Middle Case Studies Interview Questions (50 MCQs)

Question 1

A URL shortener creates 100 million links/month and each record is ~500 bytes. Roughly how much storage is needed for 5 years?

Accepted Answer

About 3 TB

Answer

About 300 GB

Answer

About 30 TB

Answer

About 30 GB

Question 2

If short codes come from base62-encoding a sequential counter, what is the main risk?

Accepted Answer

Codes are guessable, so all links can be enumerated

Answer

The collision probability between codes becomes very high

Answer

Codes get shorter over time

Answer

Base62 does not work on large numbers

Question 3

How do we prevent a KGS from giving the same key to two app servers?

Accepted Answer

Atomically move keys in batches to a used table and assign each batch to one server

Answer

Ask each server to hash the key before use

Answer

Force all servers to pick keys randomly

Answer

Return keys to the pool after they are used

Question 4

To support custom aliases alongside generated codes, which approach is safest?

Accepted Answer

A unique constraint on the key with conditional insert; error on conflict

Answer

First `SELECT`, then `INSERT` without a constraint if absent

Answer

Write the alias to cache and sync to the database later

Answer

Overwrite any existing code with the user's alias unchecked

Question 5

Which eviction policy makes most sense for caching URL shortener mappings?

Accepted Answer

LRU, because recently hot links get most clicks

Answer

FIFO, because older links are always more popular

Answer

Random, because access is perfectly uniform

Answer

No eviction, because all data fits in RAM

Question 6

If a Snowflake ID generator's clock moves backward (e.g., after NTP sync), what should it do?

Accepted Answer

Detect it and wait until the last timestamp is passed, or fail

Answer

Use the new timestamp without any check

Answer

Reset the sequence number to zero and continue

Answer

Randomly change its machine ID

Question 7

If Snowflake's 12-bit sequence is exhausted within one millisecond, what does the generator do?

Accepted Answer

Waits for the next millisecond and restarts the sequence at zero

Answer

Borrows timestamp bits for the sequence

Answer

Returns a UUID instead of a Snowflake

Answer

Lets the sequence overflow and continues from zero

Question 8

What is the main advantage of range allocation for IDs?

Accepted Answer

Each server takes a block and issues IDs without coordinator round-trips

Answer

IDs will be perfectly time-ordered across the whole system

Answer

No ID is ever wasted

Answer

No durable storage is needed at all

Question 9

What is the main drawback of the sliding window log algorithm?

Accepted Answer

It must store a timestamp for every request per client

Answer

It allows double bursts at window boundaries

Answer

It cannot be implemented in Redis at all

Answer

It only works for daily limits

Question 10

How does the sliding window counter estimate the request count?

Accepted Answer

Current window count plus a weighted share of the previous window's count

Answer

The average count over all windows in the past day

Answer

Only the current window's count, ignoring the previous

Answer

The maximum of the current and previous window counts

Question 11

If a rate limiter reads from Redis and then increments separately, what goes wrong and what's the fix?

Accepted Answer

A race condition; fix with a Lua script or atomic ops like `INCR`

Answer

A memory leak; fix by periodically restarting Redis

Answer

A deadlock; fix by raising the connection timeout

Answer

No problem, since Redis is single-threaded

Question 12

In a microservice architecture, where is general rate limiting usually placed?

Accepted Answer

In the API gateway, before traffic reaches services

Answer

Inside the database as a trigger

Answer

Only in client-side code

Answer

In each service with its own local counter

Question 13

When is leaky bucket preferred over token bucket?

Accepted Answer

When the downstream needs a steady, constant output rate

Answer

When we want to freely let large bursts through

Answer

When we have unlimited memory available

Answer

When requests have no ordering at all

Question 14

In a hybrid news feed model, which division of work is common?

Accepted Answer

Push for normal users, pull for high-follower accounts at read time

Answer

Pull for normal users, push for high-follower accounts

Answer

Push for text posts, pull for image posts

Answer

Push during the day and pull at night

Question 15

Why does each user's feed cache usually store only post IDs rather than full post content?

Accepted Answer

To save memory and so a post edit applies in one place

Answer

Because Redis cannot store long strings

Answer

Because IDs are encrypted and safer

Answer

Because post content is never displayed

Question 16

For correct message ordering in a conversation, which approach is most reliable?

Accepted Answer

A server-assigned monotonic sequence number per conversation

Answer

Sorting by the sender device's clock timestamp

Answer

Sorting by arrival order at the recipient client

Answer

Alphabetical sorting by message text

Question 17

When sender and recipient are connected to different WebSocket servers, how does a message reach the recipient?

Accepted Answer

Via a session registry and inter-server pub/sub to route the message

Answer

The sender connects directly to the recipient device's IP

Answer

The load balancer broadcasts to all servers, which all store it

Answer

The recipient polls every server each minute

Question 18

Which storage choice is common for message history in a large chat app?

