System Design interview questions and answers

1. Duplicate emails despite a check

Signup code runs SELECT to check whether the email exists and then INSERTs the user. Duplicate accounts still appear under heavy traffic. What is the right fix?

1. A.Add a short sleep between the check and the insert
2. B.Add a UNIQUE constraint (or unique index) on email and handle the duplicate-key error
3. C.Run the SELECT check twice before inserting
4. D.Wrap the SELECT and INSERT in one READ COMMITTED transaction so the check and the insert happen together

2. Updating a counter in MongoDB

Many requests add 1 to a post's likes at the same time. Which update is safe?

1. A.Read the post, add 1 in code, then save it
2. B.updateOne({ _id }, { $inc: { likes: 1 } })
3. C.updateOne({ _id }, { $set: { likes: oldLikes + 1 } })
4. D.replaceOne({ _id }, { likes: oldLikes + 1 })

3. Counting API calls in Redis

Ten servers count API calls per customer in Redis. Which command keeps the count exact under concurrency?

1. A.GET then SET with the value plus one
2. B.APPEND calls:cust42 1
3. C.INCR calls:cust42
4. D.SETNX calls:cust42 1

4. What a load balancer does

An app runs on four identical servers behind a load balancer. What is the load balancer's main job?

1. A.Spread requests across healthy servers
2. B.Store session data for all four servers
3. C.Cache database queries for the servers
4. D.Encrypt traffic between the servers and the DB

5. Latency versus throughput

Which statement correctly separates latency from throughput?

1. A.Latency is requests per second; throughput is time per request
2. B.Latency is time per request; throughput is requests per second
3. C.Both mean the same thing measured in different units
4. D.Latency only applies to databases, throughput to APIs

6. Estimating requests per second

An app has 10 million daily active users, and each makes about 20 API requests per day. What is the average request rate?

1. A.About 2,300 requests per second
2. B.About 23,000 requests per second
3. C.About 230 requests per second
4. D.About 200,000 requests per second

7. How a token bucket limits rate

A token bucket rate limiter has capacity 10 and refills 1 token per second. A client has been idle for a minute and then sends 15 requests at once. How many are allowed immediately?

1. A.15
2. B.1
3. C.10
4. D.0

8. Large file uploads

Users upload videos up to 2 GB. Sending them through the API servers ties up memory and bandwidth. What is the common design?

1. A.Give clients a pre-signed URL to upload to object storage
2. B.Store each video as a BLOB column in the main database
3. C.Raise the API servers’ request body limit to 2 GB
4. D.Split each video into chunks and store them in Redis

9. Spotting a single point of failure

A system has 6 app servers across two zones, one load balancer pair, and one PostgreSQL server with nightly backups. What is the single point of failure?

1. A.The load balancer pair
2. B.The six app servers
3. C.The two availability zones
4. D.The single PostgreSQL server

10. DNS failover is slow

A team fails over to a backup region by changing a DNS record with a TTL of 3600 seconds. Why do many users still hit the dead region for a while?

1. A.DNS changes need a full day to be approved
2. B.Load balancers ignore DNS records completely
3. C.Browsers only read DNS on the first ever visit
4. D.Resolvers and clients cache the old record up to an hour

11. Pushing live order status

A food app shows live order status that only the server updates. The browser never sends data on this channel. What is a simple fit?

1. A.A WebSocket with a custom binary protocol
2. B.Server-Sent Events over HTTP
3. C.Polling every 100 milliseconds
4. D.Email notifications for each status change

12. Getting a repository into a ViewModel

With Hilt, how do you get an OrdersRepository into OrdersViewModel?

1. A.Annotate with @HiltViewModel and use an @Inject constructor
2. B.Create the repository inside onCreate of the Activity
3. C.Use a global lateinit repository set at startup
4. D.Pass the repository through an Intent extra

13. Picking a framework for a React team

A startup's engineers all build React web apps with TypeScript. They need one mobile app for Android and iOS soon. Which choice usually fits best?

