Junior Reliability & Fault Tolerance Interview Questions (50 MCQs)

Question 1

In system design, how is a service's 'availability' usually defined?

Accepted Answer

The percentage of time the service successfully serves requests

Answer

The average time the service takes to answer each request

Answer

The number of requests the service processes per second

Answer

The number of servers the service is deployed on

Question 2

Roughly how much downtime per year does `99.9%` (three nines) availability allow?

Accepted Answer

About 8.76 hours

Answer

About 52.6 minutes

Answer

About 3.65 days

Answer

About 5.26 minutes

Question 3

Roughly how much annual downtime does `99.99%` availability allow?

Accepted Answer

About 52 minutes

Answer

About 8 hours

Answer

About 5 minutes

Answer

About 26 seconds

Question 4

Roughly how much downtime does a `99.9%` SLO allow in a 30-day month?

Accepted Answer

About 43 minutes

Answer

About 4.3 minutes

Answer

About 7.2 hours

Answer

About 2 hours

Question 5

What is a 'Single Point of Failure' (SPOF)?

Accepted Answer

A component whose failure alone takes down the whole system

Answer

The component with the highest CPU usage in the system

Answer

The first service a user's request reaches

Answer

A server that only has one version of the code deployed

Question 6

What is the main goal of redundancy in a system?

Accepted Answer

Keeping the system working when one component fails

Answer

Lowering infrastructure cost by removing extra servers

Answer

Simplifying the architecture by reducing service count

Answer

Guaranteeing all requests are answered under 10ms

Question 7

In an active-passive setup, what does the passive node usually do?

Accepted Answer

It stays ready and takes over traffic only when the active fails

Answer

It always serves half of the traffic alongside the active

Answer

It serves only read requests and rejects all writes

Answer

It takes a backup of the active and then shuts down

Question 8

What is the key characteristic of an active-active setup?

Accepted Answer

All nodes serve real traffic at the same time

Answer

One node works while others only wait for it to fail

Answer

Nodes switch on in turns on a fixed schedule

Answer

Each node runs only a different version of the app

Question 9

What does 'failover' refer to?

Accepted Answer

Switching work to a healthy component after a failure

Answer

Clearing error logs after an incident is resolved

Answer

Adding more servers when traffic goes up

Answer

Reverting code to the previous version after a bad deploy

Question 10

What does a load balancer use health checks for?

Accepted Answer

Detecting unhealthy instances and stopping traffic to them

Answer

Measuring the internet speed of end users

Answer

Encrypting traffic between client and server

Answer

Compressing responses to reduce bandwidth

Question 11

In Kubernetes, what happens when a readiness probe fails?

Accepted Answer

The pod is removed from Service endpoints but not restarted

Answer

The container is killed and restarted immediately

Answer

The whole node is removed from the cluster and drained

Answer

The deployment is rolled back to the previous version

Question 12

Why should every network call to another service have a timeout?

Accepted Answer

So a slow dependency can't hold threads and connections forever

Answer

So requests reach the destination server faster than usual

Answer

So the need to retry requests disappears completely

Answer

So responses are automatically cached on the client

Question 13

Which error is usually worth retrying?

Accepted Answer

`503 Service Unavailable`

Answer

`400 Bad Request`

Answer

`401 Unauthorized`

Answer

`404 Not Found`

Question 14

What does 'exponential backoff' mean for retries?

Accepted Answer

The wait between attempts roughly doubles on each retry

Answer

The number of retries grows exponentially with each error

Answer

All retries are sent immediately and in parallel

Answer

The wait between attempts is always a fixed value

Question 15

Why add 'jitter' (randomness) to exponential backoff?

Accepted Answer

So many clients don't retry in sync and create traffic waves

Answer

So each retry is guaranteed to be sent faster than the last

Answer

So request ordering on the server is always preserved

Answer

So there's no need to set a maximum retry count

Question 16

What is the purpose of an 'idempotency key' in a payments API?

Accepted Answer

Preventing an operation from running twice when a request is retried

Answer

Authenticating the user without needing a password

Answer

Encrypting card details while in transit over the network

Answer

Setting the request's priority in the server's queue

Question 17

What problem does the circuit breaker pattern solve?

Accepted Answer

Continuing to call a dependency that keeps failing

Answer

Persisting data durably during a datacenter power cut

Answer

Splitting traffic evenly across several servers

Answer

Encrypting communication between microservices

Question 18

What are the three standard states of a circuit breaker?

