4A_Replication & Partitioning

Question 1

What are the reasons for replication?

Answer 1

Performance:
- Scale in size; more capacity
- Scale geographically; Closer to client = lower latency
Redundancy:
- Have other copies when a component fails

Question 2

Redundancy vs Availability?

Answer 2

Availability also means ability to progress, not only read data but also able to do writes

Question 3

Definition of CAP?

Answer 3

Consistency, Availability and Partition Resistance

Question 4

What are the different consistency models?

Answer 4

Linearisability
Sequential consistency
Causal Consistency
Eventual Consistency

Question 5

What is linearisability?

Answer 5

Clients see atomic writes in same order

Question 6

What is sequential consistency?

Answer 6

Clients see writes in same order

Question 7

What is causal consistency?

Answer 7

Clients see causally related writes in same order, unrelated writes possibly in different order.

Question 8

What is eventual consistency?

Answer 8

Eventually all replicas converge.

Clients may see differences meanwhile.

Question 9

Arguments linearisability

Answer 9

Global lock on replicas

Not scalable

Question 10

Arguments Sequential Consistency

Answer 10

Need total order on writes

Hard, but some capacity possible. Use MutEx solutions

Question 11

Arguments causal consistency

Answer 11

Causal relations via vector clocks.

Hard, need per-process info

Question 12

Arguments against consistency in general?

Answer 12

Does not scale

Slow and expensive

Question 13

Replication Protocols?

Answer 13

Primary / Backup
Active Replication
Quorum Based

Question 14

What is primary / backup replication protocol?

Answer 14

Primary is implicit sequencer.
Variations:
- Send write operations to backups
- Send result of write operation to backups
- Send invalidation to backups

Question 15

What is active replication protocol?

Answer 15

All replicas perform the write operation

Need explicit sequencer / total ordering

Question 16

What is caching?

Answer 16

Able to replicate temporarily

Question 17

What are scaling techniques?

Answer 17

Bigger Machines
Virtualisation
Asynchronous communication
Replication and caching
Partitioning

Question 18

What is partitioning?

Answer 18

Split work/responsibilities over a set of servers

Common method is hierarchical

Question 19

Second common method partitioning?

Answer 19

Formulaic. Data is divided into parts according to some formula. Parts are distributed to servers via another formula using the server ID and a meta server remembers which parts are where

Question 20

What falls under service partitioning?

Answer 20

Split different things:
Obvious: Web, DB, DNS
Others: signup, login, search, user accounts, inventory

Question 21

What falls under data partitioning?

Answer 21

Splitting similar things:

Split by customer ID, last name, location. This is often hierarchically and splits are often equal in size (hashing)

Question 22

What is Hadoop?

Answer 22

Partition Data over many disks

Process data in parallel

Question 23

What is MapReduce data flow?

Answer 23

First you split your data. The map is some sort of formula function which is ran in parallel. Then the results are merged and a reduce function is applied after which the results are returned

Question 24

Why is consistent hashing used?

Answer 24

A method to divide data over servers

Question 25

What are the problems of dividing data over servers based on hashing?

Answer 25

If one server goes down, all data has to move