Middle Backend Architecture Interview Questions (50 MCQs)

Question 1

What are the three states of a circuit breaker?

Accepted Answer

Closed, Open, and Half-Open

Answer

Active, Passive, and Standby

Answer

Pending, Running, and Done

Answer

Healthy, Degraded, and Dead

Question 2

In the Half-Open state, what happens if a trial request fails?

Accepted Answer

The breaker returns to Open and the cooldown timer restarts

Answer

The breaker goes Closed and all traffic flows

Answer

The calling service is restarted

Answer

The breaker stays Open forever until manual intervention

Question 3

What does the Bulkhead pattern do?

Accepted Answer

Partitions resources (thread or connection pools) per dependency so one failure can't exhaust all

Answer

Serves all requests from one shared pool to maximize utilization

Answer

Automatically retries failed requests

Answer

Distributes traffic across regions

Question 4

Why is jitter (randomization) usually added to exponential backoff?

Accepted Answer

So clients don't retry in lockstep and create synchronized waves (thundering herd)

Answer

So the total number of retries becomes unlimited

Answer

So each request's timeout gets shorter

Answer

So request processing order is preserved

Question 5

What is a good way to make retrying `POST /payments` safe?

Accepted Answer

The client sends an `Idempotency-Key` and the server stores and replays the result for that key

Answer

The server rejects every duplicate with a 500 error

Answer

Change the method to `GET` because GET is always safe

Answer

Make the client timeout infinite so retries are never needed

Question 6

Which is a common cause of cascading failure?

Accepted Answer

A slow downstream service filling up callers' threads that have no timeouts

Answer

Using structured logging in every service

Answer

Having a readiness probe on every pod

Answer

Deploying services with rolling updates

Question 7

Which is a good example of the fallback pattern?

Accepted Answer

Returning the last cached price when the pricing service is unavailable

Answer

Retrying forever until the service responds

Answer

Returning a bare 500 error to the user

Answer

Raising the timeout to several minutes

Question 8

What is the difference between a Layer 4 and a Layer 7 load balancer?

Accepted Answer

L4 routes on IP/port at TCP/UDP level; L7 inspects HTTP content like path and headers

Answer

L4 reads HTTP content; L7 only sees IPs

Answer

L7 is always faster than L4 because it has more information

Answer

L4 is only for databases and L7 only for caches

Question 9

When does least connections outperform round robin?

Accepted Answer

When request durations vary widely or connections are long-lived

Answer

When all requests have exactly the same cost

Answer

When there is only one server behind the balancer

Answer

When servers are not stateful

Question 10

Why should a liveness probe usually not check the database or downstream services?

Accepted Answer

Because a DB outage would restart every pod uselessly and worsen the incident

Answer

Because Kubernetes forbids database connections from probes

Answer

Because liveness runs only once at startup

Answer

Because checking the database always takes over 10 seconds

Question 11

What should a service do when it receives `SIGTERM` in Kubernetes?

Accepted Answer

Stop accepting new requests, finish in-flight ones, then exit

Answer

Exit immediately with `process.exit(0)`

Answer

Ignore the signal so Kubernetes gives up

Answer

Fail all in-flight requests with a 500

Question 12

Which symptom indicates the 'distributed monolith' anti-pattern?

Accepted Answer

Services must be deployed together in lockstep and changing one breaks others

Answer

Each service has its own database and independent pipeline

Answer

Services communicate asynchronously through events

Answer

Each team fully owns one service

Question 13

Why is the database-per-service pattern recommended?

Accepted Answer

So each service owns its data and can evolve its schema independently

Answer

Because smaller databases are always cheaper

Answer

To make JOINs across services' data easier

Answer

Because ACID transactions across services become automatic

Question 14

Which is the most sensible reason to extract a module from a monolith into its own service?

Accepted Answer

It has distinct scaling needs or an independent deploy cadence owned by a separate team

Answer

Because microservices are trendy

Answer

Because the module has more than 1000 lines

Answer

So a new framework can be tried

Question 15

How does the Strangler Fig pattern work for migrating off a legacy system?

Accepted Answer

A routing layer fronts the legacy system and features are moved to the new system incrementally

Answer

The legacy system is switched off and replaced overnight

Answer

Legacy code is translated line by line into a new language

Answer

Both systems run in parallel forever with no migration

Question 16

What is the purpose of an Anti-Corruption Layer (ACL)?

Accepted Answer

Translating between an external/legacy model and our domain model so ours stays clean

Answer

Preventing SQL injection in inputs

Answer

Encrypting data between services

Answer

Automatically deleting corrupted data from the database

Question 17

What is the Dependency Rule in Clean Architecture?