Accepted Answer

Closed, Open, Half-Open

Answer

Active, Passive, Standby

Answer

Healthy, Degraded, Down

Answer

Pending, Running, Stopped

Question 19

Where does the 'bulkhead' pattern come from and what does it do?

Accepted Answer

From ship compartments; it isolates resources so one part's failure doesn't spread

Answer

From water dams; it stores incoming traffic in one big shared queue

Answer

From electrical fuses; it shuts the whole service down after errors

Answer

From bridges; it alternates requests between two datacenters

Question 20

Which is an example of 'graceful degradation'?

Accepted Answer

Showing the page without recommendations when that service is down

Answer

Returning a `500` for the whole page when recommendations are down

Answer

Restarting the entire system whenever a dependency breaks

Answer

Keeping the user on a loading screen until the fault is fixed

Question 21

What is the main goal of rate limiting in an API?

Accepted Answer

Protecting the service from overuse or abuse

Answer

Speeding up responses for users who send many requests

Answer

Guaranteeing that no request is ever rejected

Answer

Shrinking responses by removing extra fields

Question 22

Which status code fits when a client exceeds the rate limit?

Accepted Answer

`429 Too Many Requests`

Answer

`403 Forbidden`

Answer

`500 Internal Server Error`

Answer

`408 Request Timeout`

Question 23

What does the `Retry-After` header in a `429` or `503` response tell the client?

Accepted Answer

How long to wait before sending the request again

Answer

How many more times it is allowed to retry

Answer

The address of another server to send requests to

Answer

The API version that it should use instead

Question 24

In the token bucket algorithm, what allows short bursts?

Accepted Answer

Tokens saved up in the bucket up to its capacity

Answer

Sending requests to the server at a perfectly fixed rate

Answer

Resetting the counter at the start of every minute

Answer

Keeping the timestamp of every previous request

Question 25

What is the well-known weakness of the fixed window counter algorithm?

Accepted Answer

At a window boundary, a client can send up to double the limit

Answer

It needs memory per user proportional to the number of requests

Answer

It cannot be implemented without a central queue

Answer

It always processes requests with a fixed delay

Question 26

What does 'load shedding' mean?

Accepted Answer

Deliberately rejecting some requests during overload so the rest are served well

Answer

Moving all traffic to a more powerful server at peak load

Answer

Deleting old data from the database to free up space

Answer

Splitting one large request into several smaller ones

Question 27

What does 'backpressure' mean in a system?

Accepted Answer

A slow consumer signals the producer to slow down production

Answer

The producer sends data faster so the queue empties sooner

Answer

Failed messages are pushed back to the head of the queue

Answer

The server sends responses in reverse order of arrival

Question 28

What is a 'cascading failure'?

Accepted Answer

One component's failure overloading and toppling others in a chain

Answer

A failure that only happens in staging and never in production

Answer

Several back-to-back deploys that each add a new bug

Answer

An error visible only in logs with no user impact

Question 29

What does RTO (Recovery Time Objective) specify?

Accepted Answer

The maximum acceptable time to restore service after a failure

Answer

The maximum amount of data loss that is acceptable

Answer

The time interval between two consecutive backups

Answer

The service's average response time under normal conditions

Question 30

If a database is backed up every 6 hours with no replication, what is the worst-case RPO?

Accepted Answer

About 6 hours of data

Answer

Zero; no data is lost at all

Answer

About 3 hours of data

Answer

About 12 hours of data

Question 31

What does the well-known '3-2-1' backup rule say?

Accepted Answer

3 copies of data, on 2 different media, with 1 copy offsite

Answer

3 backups a day, kept 2 weeks, with 1 restore test a year

Answer

3 datacenters, 2 regions, and 1 primary database

Answer

3 synchronous replicas, 2-hour RPO, and 1-hour RTO

Question 32

Why is periodically testing backup restores important?

Accepted Answer

Because a backup that was never restored may be broken or incomplete

Answer

Because restoring makes the backup files more compressed

Answer

Because without restores, backups are auto-deleted after a month

Answer

Because restoring replaces the need for replication

Question 33

What does deploying a service across multiple Availability Zones (multi-AZ) protect against?

Accepted Answer

The failure of a single datacenter or zone within a region

Answer

A complete outage of an entire cloud region

Answer

Software bugs in the application code

Answer

Accidental deletion of data by a user

Question 34

Why is it said that 'replication is not a backup'?