1. A.React Native, so they reuse React, TypeScript and many npm libraries
2. B.Flutter, because it shares more code with a React web codebase
3. C.Two fully native apps in Swift and Kotlin written from scratch
4. D.A PWA only, because app stores do not accept hybrid apps

14. How often Ethereum makes a block

Under proof of stake, how often does Ethereum mainnet have a slot in which a new block can be proposed?

1. A.Every 12 seconds
2. B.Every 10 minutes
3. C.Every 1 second
4. D.Every 2 minutes

15. Why Ethereum moved to proof of stake

What was the main practical change when Ethereum switched from proof of work to proof of stake in 2022?

1. A.Transaction fees went to zero
2. B.Every transaction became private
3. C.Energy use dropped by over 99% since mining was no longer needed
4. D.Blocks started to need approval from a central company

16. Heavy work in an interrupt

A UART receive ISR parses a full command and writes to flash. What is the usual better design?

1. A.Raise the ISR to the lowest priority
2. B.Add more delay loops inside the ISR
3. C.Store the byte, set a flag, and parse in the main loop
4. D.Disable all interrupts for the whole program

17. printf in an interrupt

Why is calling printf inside an ISR a bad idea?

1. A.printf only works before main has started
2. B.The compiler removes printf from ISRs
3. C.It is slow, may block, and is often not reentrant
4. D.printf cannot print integers on ARM

18. ACID vs eventual consistency

A team compares a relational database with a store that offers only eventual consistency across replicas. What does eventual consistency mean?

1. A.Every read always sees the latest write
2. B.Writes are rejected until all replicas agree
3. C.Data is checked for constraints once a day
4. D.Replicas converge if writes stop, but reads may be stale

19. A logo in a browser tab

The full logo with name and tagline is unreadable as a 16×16 or 32×32 browser favicon. What is the usual approach?

1. A.Use the simplified symbol or a single letter
2. B.Shrink the full logo and hope it reads
3. C.Use a photo of the founder
4. D.Leave the favicon blank

20. Paying reverse charge tax

A company owes ₹18,000 GST under reverse charge on legal fees. Can it pay this from its ITC balance?

1. A.Yes, any ITC can pay reverse charge tax
2. B.No, reverse charge tax must be paid in cash
3. C.Yes, but only IGST credit can be used
4. D.No, reverse charge tax is paid by the lawyer

21. Async callback inside forEach

getPrice(id) returns a promise that resolves to 10 after a short delay. What does this print?

async function totalPrice(ids) {
  let total = 0
  ids.forEach(async (id) => {
    total += await getPrice(id)
  })
  return total
}

totalPrice([1, 2, 3]).then(console.log)

1. A.30
2. B.NaN
3. C.0
4. D.undefined

22. Different prices on different servers

Each of 6 API servers caches prices in a local in-memory Map for 10 minutes. After a price change, users see different prices on refresh. Why?

1. A.In-memory Maps lose their data every few seconds
2. B.The load balancer rewrites the prices it sends
3. C.JavaScript Maps are unable to store numbers
4. D.Each server has its own copy, cleared separately

23. Average latency with a cache

A cache answers in 1 ms and the database in 20 ms. On a miss the app checks the cache, then the database. With a 90% hit ratio, what is the average read latency?

1. A.3 ms
2. B.2.9 ms
3. C.2 ms
4. D.10.5 ms

24. How many short URLs fit

A URL shortener uses 7-character codes from [a-zA-Z0-9]. Roughly how many unique codes exist?

1. A.About 78 billion
2. B.About 3.5 trillion
3. C.About 10 million
4. D.About 8 billion

25. Why look at p99 latency

Average API latency is 80 ms but users complain the app is often slow. What metric is most likely to show the problem?

1. A.p99 latency
2. B.Mean latency over a day
3. C.Total requests per day
4. D.Median CPU usage

26. IP rate limits block real users

A login rate limit of 10 requests per minute per IP keeps blocking many genuine users on one Indian mobile carrier. What is the likely cause?

