Middle Scalability & Performance Interview Questions (50 MCQs)

Question 1

What relationship does Little's law express?

Accepted Answer

Average items in the system = arrival rate × average time in system (`L = λW`)

Answer

Maximum speedup = 1 divided by the serial fraction of the program

Answer

Throughput = number of servers × per-request latency

Answer

Latency always decreases linearly as throughput rises

Question 2

A service receives 2,000 RPS with an average latency of 50 ms. On average, how many requests are in flight concurrently?

Accepted Answer

About 100 requests

Answer

About 40 requests

Answer

About 1,000 requests

Answer

About 10 requests

Question 3

Each request holds a DB connection for 20 ms on average and peak load is 3,000 RPS. Roughly how many active connections are needed across all instances?

Accepted Answer

About 60 connections

Answer

About 150 connections

Answer

About 600 connections

Answer

About 3,000 connections

Question 4

Per Amdahl's law, if 10% of a job is inherently serial, what is the maximum speedup with infinite processors?

Accepted Answer

10x

Answer

90x

Answer

100x

Answer

Unlimited

Question 5

A job is 95% parallelizable. With 20 workers, what is the approximate speedup per Amdahl's law?

Accepted Answer

About 10x

Answer

About 19x

Answer

About 20x

Answer

About 5x

Question 6

100 million text posts averaging 500 bytes (including metadata) are stored daily. Roughly what is one year's raw volume?

Accepted Answer

About 18 TB

Answer

About 1.8 TB

Answer

About 180 TB

Answer

About 50 GB

Question 7

A service serves 500 image downloads per second at peak, averaging 200 KB each. Roughly what is the egress bandwidth?

Accepted Answer

About 100 MB/s, i.e. about 800 Mbps

Answer

About 10 MB/s, i.e. about 80 Mbps

Answer

About 100 Mbps, i.e. about 12.5 MB/s

Answer

About 1 GB/s, i.e. about 8 Gbps

Question 8

Users upload 2 million photos daily at 3 MB average, with 3 replicas. Roughly how much storage is needed per day?

Accepted Answer

About 18 TB

Answer

About 6 TB

Answer

About 1.8 TB

Answer

About 60 TB

Question 9

Roughly how many times slower is a spinning-disk seek than a RAM access?

Accepted Answer

About `10^5` times; ~10 ms vs ~100 ns

Answer

About 10 times; ~1 ms vs ~100 µs

Answer

About 100 times; ~10 µs vs ~100 ns

Answer

About `10^9` times; ~10 s vs ~10 ns

Question 10

An endpoint makes 40 sequential calls to a service in another region with ~80 ms RTT. What is the best fix?

Accepted Answer

Batch the calls into one request or colocate the service in the same region

Answer

Add more instances of the target service to lower the RTT

Answer

Upgrade the caller's CPU so it sends requests faster

Answer

Raise each call's timeout to 5 seconds

Question 11

Four servers each report their own p99. How should the service-wide p99 be computed?

Accepted Answer

Merge the latency histograms or sketches and compute p99 from them

Answer

Take the average of the four p99 values

Answer

Pick the maximum of the four p99 values

Answer

Report the p99 of the server with the most traffic

Question 12

A user request fans out in parallel to 100 backends and waits for all. If each backend is slow only 1% of the time, what fraction of user requests is slow?

Accepted Answer

About 63%

Answer

About 1%

Answer

About 10%

Answer

About 99%

Question 13

A service has a 99.9% success SLO over 30 days and 50 million requests per month. How many failed requests is the error budget?

Accepted Answer

About 50,000 requests

Answer

About 5,000 requests

Answer

About 500,000 requests

Answer

About 500 requests

Question 14

A service needs two independent dependencies per request, each with 99.9% availability. What is the service's maximum availability?

Accepted Answer

About 99.8%

Answer

About 99.9%

Answer

About 99.99%

Answer

About 99.0%

Question 15

Two independent replicas each have 99% availability, and the service works if either is up. Roughly what is the combined availability?

