Senior Backend Architecture Interview Questions (50 MCQs)

Question 1

What is the main problem with Two-Phase Commit (2PC) in a microservices architecture?

Accepted Answer

It's blocking: if the coordinator fails after prepare, participants hold their locks

Answer

It provides no consistency and always yields inconsistent data

Answer

It only works on NoSQL databases

Answer

It requires no coordination between services

Question 2

What is the difference between choreography and orchestration sagas?

Accepted Answer

Choreography: services react to each other's events; orchestration: a central orchestrator commands each step

Answer

Choreography uses 2PC and orchestration uses distributed locks

Answer

Orchestration doesn't need compensating transactions

Answer

Choreography only works with synchronous services

Question 3

In a saga, what is a compensating transaction?

Accepted Answer

A semantic action that undoes a committed step's effect, like a refund for a charge

Answer

A database rollback that physically reverts all previous steps

Answer

A transaction that offsets infrastructure costs

Answer

An automatic retry of the failed step

Question 4

Which ACID property do sagas inherently lack, requiring countermeasures?

Accepted Answer

Isolation

Answer

Durability

Answer

Atomicity of each local transaction

Answer

None; sagas are fully ACID

Question 5

Why is claiming 'our distributed system is CA' usually misleading?

Accepted Answer

Because network partitions are unavoidable, and when one occurs you must choose C or A

Answer

Because CA only applies to NoSQL databases

Answer

Because CAP only concerns single-node systems

Answer

Because Availability in CAP means monthly uptime

Question 6

What does PACELC add to CAP?

Accepted Answer

Even without a partition (Else), you trade off Latency versus Consistency

Answer

It removes Partition as an optional property

Answer

It adds Performance and Cost as a third axis

Answer

It proves all three CAP properties are achievable together

Question 7

What does linearizability (the strongest form of strong consistency) guarantee?

Accepted Answer

Each operation appears to take effect atomically at a point between its start and end, and all later reads see it

Answer

All replicas eventually converge

Answer

Each client sees its own writes but not others'

Answer

One client's operations apply in order but there's no global order

Question 8

A user edits their profile but sees the old value after refresh because the read hit a replica. Which guarantee is violated?

Accepted Answer

Read-your-writes

Answer

Global linearizability for all users

Answer

Durability

Answer

Partition tolerance

Question 9

What is the practical approach to achieving 'exactly-once' effects in message processing?

Accepted Answer

At-least-once delivery combined with idempotent processing or deduplication

Answer

At-most-once delivery and hoping the network doesn't fail

Answer

Using TCP because it guarantees exactly-once delivery

Answer

Raising timeouts so retries are never needed

Question 10

In a 4-layer call chain where each layer makes up to 3 attempts, how many requests can the bottom service receive in the worst case, and what's the fix?

Accepted Answer

81 requests; retry at only one layer and use a retry budget

Answer

12 requests; increase retries per layer

Answer

3 requests; no change needed

Answer

4 requests; remove timeouts

Question 11

Why should an overloaded service reject some requests early (load shedding)?

Accepted Answer

So accepted requests finish in reasonable time instead of all of them timing out

Answer

Because rejecting requests automatically adds capacity

Answer

So clients stop using the service forever

Answer

Because the load balancer doesn't work without it

Question 12

What is the benefit of deadline propagation (as gRPC does)?

Accepted Answer

Downstream services know the remaining time and stop useless work once the deadline passes

Answer

Each service keeps its own fixed timeout regardless of the caller

Answer

Deadlines make requests never time out

Answer

It's only used to log request time

Question 13

What is a hedged request and what does it cost?

Accepted Answer

If no reply by about p95, send a second request to another replica and use the first reply; adds a little load

Answer

Every request always goes to all replicas; it's free

Answer

Requests retry forever until success; it increases latency

Answer

Requests are batched and sent together; it lowers throughput

Question 14

What does Queue-Based Load Leveling gain, and what does it sacrifice?