1. A.Mobile phones send requests twice by default
2. B.The carrier caches login responses
3. C.Carrier-grade NAT puts many users behind one IP
4. D.IPv4 addresses change every minute

27. Separate models for reads and writes

An order system has complex write rules, while the dashboard needs fast, heavily joined views. The team adopts CQRS. What does that mean in practice?

1. A.Every request is written to two databases at once
2. B.Writes use one model; reads use a separate, denormalised view
3. C.Queries and commands share one table with more indexes
4. D.Reads are served only from the write database

28. Surviving screen rotation

On Android, a screen loses its loaded data when the phone is rotated. What is the recommended fix?

1. A.Lock the screen to portrait for all users
2. B.Store the data in a Jetpack ViewModel, which survives configuration changes
3. C.Store the data in a static global variable
4. D.Reload the data in onResume every time

29. Downloading a list of images

You need to download a variable number of thumbnails concurrently and collect the results. What fits best?

1. A.One async let per URL written by hand
2. B.withThrowingTaskGroup, adding one child task per URL
3. C.A for loop that awaits each download in turn
4. D.DispatchQueue.main.sync for each download

30. Messy player code

Player code is full of flags like isJumping, isAttacking and isClimbing, causing bugs where two states run at once. What is a good refactor?

1. A.Add more boolean flags to handle each combination
2. B.Use a finite state machine where the player is in exactly one state with clear transitions
3. C.Move all logic into one big Update method
4. D.Run each state on a separate thread

31. High-priority task starves

A low-priority logger holds a lock that a high-priority control task needs. A medium-priority task then runs for 200 ms, and the control task misses deadlines. What fixes this?

1. A.Use a mutex with priority inheritance for the shared lock
2. B.Give the medium task an even higher priority
3. C.Replace the lock with a binary semaphore
4. D.Make the logger task run more often

32. Two tasks stop forever

Task A takes the SPI mutex then the flash mutex. Task B takes the flash mutex then the SPI mutex. Both freeze after some hours. What is the cleanest fix?

1. A.Give task A a higher priority than task B
2. B.Add a 1 ms delay between taking the two mutexes
3. C.Use recursive mutexes for both resources
4. D.Make every task take the two mutexes in the same order

33. Gatekeeper task for a shared UART

Many tasks print to one UART. Instead of a mutex, a gatekeeper task owns the UART and others send it strings through a queue. What is a key benefit?

1. A.Output can then go faster than the UART baud rate allows
2. B.The queue removes the need for any RAM buffers at all
3. C.Each writer task then gets its own physical UART port
4. D.Writers do not block on the UART, and no inversion on it can occur

34. How much Lambda concurrency you need

A function receives 200 requests per second and each takes 500 ms. About how many concurrent executions does it need?

1. A.About 200
2. B.About 400
3. C.About 40
4. D.About 100

35. Why add jitter to retries?

Your clients retry failed calls with exponential backoff. Why should you also add random jitter to the wait time?

1. A.To stop many clients from retrying at the same moment and overloading the recovering service in synchronized waves
2. B.To make each retry faster than the previous one
3. C.To make sure every retry goes to a different region
4. D.Jitter is only needed when the requests are not idempotent

36. Timeouts in Go clients

In a Go service, what is the idiomatic way to stop an outbound HTTP call from waiting longer than 2 seconds and to cancel it if the incoming request is cancelled?

1. A.Create the request with a context from context.WithTimeout(r.Context(), 2*time.Second)
2. B.Call time.Sleep(2 * time.Second) before the request
3. C.Use a global variable that is checked after the call returns
4. D.Set GOMAXPROCS to 2

37. Two clerks edit the same record

Two support agents open the same customer record and both save changes. The second save silently overwrites the first. Which model change helps detect this?

1. A.A version column checked and incremented on update
2. B.A created_at column on the table
3. C.A UNIQUE constraint on the customer name
4. D.A separate table for each agent

38. Threads with nothing to do

A Kafka Streams app reads one 12-partition topic. You run 3 instances, each with num.stream.threads=8. How many threads actually process data?

