Middle Replication & Sharding Interview Questions (50 MCQs)

Question 1

Which approach implements read-your-writes in a system with read replicas?

Accepted Answer

Reading a user's data from the leader for a while after their last write

Answer

Adding more read replicas so lag goes down

Answer

Enabling CDN caching for all responses

Answer

Sending every read to a randomly chosen replica

Question 2

What is the simplest way to guarantee monotonic reads across several replicas?

Accepted Answer

Each user always reads from a fixed replica, e.g., by hashing their ID

Answer

Each request goes round-robin to the next replica

Answer

Reads always go to the replica with the lowest CPU

Answer

Switch replication from async to LWW

Question 3

In a partitioned system, a user sees an answer before the question itself. Which guarantee is violated?

Accepted Answer

Consistent prefix reads

Answer

Read-your-writes

Answer

Durability

Answer

Linearizable writes on a single key

Question 4

The data team wants to run heavy reports on the production database. What is usually the best option?

Accepted Answer

A dedicated reporting replica isolated from user traffic

Answer

Running reports on the leader during peak hours

Answer

Running reports on the same replicas that serve users

Answer

Sharding the database just to run the reports

Question 5

In a multi-tenant SaaS, what is the main risk of sharding by `tenant_id`?

Accepted Answer

One very large tenant can fill or overload a single shard

Answer

Queries within a tenant must hit every shard

Answer

Each tenant's data is scattered across all shards

Answer

Isolating tenants from each other becomes impossible

Question 6

Which technique is common for spreading writes to a very hot key?

Accepted Answer

Appending a random suffix (salt) so the key spreads over several partitions

Answer

Increasing the replication factor of that partition

Answer

Switching the hash algorithm from MD5 to SHA-256

Answer

Moving that key to a shard with a bigger disk

Question 7

What is the main cost of salting a hot key?

Accepted Answer

Reads must fetch all sub-keys and merge the results

Answer

Writes must be repeated on every shard in the cluster

Answer

That key can no longer be replicated

Answer

The stored data volume grows tenfold

Question 8

With a local (document-partitioned) secondary index, how is a query on the indexed field (not the shard key) executed?

Accepted Answer

It is sent to all shards and results are merged (scatter-gather)

Answer

It goes only to the shard owning that index value

Answer

It is read directly from a central global index

Answer

It cannot run until the next rebalancing

Question 9

What is the main drawback of a global (term-partitioned) secondary index?

Accepted Answer

One write may touch several partitions, and the index is often updated async

Answer

Reads on the indexed field must always hit every shard

Answer

It only works with numeric shard keys

Answer

It cannot be replicated

Question 10

With the 'fixed number of partitions' strategy (e.g., 1,000 partitions on 10 nodes), how is a new node added?

Accepted Answer

The new node takes whole partitions from existing nodes

Answer

All keys are rehashed and redistributed among nodes

Answer

The partition count is increased to 1,100

Answer

The new node only receives newly written data

Question 11

What does dynamic partitioning (as in HBase or MongoDB) do as data grows?

Accepted Answer

Splits a partition in two once it exceeds a size threshold

Answer

Rebuilds every partition from scratch each night

Answer

Always keeps partition count equal to node count

Answer

Moves excess data to cold storage

Question 12

For sensor data mostly queried as 'one device over a time range', which key fits best?

Accepted Answer

Partition by `device_id` and sort within the partition by `timestamp`

Answer

Range partition by `timestamp` only

Answer

Hash partition by `timestamp`

Answer

Partition by sensor type (`sensor_type`)

Question 13

In a cluster with heterogeneous hardware, how do vnodes help?

Accepted Answer

Stronger nodes get more vnodes so they take a bigger share

Answer

Vnodes route all requests to the strongest node

Answer

Weaker nodes are automatically removed from the cluster

Answer

Vnodes compress data on the weaker nodes

Question 14

Why can fully automatic rebalancing combined with automatic failure detection be dangerous?

Accepted Answer

A slow node is wrongly declared dead, and data movement adds load and cascades

Answer

Automatic rebalancing always deletes data

Answer

Automatic rebalancing only works during business hours

Answer

Automatic rebalancing permanently disables replication

Question 15

Which is a common way to route a request to the correct shard?

Accepted Answer

A routing tier that reads mapping metadata from a service like ZooKeeper

Answer

Sending the request to all shards and using the first response

Answer

Picking a random shard and retrying on failure

Answer

Storing the shard map in the user's browser cookie

Question 16

After an async failover, the old leader returns with unreplicated writes. What is the common behavior and its risk?

Accepted Answer

Those writes are discarded, and external systems relying on them become inconsistent

Answer

Those writes merge automatically with the new leader without conflict

Answer

The old leader becomes leader again and the new one loses its data

Answer

The cluster stays fully read-only until a manual review

Question 17

During failover with async replication, which follower is the best candidate for promotion?