Accepted Answer

Absorbs spikes in a queue and feeds consumers at a steady rate; costs higher processing latency and async responses

Answer

Reduces latency to zero but lowers throughput

Answer

Eliminates the need for any storage

Answer

Guarantees strong consistency but removes availability

Question 15

Why use consistent hashing instead of `hash(key) % N` for routing to stateful instances (like local caches)?

Accepted Answer

Changing N remaps only about `1/N` of keys rather than nearly all

Answer

Because consistent hashing is cryptographically stronger

Answer

Because `%` is slow in programming languages

Answer

Because it always perfectly balances load without virtual nodes

Question 16

What is the Inverse Conway Maneuver?

Accepted Answer

Deliberately shaping teams to match the desired architecture, since systems mirror org communication

Answer

Designing the architecture as the exact opposite of the org structure

Answer

Removing all teams to form a single team

Answer

Converting microservices into a monolith

Question 17

Which symptom indicates overly fine-grained services (nanoservices)?

Accepted Answer

A simple use case needs dozens of chatty network calls between services

Answer

Each service owns a whole bounded context

Answer

Services can be deployed independently

Answer

Services use under 10% CPU

Question 18

Why is sharing a common domain-model library across many microservices risky?

Accepted Answer

It creates hidden coupling; changing the shared model forces coordinated deploys of all services

Answer

Because libraries can't be used in microservices

Answer

Because it doubles container image size

Answer

Because it makes services compile slower

Question 19

The Order service synchronously fetches Customer data per request, so Customer outages take Order down. Which pattern reduces this temporal coupling?

Accepted Answer

Event-Carried State Transfer: keeping a local copy of needed Customer data via events

Answer

Raising the Customer call timeout to 60 seconds

Answer

Merging the Order and Customer databases into one

Answer

Calling Customer twice to be sure

Question 20

When is CQRS worth its complexity?

Accepted Answer

When read and write patterns differ greatly and need separate models or scaling

Answer

For every simple CRUD app from day one

Answer

Only when the team has fewer than three people

Answer

When strong read-after-write consistency is mandatory

Question 21

Which is a real challenge of event sourcing?

Accepted Answer

Evolving old event schemas (versioning/upcasting) and needing snapshots for long streams

Answer

Losing the history of changes

Answer

Being unable to rebuild state from events

Answer

Being unable to audit changes

Question 22

What's the biggest challenge of an active-active multi-region architecture accepting writes in both regions?

Accepted Answer

Resolving conflicting concurrent writes to the same data (e.g., LWW, CRDTs, or a home region per key)

Answer

Being unable to use DNS

Answer

Only one region can serve reads

Answer

Needing a single database with no replication

Question 23

Why do consensus-based clusters (like etcd) usually have an odd number of nodes (3 or 5)?

Accepted Answer

For majority quorums; an extra even node adds no fault tolerance, and quorums prevent split-brain

Answer

Because Raft won't compile with an even count

Answer

Because odd counts reduce hardware cost

Answer

Because one node must always stay off

Question 24

A worker takes a distributed lock, hits a long GC pause, its lease expires, and another worker takes the lock. What prevents the first worker's stale write?

Accepted Answer

A fencing token: a monotonically increasing lock number; storage rejects writes with an older token

Answer

Raising the lock TTL to several hours

Answer

Checking the system clock before writing

Answer

Acquiring two locks instead of one

Question 25

Why is ordering events from different services by wall-clock timestamps dangerous?

Accepted Answer

Clock skew and NTP adjustments make real-time order and causality unreliable

Answer

Because timestamps aren't always stored in UTC

Answer

Because timestamps increase message size

Answer

Because servers don't have clocks

Question 26

What is the advantage of tail-based over head-based sampling in tracing?

Accepted Answer

The decision is made after the trace completes, so all erroneous or slow traces can be kept

Answer

It has no buffering or memory cost

Answer

The decision is made at the start of the request and is simpler

Answer

It stores every trace without exception

Question 27

Why is burn-rate (error-budget consumption) alerting better than a fixed error-rate threshold alert?

