Middle Consistency, CAP & Consensus Interview Questions (50 MCQs)

Question 1

What does PACELC add on top of CAP?

Accepted Answer

That even without a partition you trade off latency against consistency

Answer

That during a partition you can have all three CAP properties

Answer

That partition tolerance is only needed for multi-datacenter systems

Answer

That durability must be added to CAP as a fourth letter

Question 2

With default settings, where do systems like Cassandra and DynamoDB fall in PACELC?

Accepted Answer

PA/EL

Answer

PC/EC

Answer

PA/EC

Answer

PC/EL

Question 3

How are Google Spanner or etcd described in PACELC terms?

Accepted Answer

PC/EC; they always pick consistency, even at the cost of latency or availability

Answer

PA/EL; they always prefer speed and availability over all else

Answer

PA/EC; available under partition but consistent normally

Answer

PC/EL; consistent under partition but eventual normally

Question 4

How does the "C" in CAP differ from the "C" in ACID?

Accepted Answer

CAP's C means linearizability; ACID's C means preserving application invariants

Answer

Both mean exactly the same thing, just used in two contexts

Answer

CAP's C means durability, while ACID's C means isolation

Answer

CAP's C means eventual consistency; ACID's C means strong

Question 5

What is the precise definition of Availability in the formal CAP proof?

Accepted Answer

Every request to a non-failing node must eventually get a non-error response

Answer

The system must have at least 99.99% yearly uptime

Answer

At least one node in the cluster must answer requests

Answer

Each request must be answered within 100 milliseconds

Question 6

What is the difference between linearizability and serializability?

Accepted Answer

Linearizability is about recency on single objects; serializability is about isolating multi-object transactions

Answer

They are identical; linearizability is just another name for it

Answer

Serializability guarantees real-time order but linearizability does not

Answer

Linearizability is for relational and serializability is for NoSQL

Question 7

How does sequential consistency differ from linearizability?

Accepted Answer

One shared global order respecting each process's order, but not necessarily real time

Answer

Each process sees a different write order with no shared order

Answer

All operations must run strictly in physical timestamp order

Answer

It orders one client's writes only and ignores reads

Question 8

What does causal consistency guarantee?

Accepted Answer

Causally related operations are seen in order; concurrent ones may appear in any order

Answer

All operations appear in one identical global order to every node

Answer

Every read always returns the globally latest write

Answer

Only writes made to the same key are kept in order

Question 9

In an async leader-follower setup, which technique provides read-your-writes for a user?

Accepted Answer

Track the user's last write position and read only from replicas caught up to it

Answer

Always read from a random follower to spread the load

Answer

Raising the user's browser cache TTL to several minutes

Answer

Reducing the number of followers down to one

Question 10

What is the simplest way to provide monotonic reads across multiple replicas?

Accepted Answer

Each user always reads from the same replica, e.g. by hashing their ID

Answer

Send every read to a randomly chosen replica

Answer

Route reads to whichever replica has the lowest latency

Answer

Read all replicas and pick the older value

Question 11

Which anomaly do consistent prefix reads prevent?

Accepted Answer

Seeing a reply before its original message because writes appear out of order

Answer

Seeing the user's own write after a delay of seconds

Answer

Losing a write because the leader crashed

Answer

Two values being written to the same key concurrently

Question 12

With N=3, what is the trade-off of choosing `W=3` and `R=1`?

Accepted Answer

Reads are fast, but writes fail if even one node is down

Answer

Writes are fast, but reads may return stale values

Answer

Both reads and writes continue when one node fails

Answer

No overlap between reads and writes is guaranteed

Question 13

With N=3, `R=1` and `W=1`, what consistency do you get?

Accepted Answer

Eventual consistency; a read may hit a replica that missed the write

Answer

Strong consistency, because each operation touches one node

Answer

Linearizability, because the latency is very low

Answer

Causal consistency, because only one node is involved

Question 14

What is a sloppy quorum in Dynamo-style systems?

Accepted Answer

When home replicas are unreachable, writes are accepted by other non-home nodes

Answer

The quorum is computed by plurality instead of an absolute majority

Answer

Reads are returned without waiting for any replies

Answer

Nodes vote at random on each incoming request

Question 15

What does hinted handoff do?

Accepted Answer

A stand-in node stores the write with a hint and delivers it when the owner returns

Answer

The leader hands its role to another node before shutting down

Answer

The client redirects the request to another node to cut load

Answer

Old data is moved to cold storage and removed from the node

Question 16

How does read repair work?

Accepted Answer

During a read, if replicas disagree, the newer value is written back to stale ones

Answer

A broken disk is automatically replaced with a healthy one

Answer

Failed reads are retried automatically until they succeed

Answer

The database index is rebuilt after every read

Question 17

Why are Merkle trees used in anti-entropy between replicas?

Accepted Answer

To find differing ranges by comparing hashes without transferring all data

Answer

To encrypt data before it is sent over the network

Answer

To sort writes in order of their timestamps

Answer

To elect a new leader based on tree depth

Question 18

In an LWW store, node A's clock is 10 seconds ahead. What is the risk?

