Rate limiting in Microservices - Scalability & System Analysis

Scalability Analysis - Rate limiting

Growth Table: Rate Limiting at Different Scales

Users	Requests per Second (RPS)	Rate Limiter Type	Infrastructure Changes	Challenges
100 users	~500 RPS	In-process (local) rate limiting	Single microservice instance	Simple counters, low overhead
10,000 users	~50,000 RPS	Centralized rate limiter (Redis or API Gateway)	Multiple microservice instances, shared cache	Consistency, latency in shared store
1,000,000 users	~5,000,000 RPS	Distributed rate limiting with sharded stores	Multiple rate limiter clusters, load balancers	Data partitioning, synchronization, failover
100,000,000 users	~500,000,000 RPS	Hierarchical rate limiting with edge/CDN enforcement	Global distributed caches, edge nodes, multi-region	Network bandwidth, global consistency, cost

First Bottleneck

At small scale, the first bottleneck is the in-process memory for counters in each microservice instance. As traffic grows, the bottleneck shifts to the centralized data store (like Redis) used for shared counters, which can become overwhelmed by high request rates and cause latency.

Scaling Solutions

• Local Rate Limiting: Use in-memory counters for low traffic to avoid network calls.
• Centralized Store: Use Redis or Memcached with connection pooling for moderate scale.
• Sharding: Partition keys by user or API key to distribute load across multiple Redis instances.
• Hierarchical Rate Limiting: Combine edge (CDN or API Gateway) and backend limits to reduce backend load.
• Token Bucket or Leaky Bucket Algorithms: Efficient algorithms to smooth bursts and reduce storage overhead.
• Asynchronous Updates: Use approximate counters or probabilistic data structures to reduce write load.
• Load Balancing: Distribute requests evenly to rate limiter clusters to avoid hotspots.

Back-of-Envelope Cost Analysis

• At 10,000 users with 50,000 RPS, Redis needs to handle ~50,000 ops/sec, which is near a single Redis instance limit; requires sharding or clustering.
• Each request counter uses a few bytes; for 1M users, storage for counters can reach several GBs in Redis.
• Network bandwidth for rate limiter calls grows with RPS; at 5M RPS, requires multiple high-throughput network links.
• CPU usage on microservices increases with local rate limiting logic; offloading to dedicated rate limiter services can reduce this.

Interview Tip

Start by clarifying the scale and traffic patterns. Discuss simple local rate limiting first, then explain how centralized stores become bottlenecks. Describe sharding and hierarchical approaches. Emphasize trade-offs between accuracy, latency, and cost. Use real numbers to show understanding of limits and solutions.

Self Check Question

Your database handles 1000 QPS for rate limiting counters. Traffic grows 10x to 10,000 QPS. What do you do first?

Answer: Introduce caching or sharding to distribute load. For example, add Redis read replicas or partition counters by user ID to multiple Redis instances to avoid overloading a single database.

Key Result

Rate limiting scales from simple in-memory counters at low traffic to distributed, sharded, and hierarchical systems at high traffic. The first bottleneck is usually the centralized data store for counters, which requires sharding and caching to scale efficiently.

Practice

1. What is the main purpose of rate limiting in microservices?

easy

A. To control how many requests a user can make in a given time

B. To increase the speed of the service

C. To store user data securely

D. To balance the load between servers

Solution

1. Step 1: Understand the concept of rate limiting

   Rate limiting is designed to restrict the number of requests a user or client can send to a service within a certain time frame.

2. Step 2: Identify the main goal of rate limiting

   The main goal is to prevent overload and abuse by controlling request frequency, not to speed up services or store data.

3. Final Answer:

   To control how many requests a user can make in a given time -> Option A

4. Quick Check:

   Rate limiting = Control request count [OK]

Hint: Rate limiting limits request count per time [OK]

Common Mistakes:

• Confusing rate limiting with load balancing
• Thinking rate limiting speeds up the service
• Mixing rate limiting with data storage

2. Which of the following is the correct way to represent a fixed window rate limiter allowing 100 requests per minute in pseudocode?

easy

A. if requests_in_last_minute < 100 then block else allow

B. if requests_in_last_hour > 100 then block else allow

C. if requests_in_last_minute > 100 then block else allow

D. if requests_in_last_second > 100 then allow else block

Solution

1. Step 1: Understand fixed window rate limiting logic

   Fixed window rate limiting counts requests in a fixed time window (e.g., 1 minute) and blocks if the count exceeds the limit.

