Scaling & Advanced Cache Trade-offs Quiz

Q1. EASY: In a distributed caching cluster, what is a primary benefit of using consistent hashing for key distribution when resharding (adding or removing nodes)? It eliminates the need to rebalance data when the cluster size changes
It guarantees each node has an identical share of keys at all times
It stores each key on every node to achieve consistency
It minimizes the number of keys that need to be moved when a node is added or removed

Q2. EASY: What is an example of horizontal scaling in a caching system? Upgrading the cache server with more memory and a faster CPU
Adding more cache servers in parallel to handle increased load
Using a single cache server to eliminate network overhead
Reducing the amount of cached data to keep it within one server

Q3. EASY: In system performance metrics, what does 'p99 latency' refer to? The maximum latency observed for any request
The average latency of the slowest 1% of requests
The 99th percentile latency (the time under which 99% of requests complete)
The latency of the 99th request in a batch of 100 requests

Q4. EASY: Which of the following lists cache storage mediums in order from lowest latency to highest latency? DRAM < Persistent Memory (PMEM) < SSD
SSD < DRAM < PMEM
DRAM < SSD < PMEM
SSD < PMEM < DRAM

Q5. EASY: What is a primary benefit of geo-replicating cache data across multiple regions? It guarantees identical response times for users worldwide
It eliminates the need to ever invalidate cached data
It reduces access latency by serving data from the nearest location to the user
It halves the storage requirements by duplicating data in multiple places

Q6. MEDIUM: In a replicated cache cluster using quorum-based failover, why is a quorum (majority) of nodes required to agree on leader changes? To guarantee zero increase in latency during a failover event
To prevent split-brain by ensuring only one group of nodes can elect a new leader
So that one node can always be kept idle as a hot standby leader
Because reads must be served from a majority of nodes to maintain consistency

Q7. MEDIUM: In a multi-tier cache hierarchy (for example, an L1 cache on the app server and an L2 distributed cache), what is one advantage of having an L1 cache? It requires using twice the memory but provides no performance benefit
It ensures the cache is always consistent with the database without any invalidation
It eliminates the need for an L2 cache entirely if an L1 cache is present
It reduces latency by serving frequently accessed data from a cache local to the application

Q8. MEDIUM: A certain cache key is receiving an extremely high volume of requests, creating a hotspot. Which technique can help distribute this load? Using an indirection layer to map the hot key to multiple cache keys across nodes (key fanning) to spread the traffic
Disabling caching for that hot key so all requests go directly to the database
Moving the hot key's data to a slower storage medium to reduce request frequency
Using a single dedicated cache server exclusively for that hot key

Q9. MEDIUM: Why might a large-scale cache disable compression for very small objects? Because compressing small objects always risks corrupting their data in the cache
Because the compression overhead (CPU and metadata) can outweigh any size savings for tiny objects
Because caches cannot compress objects below a certain size due to system limitations
Because small objects are too tiny for any compression algorithm to handle

Q10. MEDIUM: Cloudflare's CDN uses an edge microcache in its architecture. What is the role of this edge microcache? It permanently stores all content at every edge server to completely eliminate origin traffic
It is a cache in the end-user's browser controlled by Cloudflare to speed up repeat visits
It only caches large media files (images and videos) at the edge and ignores other content
It caches content at edge servers for a short time to absorb traffic bursts and reduce load on the origin

Q11. MEDIUM: Facebook's Memcached tier scaled to handle billions of requests per second. Which approach was key to achieving this scale? They replaced Memcached with a relational database to handle all read requests
They used a single gigantic cache server with terabytes of RAM to avoid sharding
They deployed a two-tier caching system with a local L1 cache on each app server and a large L2 Memcached cluster
They offloaded caching to user browsers (via cookies) instead of using server-side caches

Q12. HARD: In a distributed cache cluster, coordinating evictions across nodes and using an admission policy (like TinyLFU) can improve efficiency. What is the main purpose of an admission policy? To prevent cache thrashing by filtering out low-value (one-hit) items from being inserted, reducing unnecessary evictions
To replicate every cached item to all nodes upon insertion for strong consistency
To only evict items after every node in the cluster agrees on it via consensus
To bypass the cache for all write operations, caching only read operations

Q13. HARD: To reduce tail latency in a distributed service, a technique called 'request hedging' can be used. What does request hedging involve? Partitioning a single client request into multiple sub-requests processed in parallel on one server
Sending the same request to multiple servers or replicas and using whichever response arrives first
Delaying incoming requests slightly to see if a faster path becomes available before processing
Automatically canceling any request that exceeds a predefined p99 latency threshold

Q14. HARD: In a multi-tier caching setup, what is a potential issue caused by propagation delay between the L1 and L2 caches? The application may read stale data from the L1 cache if it hasn't been updated or invalidated yet with the latest write
The L2 cache will always have more up-to-date data than the origin database
The L1 cache might overwrite newer data in L2 due to asynchronous writes
Propagation delays cause both L1 and L2 caches to evict the same item at exactly the same time

Q15. HARD: In a distributed cache cluster, one shard shows a much higher eviction rate and slower p99 latency than others. What is the most likely explanation? That shard is running on faster hardware, which causes it to evict items more frequently
The consistent hashing algorithm designates one node to evict data on behalf of the entire cluster
Other shards have effectively infinite capacity, so only this shard performs evictions
That shard is handling a disproportionately large share of requests or data (a shard imbalance/hotspot), leading to more evictions and higher tail latency

system-design