Accepted Answer

About 99.99%

Answer

About 98%

Answer

About 99.5%

Answer

About 99.9%

Question 16

The customer SLA is 99.9%. How should the team's internal SLO be set?

Accepted Answer

Slightly stricter, e.g. 99.95%, so they are warned before breaching the SLA

Answer

Exactly 99.9% so the team does no extra work

Answer

Looser, e.g. 99.5%, so the team is not under pressure

Answer

100% so no error ever happens

Question 17

Which latency SLI definition is more reliable and better suited to computing an error budget?

Accepted Answer

The ratio of requests served under 300 ms to all requests

Answer

The average latency of all requests per minute

Answer

The maximum latency observed per day

Answer

The latency of the last request in each 5-minute window

Question 18

A team raised the Postgres connection pool from 20 to 500 to fix slowness, and latency got worse. What is the most likely reason?

Accepted Answer

With limited cores, the database thrashed on context switching and lock contention across hundreds of active connections

Answer

Larger pools automatically re-enable TLS for every query

Answer

Postgres switches to read-only mode above 20 connections

Answer

The application sends every query twice with a larger pool

Question 19

App latency is high with 'timeout waiting for connection' errors, but database CPU is about 15%. What is the most likely cause?

Accepted Answer

Connections are held too long, e.g. during an HTTP call inside a transaction

Answer

The database has hit its CPU ceiling and must be scaled up

Answer

The network between app and database lacks bandwidth

Answer

The database indexes are corrupted and must be rebuilt

Question 20

After moving an API to serverless functions, peak traffic causes 'too many connections' errors from the database. What is the standard fix?

Accepted Answer

Putting an external pooler like PgBouncer or RDS Proxy between functions and the database

Answer

Raising each function's memory so connections close faster

Answer

Opening a new connection for every query inside the function

Answer

Disabling the serverless platform's autoscaling entirely

Question 21

What limitation does using PgBouncer in transaction pooling mode introduce?

Accepted Answer

Session-level features like `SET`, advisory locks, or classic prepared statements are unreliable

Answer

No transaction can contain more than one query

Answer

All queries must be read-only

Answer

Clients cannot outnumber the database connections

Question 22

Workers consuming a queue mostly wait on an external API. What is the best autoscaling metric for them?

Accepted Answer

Queue depth or backlog per worker (or consumer lag)

Answer

Average CPU utilization of the workers

Answer

Memory usage of the message broker server

Answer

The number of users logged in to the website

Question 23

An autoscaler keeps adding and then removing instances every few minutes (flapping). Which setting reduces this?

Accepted Answer

A cooldown/stabilization window and different thresholds for scale-out and scale-in

Answer

Reducing the metric sampling interval to one second

Answer

Using the same threshold for scale-out and scale-in

Answer

Removing health checks from the load balancer

Question 24

A shop sees a 10x traffic jump every night at 21:00 when a sale starts, and instances take 4 minutes to boot. What is the best approach?

Accepted Answer

Scheduled/predictive scaling so capacity is ready before 21:00

Answer

Rely on reactive CPU-based autoscaling

Answer

Raise the CPU threshold for scale-out to 95%

Answer

Add instances only after receiving 503 errors

Question 25

For a report-generation endpoint that takes 2 minutes, which API pattern scales better?

Accepted Answer

Return `202 Accepted` with a job ID plus a status endpoint or webhook

Answer

Keep the HTTP request open for 2 minutes, then return `200`

Answer

Raise the load balancer timeout to 10 minutes

Answer

Run the report in the user's browser with JavaScript

Question 26

What does backpressure mean in a data processing pipeline?

Accepted Answer

A slow consumer signals the producer to slow down its sending rate

Answer

The producer sends faster so the queue drains

Answer

Old messages automatically return to the front of the queue

Answer

Compressing messages to reduce pressure on the network

Question 27

A producer collects messages until it has 500 or 20 ms pass, then sends. What is the effect?