Accepted Answer

It catches severe incidents fast and slow leaks too, while reducing noisy alerts

Answer

Because it never produces false positives

Answer

Because it requires no SLO

Answer

Because it only works on CPU

Question 28

Per SRE practice, what should paging the on-call primarily be based on?

Accepted Answer

User-visible symptoms like error rate or latency against the SLO

Answer

Any time a server's CPU exceeds 70%

Answer

Every exception logged

Answer

Every pod restart

Question 29

What is coordinated omission in load testing?

Accepted Answer

The load tool delays sending while the server is slow, under-reporting true tail latency

Answer

The team deliberately dropping failed tests from reports

Answer

Coordinating multiple load tools to send simultaneously

Answer

Ignoring static asset requests in tests

Question 30

Ten instances each report their own p99. Is the average of those ten numbers the service-wide p99?

Accepted Answer

No; percentiles can't be averaged and must be computed from aggregated histograms

Answer

Yes; the mean of p99s is exactly the overall p99

Answer

Yes, as long as instances are equally sized

Answer

No; you should report the maximum as the overall p99

Question 31

As a service's utilization approaches 100%, how does latency behave?

Accepted Answer

It rises sharply and non-linearly as queueing delay explodes

Answer

It rises linearly and gently

Answer

It stays flat until exactly 100%

Answer

It drops because caches are warmer

Question 32

For an I/O-bound service that mostly waits on other services, what's wrong with autoscaling on CPU?

Accepted Answer

CPU stays low while concurrency and latency soar; scale on in-flight requests or queue depth instead

Answer

CPU is always 100% so it scales infinitely

Answer

Kubernetes doesn't support autoscaling I/O-bound services

Answer

No problem; CPU is the best signal for every service

Question 33

A new release included a destructive migration (dropping a column) and has a bug. Why isn't a simple code rollback enough?

Accepted Answer

The old code needs the dropped column; destructive schema changes must be decoupled and done after stabilization

Answer

Because Kubernetes doesn't support rollbacks

Answer

Because rollbacks always restore data automatically

Answer

Because old code is faster than new code

Question 34

What is shadow traffic (traffic mirroring) useful for when rewriting a service, and what precaution is needed?

Accepted Answer

Copying real traffic to the new version and comparing responses; side effects like payments or emails must be disabled

Answer

Moving all users to the new version with no way back

Answer

Hiding traffic from the monitoring system

Answer

Sending traffic only at night

Question 35

Which describes the correct approach to chaos engineering?

Accepted Answer

Hypothesize steady state, inject controlled faults with a small blast radius, compare against a control

Answer

Randomly killing production servers without notice or monitoring

Answer

Running only unit tests with random data

Answer

Removing all resilience patterns to see what happens

Question 36

What benefit does cell-based architecture offer a large multi-tenant SaaS?

Accepted Answer

Each cell is an independent full copy of the stack serving a subset of tenants, limiting blast radius

Answer

All tenants share one database to simplify management

Answer

It completely removes the need for routing

Answer

Each request is sent to every cell

Question 37

A Lambda scales to a thousand instances during a spike and overwhelms a legacy downstream API. What's an appropriate architectural fix?

Accepted Answer

Cap the function's concurrency (reserved concurrency) and put a queue between it and the API

Answer

Increase function memory to make it faster

Answer

Remove the function timeout

Answer

Turn all calls into immediate retries

Question 38

During a strangler-fig migration, both the legacy system and the new service need customer data. Which approach is common to keep data in sync during the transition?

Accepted Answer

Change Data Capture (e.g., Debezium) from the source DB, applying changes to the other

Answer

Manually copying data with a monthly script

Answer

Having the new service query legacy internal tables forever

Answer

Writing to both databases from code with no guarantees

Question 39

In context mapping, what's the difference between a Conformist relationship and an Anti-Corruption Layer?

