant memory, no boundary bursts
  • Lost: Not 100% accurate (0.003% error), slightly more complex than fixed window

Implementation at src/fastapi_420/algorithms/sliding_window.py:18-30 uses the formula from the Redis GCRA algorithm paper.

Decision 2: HTTP 420 Instead of 429

What we chose: Return HTTP 420 "Enhance Your Calm" for rate limit violations.

Alternatives considered:

  • HTTP 429: Standard code, rejected because 420 has better developer experience (memorable, distinctive in logs)
  • HTTP 503: Service Unavailable, rejected because it implies the server is broken, not that the client is too fast

Trade-offs:

  • Gained: Distinctive, friendly message, easy to grep logs for "420"
  • Lost: Not IANA-registered (some strict HTTP clients might not recognize it)

The implementation at src/fastapi_420/exceptions.py:40-68 still includes standard headers (Retry-After, RateLimit-*) for compatibility.

Decision 3: Three-Layer Defense

What we chose: Per-user, per-endpoint, and global circuit breaker layers that all must pass.

Alternatives considered:

  • Single layer (per-user only): Rejected because distributed attacks bypass it
  • Two layers (per-user + global): Considered but doesn't protect individual endpoints from being overwhelmed while overall traffic is fine

Trade-offs:

  • Gained: Comprehensive protection against different attack types
  • Lost: Higher latency (3 checks instead of 1), more complex configuration

See implementation at src/fastapi_420/defense/layers.py:47-84. Each layer returns immediately on denial for fast failure.

Deployment Architecture

In production, this typically runs as:

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                   Load Balancer                      β”‚
β”‚              (AWS ALB / Nginx / Cloudflare)          β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
            β”‚                    β”‚
            β–Ό                    β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”    β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚  API Server 1   β”‚    β”‚  API Server 2   β”‚    (N servers)
β”‚  FastAPI +      β”‚    β”‚  FastAPI +      β”‚
β”‚  Rate Limiter   β”‚    β”‚  Rate Limiter   β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”˜    β””β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”˜
         β”‚                      β”‚
         β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                    β–Ό
         β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
         β”‚   Redis Cluster     β”‚
         β”‚  (Shared state)     β”‚
         β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Components:

  • Load Balancer: Distributes traffic, SSL termination, sets X-Forwarded-For header
  • API Servers: Run FastAPI with rate limiting middleware, 4-8 instances typical
  • Redis Cluster: 3-node cluster for high availability, handles 100k+ ops/sec

Infrastructure: Each API server: 2 vCPU, 4GB RAM, runs in Docker container Redis: 4GB RAM, persistence enabled, replica for failover

Error Handling Strategy

Error Types

  1. Storage connection failures - Redis is down or network partitioned. Handled at src/fastapi_420/limiter.py:265-284 with fallback to memory storage if FALLBACK_TO_MEMORY=true.

  2. Invalid configuration - Malformed rate limit strings like "abc/minute". Caught at startup by Pydantic validators (src/fastapi_420/config.py:165-179), application doesn't start.

  3. Race conditions - Multiple requests trying to increment the same counter simultaneously. Prevented by Lua scripts which execute atomically in Redis.

Recovery Mechanisms

Redis connection loss:

  • Detection: Health check fails at src/fastapi_420/storage/redis_backend.py:465-472
  • Response: Switch to fallback MemoryStorage if configured
  • Recovery: Background task retries connection every 30 seconds

Circuit breaker tripped:

  • Detection: Global request count exceeds threshold
  • Response: Reject most traffic, allow authenticated users (adaptive mode)
  • Recovery: After CIRCUIT_RECOVERY_TIME seconds, enter half-open state, gradually allow traffic

Extensibility

Where to Add Features

Want to add a new algorithm (e.g., leaky bucket)?

  1. Create src/fastapi_420/algorithms/leaky_bucket.py implementing BaseAlgorithm protocol
  2. Add to algorithm factory map at src/fastapi_420/algorithms/__init__.py:17
  3. Update Algorithm enum in src/fastapi_420/types.py:28-33
  4. Write tests in tests/test_algorithms.py

Want to add geolocation-based blocking?

  1. Extend FingerprintData in src/fastapi_420/types.py:158 with country_code field
  2. Add geo lookup in CompositeFingerprinter.extract() at src/fastapi_420/fingerprinting/composite.py:96
  3. Add blocking logic in LayeredDefense._should_bypass_circuit() at src/fastapi_420/defense/layers.py:232

Limitations

Current architectural limitations:

  1. No distributed circuit breaker - Circuit breaker state is per-process. In a 10-server deployment, each server has its own circuit. Total threshold is 10x configured value. Fix requires: Shared circuit state in Redis.

  2. No adaptive rate limits - Limits are static, don't adjust based on system load. Under heavy load, might want to reduce limits automatically. Fix requires: Monitor system metrics (CPU, memory) and dynamically calculate limits.

  3. No request cost weighting - All requests count as 1. A request that does heavy computation should count more. Fix requires: Add cost parameter to limiter.check(), multiply count by cost.

These are not bugs, they're conscious tradeoffs. Fixing them would require significant additional complexity.

Comparison to Similar Systems

vs. SlowAPI (Flask-Limiter port)

How we're different:

  • Native async support (SlowAPI uses sync code with thread pooling)
  • Three-layer defense vs single-layer
  • Multiple algorithms built-in vs fixed window only

Why we made different choices: This project targets high-throughput async APIs. SlowAPI targets traditional Flask apps. Different use cases lead to different architectures.

vs. Upstash Rate Limit (serverless-first)

How we're different:

  • Self-hosted Redis vs Upstash cloud service
  • Middleware-based vs edge function integration
  • Per-server vs globally distributed state

Why we made different choices: Upstash optimizes for serverless edge deployments (Vercel, Cloudflare Workers). This project targets traditional server deployments with more control over infrastructure.

Key Files Reference

Quick map of where to find things:

  • src/fastapi_420/limiter.py - Main orchestrator, start here to understand flow
  • src/fastapi_420/middleware.py - ASGI integration, how it hooks into FastAPI
  • src/fastapi_420/algorithms/sliding_window.py - Recommended default algorithm
  • src/fastapi_420/storage/redis_backend.py - Production storage backend
  • src/fastapi_420/storage/lua/ - Atomic Lua scripts for Redis
  • src/fastapi_420/defense/layers.py - Three-layer protection system
  • src/fastapi_420/fingerprinting/composite.py - Client identification
  • src/fastapi_420/config.py - All configuration with validation
  • examples/app.py - Complete working example

Next Steps

Now that you understand the architecture:

  1. Read 03-IMPLEMENTATION.md for detailed code walkthrough showing how each component is built
  2. Try modifying examples/app.py to add custom rate limits on new endpoints