Accepted Answer

Source code dependencies must point only inward, toward the domain

Answer

The domain layer should depend on the framework and ORM

Answer

Any layer may depend on any other layer

Answer

Dependencies must point only outward, toward the UI

Question 18

In hexagonal architecture, which is a driven (secondary) adapter?

Accepted Answer

A PostgreSQL implementation of `OrderRepository`

Answer

A REST controller that invokes a use case

Answer

A CLI command that places an order

Answer

A message consumer that triggers a use case

Question 19

In DDD, what is the role of an Aggregate Root?

Accepted Answer

The single entry point to an aggregate that enforces its invariants

Answer

The main database table every other table references

Answer

A service that manages every aggregate in the system

Answer

A base class all entities inherit from

Question 20

Per common DDD guidance, how many aggregates should one transaction modify?

Accepted Answer

One aggregate; coordinate across aggregates via domain events and eventual consistency

Answer

All related aggregates in one large transaction

Answer

At least two aggregates to preserve consistency

Answer

None; aggregates are never persisted

Question 21

What is a Domain Event?

Accepted Answer

A record of something meaningful that happened in the domain, like `OrderPlaced`

Answer

A command to do something in the future, like `PlaceOrder`

Answer

A user's click event in the browser

Answer

An error recorded in the system log

Question 22

How does the Backend for Frontend (BFF) pattern differ from a general API Gateway?

Accepted Answer

A dedicated backend per client type (web, mobile) shaping data for its needs

Answer

A BFF only serves static frontend files

Answer

A BFF replaces the database for the frontend

Answer

A BFF is the same as an L4 load balancer

Question 23

What is the difference between client-side and server-side service discovery?

Accepted Answer

Client-side: the client queries the registry and picks an instance; server-side: a router/LB does it

Answer

Client-side is only for browsers and server-side only for backends

Answer

In server-side, each service hard-codes addresses

Answer

There is no difference

Question 24

What mechanism does a service mesh (like Istio or Linkerd) typically use?

Accepted Answer

A sidecar proxy next to each service handling mTLS, retries, and telemetry

Answer

A library that must be imported into each service's code

Answer

A central database storing every request

Answer

Replacing HTTP with a proprietary protocol

Question 25

What is OpenTelemetry?

Accepted Answer

A vendor-neutral standard and SDKs for producing and exporting traces, metrics, and logs

Answer

A time-series database for storing metrics

Answer

A graphical dashboard for viewing logs

Answer

A protocol for service telephony

Question 26

In distributed tracing, how do traces and spans relate?

Accepted Answer

A trace is a request's full journey, composed of nested spans for each operation

Answer

A span is the full journey and a trace is one operation

Answer

Traces are for errors and spans for successful requests

Answer

They are synonyms

Question 27

Which metrics does the RED method recommend for monitoring services?

Accepted Answer

Request Rate, Errors, and Duration

Answer

RAM, Ethernet, and Disk

Answer

Retries, Events, and Deployments

Answer

Readiness, Exceptions, and Deadlocks

Question 28

What is the USE method designed to analyze?

Accepted Answer

Resources (CPU, memory, disk) by Utilization, Saturation, and Errors

Answer

Frontend user experience

Answer

Code quality via Unit tests, Style, and Errors

Answer

Monthly cloud infrastructure cost

Question 29

With a 99.9% availability SLO over 30 days, roughly how big is the error budget?

Accepted Answer

About 43 minutes

Answer

About 4.3 minutes

Answer

About 7.2 hours

Answer

About 1 day

Question 30

Why is putting `userId` as a label on Prometheus metrics a bad idea?

Accepted Answer

High cardinality explodes the number of time series and memory usage

Answer

Prometheus accepts only numeric labels

Answer

Prometheus deletes labels after an hour

Answer

Labels cause metrics to be encrypted

Question 31

What problem does consumer-driven contract testing (e.g., with Pact) solve?

Accepted Answer

Catching breaking API changes between services without full e2e runs

Answer

Load-testing a service under high traffic

Answer

Checking legal contracts with customers

Answer

Generating test data in the database

Question 32

Which is an example of a fake (as a test double)?

Accepted Answer

An `InMemoryUserRepository` that really stores and retrieves data in a Map

Answer

A function that always returns `null` with no logic

Answer

An object that only verifies `save` was called twice

Answer

A parameter passed in but never used

Question 33

What is a spy in testing?

Accepted Answer

A test double that records calls so you can assert on them afterwards

Answer

A tool for checking code security

Answer

An in-memory database implementation

Answer

A test that runs in production

Question 34

What is the main long-term risk of feature flags?