Accepted Answer

The follower with the most advanced replication log position

Answer

The follower with the most free disk space

Answer

The follower that was started up earliest

Answer

The follower with the fewest client connections

Question 18

Which mechanism is used to prevent split-brain damage after failover?

Accepted Answer

Fencing (e.g., STONITH) that stops the old leader from writing

Answer

Dropping replication factor to 1 during failover

Answer

Raising the failure-detection timeout to several hours

Answer

Enabling LWW on all tables

Question 19

Which scenario is a good fit for multi-leader replication?

Accepted Answer

A multi-datacenter app or an offline-first client like a phone calendar

Answer

A small single-server database with low traffic

Answer

A banking system where every transaction must be serializable immediately

Answer

A purely read-only database with no writes

Question 20

What is the simplest strategy for reducing conflicts in multi-leader replication?

Accepted Answer

Always route all writes for a record or user to one designated leader

Answer

Increase the number of leaders as much as possible

Answer

Make inter-leader replication sync without any timeout

Answer

Set every server's clock one minute ahead

Question 21

What can a vector clock (version vector) detect that a plain timestamp cannot?

Accepted Answer

Whether two writes were concurrent or one happened after the other

Answer

The exact write time in milliseconds

Answer

Which user performed the write

Answer

How much data each write changed

Question 22

What is the key property of CRDTs (Conflict-free Replicated Data Types)?

Accepted Answer

Their merge is commutative, associative, and idempotent

Answer

They can only be written on a single leader

Answer

Every write requires consensus among all replicas

Answer

They resolve conflicts by highest timestamp

Question 23

What does a sloppy quorum with hinted handoff do?

Accepted Answer

When home nodes are down, writes are stored temporarily on other nodes and handed back later

Answer

It reduces the replicas required for reads to zero

Answer

It rejects all writes until home nodes return

Answer

It forces all replicas into sync replication

Question 24

In leaderless replication, how do read repair and anti-entropy differ?

Accepted Answer

Read repair fixes stale replicas during reads; anti-entropy compares all data in the background

Answer

Read repair runs only on leaders, anti-entropy only on followers

Answer

Both do exactly the same thing under different names

Answer

Read repair deletes data, anti-entropy compresses it

Question 25

In a leaderless system with N=3, what is the trade-off of W=3, R=1?

Accepted Answer

Reads are fast, but even one node failure blocks writes

Answer

Writes are fast, but reads may return stale data

Answer

Both reads and writes tolerate two node failures

Answer

The quorum condition fails and reads are always stale

Question 26

Which approach is common for avoiding cross-shard joins?

Accepted Answer

Denormalizing data or co-locating related data under the same shard key

Answer

Running joins by locking all shards simultaneously

Answer

Converting every join into a SQL subquery

Answer

Raising the replication factor so the join becomes local

Question 27

In a sharded database, what is usually done with small, rarely changing tables like `countries` or `currencies`?

Accepted Answer

Replicate them to every shard as reference tables

Answer

Split them across shards by `id`

Answer

Keep them only on the first shard

Answer

Delete them and hardcode them in the code

Question 28

Why is two-phase commit (2PC) expensive for cross-shard transactions?

Accepted Answer

Extra round-trips and holding locks until the coordinator decides, blocking if it fails

Answer

Because 2PC always rolls back half the transactions

Answer

Because 2PC only works on NoSQL databases

Answer

Because 2PC disables replication

Question 29

Which sequence is common for resharding with minimal downtime?

Accepted Answer

Dual-write to old and new layouts, backfill old data, verify, then switch reads

Answer

Stop the service fully, copy the data, and restart

Answer

Switch reads to the new shards before copying any data

Answer

Drop the old shards and wait for new ones to fill

Question 30

In directory-based sharding, how is the directory kept from becoming a bottleneck?

Accepted Answer

Replicating the directory and caching the mapping in the app tier

Answer

Querying the directory for every row of every query

Answer

Hosting the directory on the busiest shard

Answer

Deleting the directory after the first rebalancing

Question 31

For a 'top 10 best-selling products' query over hash-sharded data, what is the correct approach?

Accepted Answer

Fetch the top 10 from each shard and merge to pick the final 10

Answer

Take 10 items from the first shard that responds

Answer

Take the single top item from each shard

Answer

Take 10 divided by the shard count from each shard

Question 32

What is a reliable way to measure replication lag?

Accepted Answer

Periodically writing a timestamp to a heartbeat table on the leader and reading it on replicas

Answer

Comparing open connection counts on leader and replica

Answer

Measuring CPU usage on the replica

Answer

Counting tables on each replica

Question 33

What is the problem with statement-based replication compared to row-based?

Accepted Answer

Nondeterministic functions like `NOW()` or `RAND()` yield different results on replicas

Answer

Its log is always much larger than row-based

Answer

It cannot replicate `INSERT` statements

Answer

It does not work on replicas of the same version

Question 34

What advantage does a logical (row-based) replication log have over physical WAL shipping?