1. A.12, because there are only 12 tasks
2. B.24, all threads share the work
3. C.8, one instance does all the work
4. D.3, one thread per instance

39. Punctuating bullet lists

A bulleted list has items that are short fragments, such as 'Free delivery' and 'Easy returns'. Some end in full stops and some do not. What is the best edit?

1. A.Make them consistent; fragments usually take no full stop
2. B.Add a full stop and 'and' to every item
3. C.Add a semicolon to each item and a full stop to the last
4. D.Leave it, since bullet punctuation does not matter

40. Cash paid under the wrong head

A taxpayer mistakenly deposited ₹50,000 under CGST instead of IGST in its cash ledger. What can it do?

1. A.Move it to IGST through Form PMT-09
2. B.Nothing; it can only be refunded after a year
3. C.Use the CGST cash to pay IGST directly
4. D.Ask the bank to reverse the payment

41. Burst at a fixed window boundary

A fixed-window limiter allows 100 requests per calendar minute per client. What is the most requests a client can get accepted within a 2-second span?

1. A.100
2. B.101
3. C.200
4. D.6,000

42. Pool exhausted during payments

Checkout opens a transaction, inserts the order, calls the payment provider over HTTP (2 to 5 seconds), then commits. With a pool of 20 connections, checkout times out at peak. What is the best redesign?

1. A.Raise the pool size to 500 connections and set the database max_connections high enough to cover every single app instance
2. B.Use READ UNCOMMITTED so the transaction finishes faster
3. C.Wait forever for a free connection instead of timing out
4. D.Commit the order as pending first, call the provider with an idempotency key, then update it in a new short transaction

43. One company sees another's invoices

In a multi-tenant SaaS, a new report query forgot `WHERE tenant_id = ?`. Company A saw Company B's invoices. Which approach prevents this whole class of bug?

1. A.Ask reviewers to look for tenant_id in PRs
2. B.Give each tenant a different colour theme
3. C.Hide each tenant's report URLs from others
4. D.Enforce tenant scope centrally, e.g. with RLS

44. Keeping OAuth tokens out of the browser

A security review says a single-page app should not hold OAuth refresh tokens in JavaScript at all. Which pattern does that?

1. A.A backend-for-frontend holds tokens and gives the SPA a session cookie
2. B.Store the refresh token in sessionStorage instead
3. C.Encrypt the refresh token with a key in the JavaScript bundle
4. D.Use the implicit flow so no refresh token is issued

45. CROSSSLOT errors after moving to Cluster

After moving to Redis Cluster, MGET cart:42 wishlist:42 fails with CROSSSLOT. What is the fix if both must be read together?

1. A.Increase the number of hash slots
2. B.Run the same command on a replica
3. C.Use hash tags, like {user:42}, in both keys
4. D.Rename both keys to use only digits

46. One Redis key gets half of all traffic

During a cricket final, one key holding the live score gets 300,000 reads a second, and the Redis node holding it is maxed out while others idle. What helps?

1. A.Move the key to a node with more disk
2. B.A brief local cache per app, or key copies
3. C.Give the key a longer name to spread it
4. D.Delete the key straight after every read

47. Keeping wallet balances only in Redis

A team wants to keep wallet balances only in Redis, because it is fast. What is the main risk?

1. A.Async replication and persistence gaps can lose data
2. B.Redis cannot store decimal numbers in string values
3. C.Redis allows only 1,000 keys in each logical database
4. D.Redis does not support atomic integer increments

48. A join across two services' databases

With database-per-service, the orders service needs customer names for its order list, but customers live in another service's database. What is a common solution?

1. A.Query the other service's tables directly
2. B.Merge both services under one shared database user
3. C.Ask the frontend to join the data by hand
4. D.Keep a local copy of customer names updated by events

49. Adding a cache server without reshuffling

Cache keys are assigned with server = hash(key) % N. Going from 4 to 5 servers moves most keys and causes a flood of cache misses. What reduces this?