Accepted Answer

Because accidental deletes or corruption get copied to replicas too

Answer

Because replicas are always several days behind the primary

Answer

Because replication only works on NoSQL databases

Answer

Because replicas can't serve any read requests

Question 35

What is chaos engineering?

Accepted Answer

Deliberately injecting controlled failures to find system weaknesses

Answer

Writing code without clear structure to develop faster

Answer

Deploying changes to production without any testing

Answer

Generating random traffic to test the app's UI

Question 36

What happens in a canary deployment?

Accepted Answer

The new version first gets a small share of traffic and is monitored

Answer

The new version replaces all servers at the same time

Answer

The new version runs only in the staging environment

Answer

The new version is visible only to the dev team

Question 37

What is the main benefit of blue-green deployment?

Accepted Answer

Fast rollback by switching traffic back to the old environment

Answer

Needing no extra servers at all during a deploy

Answer

Completely removing the need for pre-release testing

Answer

Running two versions in parallel forever

Question 38

What reliability benefit do feature flags provide?

Accepted Answer

Turning off a faulty feature without redeploying

Answer

Automatically adding servers when a feature turns on

Answer

Removing the need for code review on new features

Answer

Encrypting new feature code in the final bundle

Question 39

When the error rate spikes right after a deploy, what is usually the best first action?

Accepted Answer

Roll back to the previous version, then investigate

Answer

Debug directly on the production servers

Answer

Write a quick fix and deploy it without testing

Answer

Wait for the errors to resolve by themselves

Question 40

What are the classic three pillars of observability?

Accepted Answer

Metrics, logs, and traces

Answer

CPU, memory, and disk

Answer

Alerts, dashboards, and runbooks

Answer

Unit tests, integration tests, and e2e tests

Question 41

What are the four 'golden signals' from Google's SRE book?

Accepted Answer

Latency, traffic, errors, saturation

Answer

Uptime, downtime, MTTR, MTBF

Answer

CPU, memory, network, disk

Answer

Logs, metrics, traces, events

Question 42

What is the difference between an SLA and an SLO?

Accepted Answer

An SLA is a customer contract with penalties; an SLO is an internal target

Answer

An SLA is an internal target; an SLO is a customer contract

Answer

An SLA measures only latency; an SLO measures only availability

Answer

There is no difference; they are two names for one idea

Question 43

What is an 'error budget'?

Accepted Answer

The allowed unreliability derived from the SLO, e.g. `0.1%` for `99.9%`

Answer

The team's financial budget for buying monitoring tools

Answer

The number of bugs QA may report in each sprint

Answer

The maximum number of errors a log file can hold

Question 44

If a request must pass through two services in series, each with `99.9%` availability, what is the overall availability?

Accepted Answer

About `99.8%`

Answer

Still `99.9%`

Answer

About `99.99%`

Answer

About `99.0%`

Question 45

Two independent parallel replicas each have `99%` availability, and one is enough. What is the overall availability?

Accepted Answer

`99.99%`

Answer

`98.01%`

Answer

`99.5%`

Answer

`99.9%`

Question 46

How is availability computed from MTBF and MTTR?

Accepted Answer

`MTBF / (MTBF + MTTR)`

Answer

`MTTR / (MTBF + MTTR)`

Answer

`MTBF × MTTR`

Answer

`MTBF - MTTR`

Question 47

Why use percentiles like p99 instead of the average to measure latency?

Accepted Answer

Because the average hides the slowness many users experience (tail latency)

Answer

Because computing p99 is simpler and cheaper than the average

Answer

Because p99 is always lower than the average and looks better

Answer

Because the average only makes sense for failed requests

Question 48

In distributed systems, what is a 'heartbeat' used for?

Accepted Answer

Periodically sending a message to show a node is alive

Answer

Synchronizing the system clocks of all servers

Answer

Periodically rotating encryption keys between nodes

Answer

Counting the number of active users per second

Question 49

Why does making application servers stateless simplify failover?

Accepted Answer

Because any healthy instance can serve any request without local state

Answer

Because stateless servers never fail at all

Answer

Because statelessness removes the need for a load balancer

Answer

Because data is copied into every server's memory

Question 50

How is the difference between 'fault tolerance' and 'high availability' usually described?

Accepted Answer

Fault tolerance means no interruption; HA means minimal downtime with fast recovery

Answer

Fault tolerance is only for databases; HA is only for web servers

Answer

HA means zero interruption; fault tolerance means hours of downtime

Answer

Both mean exactly the same and are interchangeable