Accepted Answer

It is decoupled from the storage engine and usable for CDC or different DB versions

Answer

It always applies faster than any other replication

Answer

It eliminates replication lag entirely

Answer

It removes the need for a leader

Question 35

In Cassandra with primary key `((user_id), created_at)`, what does each part do?

Accepted Answer

`user_id` picks the partition and `created_at` orders rows within it

Answer

`created_at` picks the partition and `user_id` is the inner order

Answer

Both are hashed together and there is no ordering

Answer

`user_id` is just a local secondary index

Question 36

The `users` table is sharded by `user_id`. How do we enforce `email` uniqueness?

Accepted Answer

A separate table sharded by `email` itself that acts as a unique lookup

Answer

Declaring `UNIQUE` on `email` in each shard is enough

Answer

Checking for the email only in the app cache

Answer

Picking a random shard and checking the email there

Question 37

Why are Snowflake-style IDs (timestamp + machine id + sequence) popular in sharded systems?

Accepted Answer

They are unique without central coordination and roughly time-sortable

Answer

They are fully random and carry no time information

Answer

They guarantee all IDs land on a single shard

Answer

They are always shorter and more random than UUIDv4

Question 38

Why do many teams define many logical shards (e.g., 4,096) on a few physical servers from day one?

Accepted Answer

So later scaling just moves logical shards without changing key mapping

Answer

Because each physical server needs at least 1,000 shards

Answer

So that replication is no longer needed

Answer

Because queries over more shards are always faster

Question 39

In geo-partitioning by a user's home region, which operation stays slow?

Accepted Answer

Interactions between users in different regions or global queries

Answer

Reading the user's own profile from their region

Answer

Writing the user's settings in their own region

Answer

Logging the user in within their local region

Question 40

Which approach allows reading from replicas while still preserving read-your-writes?

Accepted Answer

The client sends its last write's log position (LSN/GTID) and the replica answers only once caught up

Answer

The replica always answers immediately and the client refreshes if stale

Answer

Restart all replicas every second

Answer

Use the client's wall clock as the data version

Question 41

Why do systems like etcd or CockroachDB use a consensus algorithm like Raft for leader election?

Accepted Answer

So majority voting ensures at most one valid leader per term

Answer

So they can have several leaders in the same term

Answer

To convert replication from sync to async

Answer

To remove the need for heartbeats

Question 42

What is the trade-off of cascading (chained) replication, where some replicas feed from other replicas?

Accepted Answer

It reduces leader network/CPU load but downstream replicas lag more

Answer

It brings lag to zero for every replica

Answer

It allows writes on every replica

Answer

It removes the need for failover entirely

Question 43

What is a delayed replica (e.g., intentionally 1 hour behind) used for?

Accepted Answer

Fast recovery from human errors like `DROP TABLE` before they reach it

Answer

Lowering read latency for distant users

Answer

Increasing the cluster's write capacity

Answer

Replacing the leader during automatic failover

Question 44

In an LWW system, node A's clock is two seconds ahead. What happens to a subsequent write on node B?

Accepted Answer

It may lose despite being newer and be silently discarded

Answer

It always wins because it truly happened later

Answer

The system detects the conflict and notifies the user

Answer

Both writes are kept and returned as siblings

Question 45

Why are deletes recorded as tombstones in leaderless replication instead of removing data immediately?

Accepted Answer

So a replica that missed the delete does not resurrect the old data

Answer

Because physically deleting data from disk is impossible

Answer

So disk usage always stays constant

Answer

Because tombstones replicate faster than any other write

Question 46

A key is extremely hot for reads only (like a celebrity's profile). What is the most effective fix?

Accepted Answer

Cache it in a front tier or spread reads across that shard's replicas

Answer

Salt the key and spread writes over several partitions

Answer

Change the shard key of the entire table

Answer

Lower the replication factor of that shard

Question 47

How is a schema migration usually performed across 200 shards?

Accepted Answer

Backward-compatible changes are rolled out gradually, shard by shard

Answer

All shards are locked at once so the migration is atomic

Answer

New code is deployed everywhere before any migration

Answer

Only the first shard is migrated and others sync themselves

Question 48

With N=5, W=3, R=3 in a leaderless system, how many node failures can be tolerated while both quorum reads and writes succeed?

Accepted Answer

2 nodes

Answer

1 node

Answer

3 nodes

Answer

0 nodes

Question 49

One read replica has fallen 5 minutes behind under heavy load. What should the load balancer do?

Accepted Answer

Remove that replica from the read pool until lag drops below a threshold

Answer

Send it more traffic so it catches up faster

Answer

Promote it to the new leader immediately

Answer

Hold the other replicas back so they all match

Question 50

In multi-leader setups, what advantage does an all-to-all topology have over a circular one?

Accepted Answer

One node failure does not break replication among the others

Answer

It guarantees messages always arrive in correct causal order

Answer

It reduces the number of replication messages

Answer

It eliminates conflicts entirely