Accepted Answer

Truly newer writes from other nodes are silently dropped for up to 10 seconds

Answer

All writes from node A are automatically rejected

Answer

The cluster goes fully offline until the clock is fixed

Answer

Reads from node A always return errors

Question 19

When vector clocks show two versions of a key are concurrent, what do systems like Riak do?

Accepted Answer

Keep both as siblings so the app or a CRDT can merge them

Answer

Keep the version with the larger vector without checking

Answer

Delete both versions and return a default value

Answer

Reject the second write to the client with a conflict error

Question 20

If `L(A) < L(B)` under Lamport clocks, what can you conclude?

Accepted Answer

Only that B did not happen before A; A may have caused B or be concurrent

Answer

A definitely caused B, so happened-before holds

Answer

A and B definitely happened concurrently

Answer

A occurred strictly earlier than B in physical time

Question 21

What does a Hybrid Logical Clock (HLC) combine?

Accepted Answer

Closeness to physical time with Lamport-style causal ordering

Answer

Atomic-clock accuracy with the low cost of NTP

Answer

A full vector clock in a fixed one-byte size

Answer

A GPS hardware clock with a database transaction counter

Question 22

Which clock should you use to measure a timeout or lease duration on one machine?

Accepted Answer

A monotonic clock, since NTP adjustments can't make it jump back

Answer

The wall clock, since it matches real-world time

Answer

The client's timestamp sent in the request header

Answer

The database server's clock read through a query

Question 23

Why does Raft use randomized election timeouts (e.g. 150–300 ms)?

Accepted Answer

To reduce split votes so one candidate usually wins first

Answer

To balance the nodes' CPU usage over time

Answer

To harden the cluster against timing attacks

Answer

So the leader can estimate the number of followers

Question 24

In Raft, a node only grants its vote to a candidate if which condition holds?

Accepted Answer

The candidate's log is at least as up-to-date as its own

Answer

The candidate has the lowest numeric ID in the cluster

Answer

The candidate has the most free RAM available

Answer

The candidate was already leader in the previous term

Question 25

In Raft, when is a log entry considered committed?

Accepted Answer

When the leader has replicated it to a majority of nodes

Answer

As soon as the leader writes it to its local log

Answer

When every node in the cluster without exception stores it

Answer

When the client confirms it received the response

Question 26

In a five-node Raft cluster, the leader lands on the minority (2-node) side of a partition. What happens?

Accepted Answer

The old leader can't commit and the majority elects a new leader with a higher term

Answer

Both sides accept writes independently and merge later

Answer

The whole cluster accepts no writes until the partition heals

Answer

The old leader keeps committing with its two nodes

Question 27

Why can reading directly from a Raft leader's local state return stale data?

Accepted Answer

It may be partitioned off while a new leader was elected without its knowledge

Answer

Because the leader only writes data to disk after restarting

Answer

Because followers apply writes before the leader does

Answer

Because the leader's state machine is never updated

Question 28

When a Raft follower's log conflicts with the leader's at some index, what happens?

Accepted Answer

The follower's conflicting entries onward are deleted and replaced with the leader's

Answer

The leader replaces its own log with the follower's entries

Answer

Both logs are merged and duplicates are removed

Answer

The follower is permanently removed from the cluster

Question 29

A team grows a three-node etcd cluster to four nodes to make it "more resilient". What is the result?

Accepted Answer

Fault tolerance stays at 1 node, while write latency and failure odds rise

Answer

Fault tolerance grows from 1 node to 2 nodes

Answer

Write throughput roughly doubles

Answer

The cluster can no longer elect any leader

Question 30

How does a fencing token prevent a stale lock holder from corrupting data?

Accepted Answer

Each lock grant returns an increasing number, and storage rejects older tokens

Answer

An encrypted token lets only trusted servers acquire the lock

Answer

The token makes the lock never expire at all

Answer

The client broadcasts the token to all nodes before each write

Question 31

A worker holding a TTL-based lock hits a long GC pause after acquiring it. What is the main danger?

Accepted Answer

The lock expires and another takes it, while the first still thinks it holds it

Answer

The lock server crashes because the client paused

Answer

The lock stays with the first worker forever

Answer

All workers lose the lock at once and deadlock

Question 32

When is a simple Redis lock with a TTL (no fencing) usually acceptable?

Accepted Answer

When the lock is only for efficiency and an occasional duplicate run is harmless

Answer

When the lock prevents a financial transaction being charged twice

Answer

When two concurrent writers would permanently corrupt a file

Answer

When the lock is responsible for username uniqueness

Question 33

Why is 2PC called a "blocking" protocol?

Accepted Answer

If the coordinator crashes after prepare, participants are stuck in doubt until it returns

Answer

Because all clients must wait during any transaction

Answer

Because it fully locks the network for other transactions

Answer

Because only one transaction can run at any moment

Question 34

What 2PC problem does Three-Phase Commit (3PC) target, and what is its main limitation?

