Cloud Storage - Mabel Flashcards

Question

What is storage replication like within/inbetween stamps?

Answer 1

Inter-stamp replication (asynchronous) of the partition layer Intra-stamp replication (synchronous) of the stream layer

Answer 2

1. Optimise latency 2. Resillience to natural catastrophes

Answer 3

Latency - S3 can take around 100-200ms to query as opposed to 1-9ms in a typical database

Answer 4

Similar model to object storage Objects are now **400KB** (Dynamo) Streaming systems use systems like S3 Blob storage and download the smaller objects in a "stream"

Answer 5

get(key, context) -> value, context put(key,value, context) | delete also is in API

Answer 6

* Its API is considerably simpler than that of a relational database (which comes with query languages) * It does not ensure atomic consistency; instead, it guarantees eventual consistency * A key-value store scales out well, in that it is very fast also at large scales. | AP

Answer 7

Amazon Dynamo is a specific key-value based on the Chord protocol, which is a **Distributed Hash Table.** Distributed Hash Tables are generally highly scalable, robust against failure and self organizing. However, they do not provide any support for range queries, guarantees on data integrity (which is pushed to the user), and do not deal with any security issues. On the physical level, a distributed hash table is made of nodes (the machines we have in a data center, piled up in racks) that work following a few design principles.

Answer 8

1. **Incremental stability.** Nodes can enter and leave in a way that does not afect the stability of the system 2. **Symmetry.** All machines run the same software and have exactly the same behaviour 3.** Decentralization**. There is no “central node” that orchestrates the others 4. **Heterogeneity**. The nodes may have different CPU power, amounts of memory, etc. | Only peer-to-peer network shown in lectures

Answer 9

Dynamo generates **128** bit hashes (16 bytes) IDs are organised in a ring and machines get a random positition on ring. When objects are hashed, they are stored on the machine that is reached first when travelling clockwise around the ring ## Footnote pages 67 onwards in textbook

Answer 10

The new machine gets assigned a position. This position is in the domain of responsibility of some existing node m. This ring interval is now going to be split: everything between m and n remains within the responsibility of m, but the other half of the domain of responsibility is newly the responsibility of n. **The corresponding data needs to be transferred from m to n.**

Answer 11

**Graceful Exit** The node (n) informs the rest of the cluster before leaving. Then, it will transfer all its interval of responsibility to the next node clockwise (m) on the ring. **Abrupt Exit** Data will be gone if no data replication is in place.

Answer 12

Data within an interval will not be stored only the node that follow the interval clockwise, but on the next N nodes that follow the interval clockwise. Equivalently explained, a node is responsible not only for the interval that follows it counterclockwise, but for the next N intervals that follow it counterclockwise.

Answer 13

Preference Lists Each node knows, for every key (or key range), which node(s) are responsible (and hold a copy) of it. This is done by associating every key (key range) with a list of nodes, by decreasing priority (going down the ring clockwise) (the highest priority node is the first in the ring and the first assigned machine)

Answer 14

Upon reading a value, R is the minimum number of nodes (among the N nodes responsible for the key) from which a copy of the value must be obtained and compared. Upon writing a new value, W is the minimum number of nodes (among the N nodes responsible for the key) that must have acknowledged receipt and confirmed storage of the value before the put or delete transaction succeeds. At least N nodes in each preference list

Answer 15

1. Connect to load balancer 2. Directed to a random node on the ring 3. Random node connects you to "coordinator node" - node first om preference list 4. Connects to N-1 nodes hosting replicas 5. Stops when R of them have responded | R is the min number of nodes read from

Answer 16

Pro * Highly scalable * Robust against failure * Self organising Con * Lookup, no search * Data Integrity * Security Issues

Answer 17

1. Poor distribution - All machines are in one section of ring 2. Heterogeneous Performance Tokens!! :) We split up the ring into tokens (some mulptiple of the number of machines) and assign tokens randomly to nodes.

Answer 18

An added node gets assigned tokens to take over from other nodes (not all tokens have to come from same node) A deleted node gets its tokens redistributed to other nodes in the network (not all tokens have to go to the same node)

Answer 19

In AP systems, where partitions may exist, vector clocks are used to reconcile versions following a DAG structure. A vector clock can logically be seen as a map from nodes (machines) to integers, i.e., the version number is incremented per machine rather than globally. A version number increases for a machine when this machine processes the corresponding entry (key-value) and updates it | pages 74-76 of the textbook illustrate this + tutorial 2

Answer 20

* If no partial ordering exists between nodes, then those nodes should be on the same level, i.e., they are both valid versions. * If there is an edge between two nodes, then the parent node should be smaller than the child. * You cannot have skip connections, i.e., there cannot be an edge from an ancestor node (excluding the parent node) directly to a child node. * Transitive edges shouldn't be present in the version DAG. * Each edge represents the update of exactly one entry of the vector clock.

Answer 21

Vector clocks can be compared to each other with a partial order relation ≤ A vector clock is <= another vector clock if for each machine, the associated integer in the first vector clock is also smaller than or equal to the associated integer in the second vector clock. e.g {"A":2,"C":1} ≤ {"A":3,"B":1,"C":1}

Answer 22

* From cheapset to most expensive: HDD SSD SDRAM * From slowest to fastest (in terms of read/write speed): HDD SSD SDRAM * By their capacity in increasing order: SDRAM SSD HDD

Cloud Storage - Mabel Flashcards

Mabel to Fill (48 cards)