Accepted Answer

Conformist adopts the upstream model as-is; an ACL translates it into our own model

Answer

Conformist translates and ACL adopts the model as-is

Answer

Both mean two teams share one model

Answer

Conformist is only for events and ACL only for REST

Question 40

In hexagonal/clean architecture, where is the transaction boundary typically defined?

Accepted Answer

In the application service (use case), via a port like Unit of Work without ORM coupling

Answer

Inside domain entities by calling `BEGIN` directly

Answer

In the HTTP controller

Answer

Separately in every repository method

Question 41

`Product` is modeled as one aggregate containing all its `Review`s and saved with optimistic locking. What goes wrong under high traffic?

Accepted Answer

Frequent version conflicts since every new review modifies the whole aggregate; Review should be its own aggregate

Answer

No problem because optimistic locking always succeeds

Answer

The database automatically deletes reviews

Answer

A larger aggregate is always faster due to fewer joins

Question 42

A service uses DNS for discovery but keeps connecting to old IPs after instances are replaced. What's the most likely cause?

Accepted Answer

DNS cached in the runtime/client beyond TTL, or long-lived connections that never re-resolve

Answer

DNS inherently doesn't support IPv4

Answer

L7 load balancers disable DNS

Answer

New instances always get the same IP as before

Question 43

Why is gRPC traffic behind an L4 load balancer often unevenly distributed?

Accepted Answer

gRPC uses long-lived multiplexed HTTP/2 connections, and L4 balances connections, not requests

Answer

gRPC only runs over UDP

Answer

L4 balancers can't connect to ports above 1024

Answer

Protobuf is unreadable by the balancer so requests are rejected

Question 44

Which is a common anti-pattern in API Gateway usage?

Accepted Answer

Putting business logic and domain orchestration in the gateway, making it a shared bottleneck for teams

Answer

Applying rate limiting in the gateway

Answer

Terminating TLS in the gateway

Answer

Validating tokens in the gateway

Question 45

Two requests with the same `Idempotency-Key` arrive concurrently and both see 'key not found'. How do you prevent double execution?

Accepted Answer

Atomically insert the key as `in-progress` via a unique constraint or `SET NX` before executing

Answer

Checking the key twice in a row

Answer

Storing the key after the operation completes

Answer

Adding a random delay before checking the key

Question 46

A service peaks at 1000 req/s, each running one 20ms query. Roughly how large should the DB connection pool be across all instances, and why is a huge pool bad?

Accepted Answer

About 20 plus headroom; an oversized pool causes contention and context switching in the database

Answer

About 20,000; bigger pools always yield more throughput

Answer

Exactly 1; databases support only one connection

Answer

About 1000; one connection per request per second

Question 47

10 million files averaging 1MB are uploaded daily and stored with a replication factor of 3. Roughly how much raw storage is needed per year?

Accepted Answer

About 11 PB

Answer

About 3.6 TB

Answer

About 30 TB

Answer

About 110 PB

Question 48

Why should dynamic config changes (e.g., via a config service) be rolled out gradually like code?

Accepted Answer

Because one bad config can take down the whole fleet at once; many outages stem from config changes

Answer

Because config never affects system behavior

Answer

Because config is only read at restart

Answer

Because gradual rollout encrypts the config

Question 49

In a modular monolith, how do you prevent module boundaries from eroding over time?

Accepted Answer

Export only each module's public API and enforce dependency rules with architecture tests or lint in CI

Answer

Let all modules access each other's tables directly

Answer

Document boundaries in a wiki and rely on code review

Answer

Put all code in a single folder for simplicity

Question 50

In which scenario is redrawing service boundaries to avoid a distributed transaction the best option?

Accepted Answer

When two services constantly must uphold a shared invariant atomically and immediately

Answer

When services have fully separate teams and domains and eventual consistency is acceptable

Answer

When only one service is involved in the operation

Answer

When the operation is read-only