1. A.A larger modulus such as % 1000
2. B.Sorting keys before assignment
3. C.Consistent hashing on a ring
4. D.Using MD5 instead of the current hash

50. What a Bloom filter can promise

A sign-up page checks a Bloom filter of taken usernames before querying the database. What can the filter say for sure?

1. A."Taken" is reliable; "not taken" can sometimes be wrong
2. B.Both answers are always exact, just faster than a DB
3. C."Definitely not taken" is reliable; "maybe taken" can be wrong
4. D.It can list every taken username on request

51. Stopping double charges on retries

A mobile app retries a UPI payment request when the network drops, and some customers are charged twice. What server-side fix uses a key-value store?

1. A.Store each idempotency key with its result and replay it on repeats
2. B.Hash the amount and reject any repeated amounts
3. C.Keep a Set of customer IDs who paid today and block them
4. D.Retry the bank call until it returns a success

52. One tenant floods the queue

A SaaS uses one job queue for all tenants. One customer uploads 2 million rows, and every other tenant's small jobs wait hours. What is a good fix?

1. A.Isolate tenants so each gets a fair share of workers
2. B.Add a single priority level for all small jobs
3. C.Raise visibility timeout so big jobs finish first
4. D.Move the large tenant's data to a bigger database

53. One hot RabbitMQ queue

One RabbitMQ queue handles 40,000 messages per second and tops out, even though the cluster has idle nodes and CPU. Why, and what helps?

1. A.The exchange is the bottleneck; switch to fanout
2. B.A queue lives on one node and core; split it across queues
3. C.Prefetch is too low; set it to unlimited
4. D.The disk is too slow; turn off persistence

54. A fixed concurrency limit

A service caps concurrent requests to its database at a fixed 200. On some days the database copes with 400; on others it slows down at 120. What approach handles both?

1. A.A higher fixed limit of 1,000 requests
2. B.Removing the limit and relying on timeouts
3. C.A fixed limit of 120 for every single day
4. D.A limit that adapts to observed latency

55. Fan-out for a celebrity account

A feed uses fan-out on write. A cricketer with 40 million followers posts, and feed writes back up for an hour. What design is common?

1. A.Fan out to all 40 million followers in one transaction
2. B.Fan out for normal users; merge celebrity posts at read time
3. C.Stop showing celebrity posts in feeds
4. D.Give the celebrity a separate database server

56. Offline edits that must merge

A notes app lets users edit the same list offline on two phones and must merge changes without losing either side. What fits best?

1. A.Last-write-wins on the whole document
2. B.Locking the note on the first device to open it
3. C.Rejecting the second phone's changes on sync
4. D.A CRDT that merges concurrent edits deterministically

57. Designing a wallet ledger

A payments team stores each wallet's balance as one mutable number. Audits cannot explain mismatches. What design is standard for money?

1. A.An append-only double-entry ledger of postings
2. B.A balance column updated with UPDATE statements
3. C.A nightly batch that recalculates balances from logs
4. D.A cache of balances refreshed every minute

58. Passing a class into a Task

In Swift 6 language mode, sending a mutable non-Sendable class instance from the main actor into a detached Task gives an error. Why?

1. A.Both sides could mutate it at once, a data race
2. B.Classes cannot be used inside any task at all
3. C.Detached tasks only accept String values
4. D.The class must first be marked with @objc

59. Scaling stopped at a hidden limit

During a big sale, auto scaling stopped adding instances at 64 vCPUs although max size allowed more. What most likely happened, and how do you prevent it?

1. A.Auto scaling is limited to 64 vCPUs for all users
2. B.An account vCPU quota was hit; raise quotas ahead
3. C.The load balancer refused more targets
4. D.The AMI was too large to launch more copies

60. Two leaders after a GC pause

A worker holding a leader lease in etcd paused for 40 s in garbage collection. Its lease expired, a new leader was elected, and then the old worker woke up and wrote to storage, corrupting data. What prevents this?

1. A.Make the lease longer than any possible pause
2. B.Use fencing tokens: each new lease gets a higher number and storage rejects writes with an older token
3. C.Check the lease once before starting the job
4. D.Run the leader on a bigger machine