Accepted Answer

A wide-column store like Cassandra partitioned by conversation ID

Answer

A single MySQL table with no indexes

Answer

Only Redis without persistence

Answer

A separate JSON file per user on the web server

Question 19

A client sent a message, got no ack, and resends. How do we avoid a duplicate message?

Accepted Answer

The client generates a unique message ID and the server dedups on it

Answer

The server drops messages with identical text

Answer

The client never retries

Answer

The server waits 10 seconds before storing

Question 20

In very large groups, what's wrong with broadcasting each member's presence change to all, and the fix?

Accepted Answer

It creates O(n²) traffic; send presence lazily or only for visible members

Answer

It weakens security; encrypt presence

Answer

It breaks message order; use sequences

Answer

No problem, since presence data is tiny

Question 21

Which strategy suits retrying notification sends on transient provider errors?

Accepted Answer

Exponential backoff with jitter and a max attempt cap

Answer

Immediate, infinite retries until success

Answer

Retrying at a fixed one-millisecond interval

Answer

No retry; drop the notification

Question 22

The queue delivers at-least-once. How do we avoid sending a user the same notification twice?

Accepted Answer

Each event carries a notification ID that the worker checks before sending

Answer

Switch the queue to at-most-once and ignore failures

Answer

Run only one worker for the whole system

Answer

Compare notifications by their text

Question 23

How does the frontier enforce politeness so one site isn't flooded with concurrent requests?

Accepted Answer

A separate queue per host consumed by one worker with a delay

Answer

All URLs go into one shared FIFO queue

Answer

URLs for each host are randomly dropped

Answer

The crawler downloads only one page per site

Question 24

About 30% of web pages are duplicates or near-duplicates. How does a crawler detect duplicate content?

Accepted Answer

Checksums for exact copies and SimHash/MinHash for near-duplicates

Answer

Character-by-character comparison against every prior page

Answer

Comparing only page URLs

Answer

Comparing HTML file sizes

Question 25

How does a crawler avoid getting stuck in a spider trap (like a calendar with endless links)?

Accepted Answer

Limits on URL length, depth, and pages per domain

Answer

Adding workers until the trap is exhausted

Answer

Disabling the Bloom filter for that site

Answer

Following only `https` links

Question 26

Why do large crawlers run a dedicated DNS resolver or cache?

Accepted Answer

DNS lookups are slow and often synchronous, bottlenecking throughput

Answer

Because sites block crawlers without dedicated DNS

Answer

Because DNS caches page content

Answer

Because `robots.txt` is read via DNS

Question 27

Why are top-k suggestions cached at each trie node in autocomplete?

Accepted Answer

So a query is answered in O(prefix length) without traversing the subtree

Answer

To shrink the trie's memory footprint

Answer

So frequencies never need updating

Answer

So the trie can correct misspellings

Question 28

How are autocomplete term frequencies usually updated?

Accepted Answer

Periodic aggregation of query logs and offline trie rebuilds

Answer

Synchronously updating the serving trie on every search

Answer

Manual list editing by a content team

Answer

Never; the trie is built once and stays fixed

Question 29

What's the problem with sharding the trie purely by first letter?

Accepted Answer

Uneven distribution; letters like `s` have far more terms than `x`

Answer

No problem; it is always balanced

Answer

Queries must be sent to every shard

Answer

A trie cannot live on more than one server

Question 30

In a key-value store with N=3 replicas, which R and W setting guarantees reading the latest write?

Accepted Answer

W=2 and R=2, because R + W > N

Answer

W=1 and R=1, because it is fastest

Answer

W=1 and R=2, because reads matter more

Answer

W=0 and R=3, because all replicas are read

Question 31

In Dynamo-style key-value stores, what are vector clocks used for?

Accepted Answer

Detecting whether two versions are concurrent or one supersedes the other

Answer

Synchronizing the physical clocks of all nodes

Answer

Encrypting stored values

Answer

Electing a leader in the cluster

Question 32

What problem does hinted handoff solve in a distributed key-value store?

Accepted Answer

When a replica is temporarily down, another node holds the write and hands it over later

Answer

Permanently purges deleted keys from disk

Answer

Load balances requests across data centers

Answer

Transfers user passwords between nodes

Question 33

Why are Merkle trees used for anti-entropy between replicas?

Accepted Answer

To find and sync only differing ranges without comparing every key

Answer

To store data more compactly

Answer

To guarantee write ordering

Answer

So a new node automatically becomes leader

Question 34

In a decentralized key-value cluster, how do nodes typically learn about each other's failures?