Accepted Answer

Accumulating stale flags that multiply code paths and test combinations

Answer

Growing the database to terabytes

Answer

Making rollbacks impossible

Answer

Weakening TLS security

Question 35

Which approach renames a column without downtime during rolling deployments?

Accepted Answer

Expand/contract: add the new column, dual-write, migrate, then drop the old one in a later release

Answer

Rename the column directly at the same time as deploying new code

Answer

Stop all instances, rename, then restart

Answer

Drop the old column before deploying the new code

Question 36

What is the best way to manage secrets (like API keys) in production?

Accepted Answer

Using a secret manager (like Vault or AWS Secrets Manager) with scoped access and rotation

Answer

Committing the `.env` file to a private repository

Answer

Baking secrets into the Docker image

Answer

Sharing secrets in the team Slack for quick access

Question 37

In multi-tenancy, which model offers the most isolation and the highest operational cost?

Accepted Answer

Database-per-tenant

Answer

Shared schema with a `tenant_id` column

Answer

Schema-per-tenant in a shared database

Answer

Storing the tenant in the user's cookie

Question 38

What is the noisy neighbor problem in a multi-tenant system and how is it mitigated?

Accepted Answer

One heavy tenant hogs shared resources; mitigated with per-tenant quotas and rate limits

Answer

One tenant's logs appear in another's; fixed by deleting logs

Answer

Servers have noisy fans; fixed with better cooling

Answer

Tenants can't message each other; fixed with a message queue

Question 39

Which workload is the poorest fit for serverless (like AWS Lambda)?

Accepted Answer

Long-running processing with steady high load and strict, consistent low latency

Answer

Processing an image after a file upload

Answer

Webhooks with sporadic, low traffic

Answer

Lightweight daily cron jobs

Question 40

What is a common problem when serverless functions connect to a relational database?

Accepted Answer

Thousands of concurrent instances can exhaust the database's connection limit

Answer

Serverless functions can't use TCP

Answer

Relational databases are read-only from serverless

Answer

Queries always run twice in serverless

Question 41

Why should background worker jobs be designed to be idempotent?

Accepted Answer

Because most queues are at-least-once and a job may run more than once

Answer

Because workers never fail

Answer

Because idempotent jobs are always faster

Answer

Because queues never deliver messages in order

Question 42

Which problem does the Transactional Outbox pattern solve?

Accepted Answer

Dual write: saving to the DB and publishing an event aren't atomic

Answer

Slow SELECT queries

Answer

Sending email to the user's Outbox folder

Answer

Message size limits in queues

Question 43

A service receives 200 requests/second with 50ms average latency. By Little's law, how many requests are in flight on average?

Accepted Answer

10

Answer

4

Answer

250

Answer

4000

Question 44

An app has 1 million daily active users, each making 10 requests per day. Roughly what is the average QPS?

Accepted Answer

About 120

Answer

About 12

Answer

About 1,200

Answer

About 10,000

Question 45

A request fans out in parallel to 100 services, each with a 1% chance of a slow (above p99) response. Roughly how likely is the overall request to be slow?

Accepted Answer

About 63%

Answer

About 1%

Answer

About 10%

Answer

Exactly 100%

Question 46

What is the distinctive property of the token bucket rate-limiting algorithm?

Accepted Answer

It allows bursts up to bucket capacity while capping the average rate

Answer

It makes output perfectly smooth and accepts no bursts

Answer

It only counts requests in fixed calendar windows

Answer

It randomly accepts or rejects each request

Question 47

A service has 5 instances, each with an in-memory per-user limit of 100 req/min. What is the problem?

Accepted Answer

A user can effectively send up to 500 req/min; the counter should live in a shared store like Redis

Answer

No problem; the limit is enforced at exactly 100

Answer

The limit drops to 20 req/min

Answer

The load balancer rejects all requests

Question 48

Per 12-Factor, how should an app emit logs?

Accepted Answer

As an event stream to `stdout`; collection and routing are the environment's job

Answer

Into rotating files on local disk managed by the app

Answer

Directly into the app's main database

Answer

Only on errors, by sending emails

Question 49

What does the Disposability factor in 12-Factor require?

Accepted Answer

Fast startup and graceful shutdown so processes can be created or killed at any moment

Answer

Automatically deleting user data after 30 days

Answer

Using cheap disposable servers

Answer

Clearing logs after each deploy

Question 50

A request synchronously passes through 3 services, each with 99.9% availability. What's the approximate overall availability?

Accepted Answer

About 99.7%

Answer

Exactly 99.9%

Answer

About 99.99%

Answer

About 97%