Accepted Answer

It reduces blocking, but partitions and unbounded delays can make it inconsistent

Answer

It reduces latency, but only works with two participants

Answer

It removes the coordinator, but gives up durability

Answer

It fixes deadlocks, but needs atomic clocks

Question 35

What is the difference between orchestration and choreography sagas?

Accepted Answer

Orchestration has a central coordinator issuing steps; choreography has services reacting to each other's events

Answer

Orchestration uses 2PC while choreography uses Raft

Answer

Choreography is only usable for two-step sagas

Answer

Orchestration defines no compensating actions at all

Question 36

Which ACID property does a saga lack by default?

Accepted Answer

Isolation; other transactions can see the saga's intermediate state

Answer

Durability; saga steps are never saved to disk

Answer

Atomicity at the level of each local transaction

Answer

Consistency within each local database

Question 37

What problem does the Transactional Outbox pattern solve?

Accepted Answer

Dual writes; atomically updating the DB and publishing by storing the event in the same transaction

Answer

Slow queries, by saving their results in a cache table

Answer

Full message queues, by auto-deleting old messages

Answer

Missing backups, by copying tables into another database

Question 38

To implement idempotency keys correctly on a payment endpoint, what should the server store?

Accepted Answer

The key, a hash of the request body, and the final response for replays

Answer

Only the key, so duplicate requests are rejected with `500`

Answer

Only the client IP, so a second request from it is refused

Answer

Only the request timestamp, so older requests are refused

Question 39

What is so-called "exactly-once processing" usually built from in practice?

Accepted Answer

At-least-once delivery plus idempotent processing or dedupe by message ID

Answer

At-most-once delivery plus a longer network timeout

Answer

Sending every message exactly once with no ack needed

Answer

Using UDP instead of TCP to avoid retransmission

Question 40

What scope does Kafka's exactly-once semantics (idempotent producer and transactions) cover?

Accepted Answer

Read from Kafka, process, write back to Kafka; not external side effects

Answer

Any external side effect like sending emails or calling APIs

Answer

Only messages that are delivered without any retry

Answer

Every database the consumer happens to connect to

Question 41

What property of ZooKeeper ephemeral znodes makes them useful for leader election?

Accepted Answer

It is deleted automatically when the client's session ends, e.g. on crash

Answer

Its contents can never be modified at all

Answer

It is stored only on the leader and never replicated

Answer

It gets a new version number on every read

Question 42

In the ZooKeeper lock recipe, why does each client watch only the sequential znode just before its own?

Accepted Answer

To avoid the herd effect; releasing the lock wakes only one client

Answer

To minimize ZooKeeper's disk consumption

Answer

Because ZooKeeper allows only one watch per session

Answer

So that client ordering gets randomly shuffled

Question 43

In etcd, how do you safely update a key only if it hasn't changed since you read it?

Accepted Answer

With a `Txn` that compares the key's `mod_revision` and then does a `Put`

Answer

By re-reading the key and writing right after it

Answer

By setting a short TTL on the key before writing

Answer

By temporarily disabling replication in the cluster

Question 44

How does optimistic concurrency control with a `version` column prevent lost updates?

Accepted Answer

`UPDATE ... WHERE id=? AND version=?`; zero rows changed means someone wrote first

Answer

It exclusively locks the whole table before every read

Answer

It copies every update into a separate timestamped table

Answer

It applies concurrent writes in alphabetical order

Question 45

What property does a CRDT (Conflict-free Replicated Data Type) have?

Accepted Answer

Replicas update independently and merging always converges to the same state

Answer

Every write must be approved by a majority before applying

Answer

Only one replica may write and the rest are read-only

Answer

Conflicts are resolved by asking the user to choose

Question 46

What is the biggest challenge of multi-leader replication between two datacenters?

Accepted Answer

Resolving conflicts when one record changes in both datacenters at once

Answer

Being unable to read from either datacenter

Answer

Each write's latency growing to several seconds

Answer

Having to shut one datacenter down every night

Question 47

In Basic Paxos, when is a value considered "chosen"?

Accepted Answer

When a majority of acceptors accept a proposal with the same number

Answer

When the proposer sends it to all of the learners

Answer

When every acceptor without exception accepts it

Answer

When the proposer with the highest node ID proposes it

Question 48

For a quorum write with N=5 and W=3, why is latency usually lower than writing to all 5?

Accepted Answer

It waits only for the 3 fastest replies, so the slowest nodes don't count

Answer

Because it is sent to only 3 nodes and the other 2 never get it

Answer

Because the data is compressed before it reaches 3 nodes

Answer

Because 3 nodes always live in the nearest datacenter

Question 49

Why must saga steps and compensating actions be idempotent?

Accepted Answer

Messages may be redelivered and the orchestrator retries steps after a crash

Answer

Because without idempotency a saga can't define compensations at all

Answer

Because idempotency gives a saga full isolation

Answer

Because each saga step must take exactly one millisecond

Question 50

Why don't teams usually store bulk application data directly in ZooKeeper or etcd?

Accepted Answer

Every write goes through consensus, and their capacity is designed for metadata

Answer

Because these systems provide no durability guarantee

Answer

Because they only support numeric data values

Answer

Because data is automatically purged after 24 hours