Accepted Answer

Higher throughput at the cost of up to ~20 ms extra latency per message

Answer

Lower latency and lower throughput for all messages

Answer

Eliminating network overhead entirely at no cost

Answer

Guaranteed exactly-once delivery for all messages

Question 28

Which compression policy is most sensible for a website?

Accepted Answer

Pre-built Brotli for static assets, light gzip/Brotli for dynamic JSON, no recompressing JPEGs

Answer

Maximum Brotli level on the fly for every response including images

Answer

No compression, because CPU is always pricier than bandwidth

Answer

Compress only responses smaller than 100 bytes

Question 29

We denormalized the author's name into every comment row to avoid a join. What is the main cost of this decision?

Accepted Answer

Renaming an author must update all their comments, which may be temporarily inconsistent

Answer

Reading comments becomes slower because they hold more data

Answer

Indexes can no longer be built on the comments table

Answer

The database no longer allows concurrent writes to comments

Question 30

After adding read replicas, users sometimes see their just-edited profile with old values. What is a common fix?

Accepted Answer

Route a user's reads of their own data to the primary for a short time after a write

Answer

Remove all replicas and only scale up the primary

Answer

Disable the browser cache for every page

Answer

Switch replication from async to a daily cron job

Question 31

Why are LSM-tree-based databases like Cassandra or RocksDB well suited to write-heavy workloads?

Accepted Answer

Writes go to a memtable and sequential log first, and are merged in batches later

Answer

Every write immediately updates its place in an on-disk B-tree

Answer

These databases eliminate writing to disk entirely

Answer

Reads in them are always faster than writes

Question 32

A dashboard runs a heavy aggregation over millions of rows on every load, and the data can tolerate minutes of staleness. What is the best optimization?

Accepted Answer

Precompute results in a materialized view or summary table refreshed periodically

Answer

Run the same query on a read replica without any change

Answer

Increase the query timeout so users do not see errors

Answer

Add indexes on every column of the table

Question 33

A cache has a 90% hit ratio; a hit takes ~1 ms and a miss (database) ~100 ms. What is the approximate average latency?

Accepted Answer

About 11 ms

Answer

About 1 ms

Answer

About 50 ms

Answer

About 90 ms

Question 34

Why does a service's latency rise sharply as utilization approaches 100%?

Accepted Answer

Queueing delay grows non-linearly, roughly proportional to `1/(1−ρ)`

Answer

The CPU automatically halves its frequency at high utilization

Answer

The load balancer sends requests twice at high utilization

Answer

Latency stays constant at high utilization and only throughput drops

Question 35

Why are servers usually planned for a target peak utilization of about 60–70%, not 95%?

Accepted Answer

To leave headroom for spikes, instance failures, and autoscaling lag while keeping latency low

Answer

Because operating systems do not allow more than 70% CPU usage

Answer

Because cloud providers charge double above 70%

Answer

Because high utilization always causes data loss

Question 36

Peak is 12,000 RPS and load tests show each instance handles ~400 RPS at 70% CPU with acceptable latency. Tolerating loss of one of three AZs, roughly how many instances are needed?

Accepted Answer

About 45 instances

Answer

About 30 instances

Answer

About 20 instances

Answer

About 90 instances

Question 37

An app has 20M DAU; each user does 50 reads and 2 writes daily, and peak is 3x average. Roughly what is peak read QPS?

Accepted Answer

About 35,000

Answer

About 12,000

Answer

About 3,500

Answer

About 350,000

Question 38

50 million distinct objects averaging 2 KB are read daily, and per the 80/20 rule 20% of them get most traffic. Roughly how much cache memory is appropriate?

Accepted Answer

About 20 GB

Answer

About 100 GB

Answer

About 2 GB

Answer

About 200 GB

Question 39

A database is spread over 16 shards, but one shard is always at 100% CPU because a celebrity's post lives there. What is this called?

Accepted Answer

A hot key / hot partition; adding shards alone does not fix it

Answer

Split brain; two shards both think they are leader

Answer

Cold start; the shard has not warmed up yet

Answer

Write skew; two transactions wrote conflicting data

Question 40

A startup's database is nearing its limit. Which order of actions is usually most sensible?