2. Step 2: Match the correct condition for allowing or blocking

   If requests exceed 100 in the last minute, block; otherwise, allow. if requests_in_last_minute > 100 then block else allow matches this logic exactly.

3. Final Answer:

   if requests_in_last_minute > 100 then block else allow -> Option C

4. Quick Check:

   Fixed window limit = block if over limit [OK]

Hint: Block when requests exceed limit in fixed window [OK]

Common Mistakes:

• Using wrong time window (hour instead of minute)
• Reversing the condition (blocking when under limit)
• Allowing requests when they should be blocked

3. Given this pseudocode for a token bucket rate limiter:

bucket_capacity = 5
refill_rate = 1 token per second
current_tokens = 3
request_tokens = 2
if current_tokens >= request_tokens:
    current_tokens -= request_tokens
    allow request
else:
    block request

What happens if a request for 4 tokens arrives immediately?

medium

A. Request is allowed and tokens reduce to -1

B. Request is blocked because refill rate is too low

C. Request is allowed and tokens reduce to 1

D. Request is blocked because not enough tokens

Solution

1. Step 1: Check current tokens against requested tokens

   Current tokens are 3, request needs 4 tokens, which is more than available.

2. Step 2: Determine if request is allowed or blocked

   Since current tokens (3) < request tokens (4), the request is blocked.

3. Final Answer:

   Request is blocked because not enough tokens -> Option D

4. Quick Check:

   Tokens < request = block [OK]

Hint: Allow only if tokens ≥ requested tokens [OK]

Common Mistakes:

• Allowing request when tokens are insufficient
• Ignoring token count and refill rate
• Assuming tokens can go negative

4. A microservice uses a sliding window rate limiter but users report some requests are blocked even when they seem under the limit. Which is the most likely cause?

medium

A. The sliding window is not updating timestamps correctly

B. The service has too many servers without shared state

C. The rate limit is set too high

D. The users are sending requests too slowly

Solution

1. Step 1: Understand sliding window rate limiter behavior

   Sliding window requires accurate tracking of request timestamps across all servers to count requests correctly.

2. Step 2: Identify issue with multiple servers and no shared state

   If servers do not share state, each counts requests independently, causing incorrect blocking even if total requests are under limit.

3. Final Answer:

   The service has too many servers without shared state -> Option B

4. Quick Check:

   Multiple servers need shared state for sliding window [OK]

Hint: Sliding window needs shared state across servers [OK]

Common Mistakes:

• Blaming slow user requests
• Assuming rate limit is too high causes blocking
• Ignoring distributed state issues

5. You design a rate limiter for a microservice that must handle 10 million users, each allowed 100 requests per hour. Which approach best balances accuracy and scalability?

hard

A. Use distributed token buckets with local caches and periodic sync

B. Use a centralized fixed window counter stored in a single database

C. Use client-side rate limiting without server checks

D. Use a sliding window log storing every request timestamp centrally

Solution

1. Step 1: Analyze scalability needs for 10 million users

   A centralized database (Use a centralized fixed window counter stored in a single database) or storing every timestamp centrally (Use a sliding window log storing every request timestamp centrally) will cause bottlenecks and high latency.

2. Step 2: Evaluate distributed token bucket with local caches

   Distributed token buckets with local caches reduce central load and sync periodically, balancing accuracy and scalability well.

3. Step 3: Consider client-side rate limiting

   Client-side (Use client-side rate limiting without server checks) is unreliable as clients can bypass limits.

4. Final Answer:

   Use distributed token buckets with local caches and periodic sync -> Option A

5. Quick Check:

   Distributed token bucket = scalable + accurate [OK]

Hint: Distributed token buckets scale best for millions [OK]

Common Mistakes:

• Choosing centralized storage causing bottlenecks
• Relying only on client-side limits
• Storing all request logs centrally causing overload