Accepted Answer

A gossip protocol; each node exchanges heartbeat state with a few random nodes

Answer

Every node pings all other nodes each second

Answer

An admin records failures manually in a config file

Answer

Clients email failures to every node

Question 35

For large video uploads, why use pre-signed URLs to send directly to object storage?

Accepted Answer

So bulky data bypasses app servers and reduces their load

Answer

So the video is automatically transcoded

Answer

So no user authentication is needed

Answer

So the video never reaches the CDN

Question 36

How do we dramatically reduce the transcoding time of a two-hour video?

Accepted Answer

Split it into segments and process in parallel across workers as a DAG

Answer

Transcode the whole video serially on one bigger server

Answer

Produce only one resolution and drop the rest

Answer

Defer transcoding until the first playback request

Question 37

CDN cost is a large share of streaming cost. Which optimization is common?

Accepted Answer

Serve only popular videos from the CDN and long-tail ones from origin

Answer

Send every video at the highest bitrate to all users

Answer

Drop the CDN and serve everything from one data center

Answer

Delete videos from the CDN after each play

Question 38

Two devices edited the same file concurrently while offline. What does a system like Dropbox usually do?

Accepted Answer

Accepts the first version and saves the second as a 'conflicted copy'

Answer

Automatically merges both versions line by line

Answer

Deletes both versions

Answer

Keeps whichever version is larger

Question 39

How do Dropbox clients learn about file changes on another device without constantly polling all metadata?

Accepted Answer

A notification service via long polling or WebSocket that just signals changes

Answer

Each client downloads the full file list every second

Answer

The server emails the files

Answer

Clients scan every device on the internet peer-to-peer

Question 40

Millions of drivers send location every 4 seconds. Where should this data live?

Accepted Answer

An in-memory store like Redis with a geo index and short TTL

Answer

A relational database with a disk write per update

Answer

A log file on the driver's device

Answer

A data warehouse with daily loads

Question 41

A user sits near a geohash cell boundary. Why isn't searching only their own cell enough?

Accepted Answer

Very close points may sit in an adjacent cell with a different prefix; search the 8 neighbors too

Answer

Because geohash only works in the northern hemisphere

Answer

Because geohash cells move over time

Answer

Because a longer prefix always reduces accuracy

Question 42

In a proximity service (like Yelp), business data rarely changes but searches are heavy. Which design fits?

Accepted Answer

Read replicas and an in-memory geo index rebuilt periodically

Answer

Send every query to a single primary database

Answer

Have every search full-scan the table

Answer

Fetch business data from an external API on every request

Question 43

Why do payment systems use a double-entry ledger?

Accepted Answer

Each transaction has equal debit and credit, so totals always sum to zero and are auditable

Answer

Because every transaction is processed twice for speed

Answer

Because it computes balances without a database

Answer

Because it allows editing past records

Question 44

What does reconciliation do in a payment system?

Accepted Answer

Periodically comparing internal records with PSP and bank reports to find mismatches

Answer

Encrypting users' card information

Answer

Computing currency exchange rates in real time

Answer

Load balancing across several payment gateways

Question 45

In a flash sale with 100 items and a million concurrent users, which design prevents oversell and protects the database?

Accepted Answer

Atomic stock decrement in Redis and enqueueing successful orders

Answer

Each request does `SELECT` then `UPDATE` directly on the database

Answer

Keep stock in each app server's local memory

Answer

Accept all requests and cancel extra orders later

Question 46

A user picked a seat but hasn't paid yet. What's the most common design?

Accepted Answer

Hold the seat with a TTL (e.g., 10 minutes) and release it if unpaid

Answer

Lock the seat for the user forever

Answer

Let multiple users pay for the same seat at once

Answer

Show the seat only after payment

Question 47

What's the difference between pull (like Prometheus) and push models for metrics collection?

Accepted Answer

In pull, the collector scrapes service endpoints; in push, services send metrics

Answer

Pull is only for logs and push only for traces

Answer

In pull, end users enter metrics manually

Answer

They are identical apart from the name

Question 48

To retain metrics for several years at reasonable cost, what is common?

Accepted Answer

Downsampling old data to lower resolution (e.g., 10s to 1h)

Answer

Keeping all raw points at full resolution forever

Answer

Deleting all metrics older than one day

Answer

Converting metrics to text and storing them in logs

Question 49

A distributed job scheduler usually guarantees at-least-once. What does that imply for jobs?

Accepted Answer

Jobs must be idempotent since they may run more than once

Answer

Jobs never run unless triggered manually

Answer

Each job runs exactly once so no care is needed

Answer

Jobs must finish in under one second

Question 50

What property does a count-min sketch have in the top-K heavy hitters problem?

Accepted Answer

Small fixed memory; may overestimate counts but never underestimates

Answer

Exact counts for all items with linear memory

Answer

May underestimate counts but never overestimates

Answer

Works only on sorted data