Accepted Answer

Optimize queries/indexes, then cache and read replicas, then scale up, and shard last

Answer

Shard onto 64 nodes immediately, then review queries

Answer

Migrate entirely to a new NoSQL database before measuring anything

Answer

Drop all indexes to make writes faster

Question 41

Does keeping a local in-memory cache in each instance make a service no longer stateless?

Accepted Answer

No, as long as the cache is only an optimization and the source of truth lives elsewhere

Answer

Yes, any data in memory makes the service stateful

Answer

No, because local caches are always synchronized across instances

Answer

Yes, and that is why the load balancer must be sticky

Question 42

What is load shedding under overload?

Accepted Answer

Quickly rejecting some requests (e.g. with `503`) so the rest are served well

Answer

Accepting every request and putting it in an unbounded queue

Answer

Routing all requests to a single backup server

Answer

Shutting down the service entirely until the peak ends

Question 43

Why is `OFFSET 1000000 LIMIT 20` slow, and what is the scalable alternative?

Accepted Answer

The database must read and discard a million rows; keyset pagination with `WHERE id > ?` is better

Answer

`OFFSET` always locks the table; use `LIMIT` without `OFFSET`

Answer

`LIMIT 20` is too small; increase it to 1000

Answer

`OFFSET` is only fast in NoSQL databases; migrate

Question 44

In high-volume internal microservice traffic, what is the usual benefit of replacing JSON with Protobuf?

Accepted Answer

Smaller payloads and faster serialize/parse due to a binary, schema-based format

Answer

Completely removing the need for API versioning

Answer

Better human readability of messages in logs

Answer

Automatic encryption of all messages without TLS

Question 45

In a load test, raising concurrent users from 200 to 800 kept throughput flat while p99 quadrupled. What does this mean?

Accepted Answer

The system is saturated and extra concurrency just waits in queues

Answer

The system still has lots of spare capacity and load should go higher

Answer

The load tool is broken because throughput should have quadrupled

Answer

Network latency between client and server has quadrupled

Question 46

Asset traffic is 40 Gbps. If a CDN's hit ratio rises from 90% to 98%, how does origin traffic change?

Accepted Answer

From about 4 Gbps to about 0.8 Gbps, i.e. 5x less

Answer

From about 36 Gbps to about 39 Gbps, slightly more

Answer

From about 4 Gbps to about 3.2 Gbps, i.e. 8% less

Answer

It does not change because the origin always sees all traffic

Question 47

A service opens a new connection for every HTTPS call to another service. What does enabling keep-alive and connection reuse do?

Accepted Answer

It removes repeated TCP and TLS handshakes, reducing latency and CPU

Answer

Responses are cached automatically and no longer reach the server

Answer

Network bandwidth doubles

Answer

The need for a load balancer between services disappears

Question 48

Why can event-loop servers with non-blocking I/O (like Node.js or Nginx) handle thousands of concurrent connections with few threads?

Accepted Answer

No thread blocks while waiting on I/O; one thread switches between ready connections

Answer

Each connection runs on its own dedicated CPU core

Answer

These servers reject requests until load decreases

Answer

Heavy CPU computations are automatically parallelized in them

Question 49

A service is CPU-bound and you want to find which functions consume the most CPU time. What is the most suitable tool?

Accepted Answer

A sampling profiler, analyzed with a flame graph

Answer

Adding logs at the start and end of every function in production

Answer

Looking at average latency on the load balancer dashboard

Answer

Running a load test at twice the current traffic

Question 50

In a system with very diverse read patterns and fairly simple writes, what trade-off does separating the write model from optimized read models (CQRS) involve?

Accepted Answer

Each read model scales and is optimized independently, but eventual consistency and sync complexity are added

Answer

All reads become strongly consistent at no extra cost

Answer

The need for any database for writes is removed

Answer

Writes get faster because validation is no longer needed