Aerospike: Key Value Data Access

Aerospike aer . o . spike [air-oh- spahyk] noun, 1. tip of a rocket that enhances speed and

stability

KVS Data Access

Topics

➤ Structured v. Unstructured Data➤ Database Hierarchy and Definitions➤ Data Access Patterns

Structured DatabasesFor performance, many early databases were structured. Every table has a defined schema. Changes to the schema required a DBA, possibly a Change Control Board (CCB).

id (10

bytes)

lname(40

bytes)

fname(40

bytes)

address(60

bytes)

city(20

bytes)

state(20

bytes)

Phone(20 bytes)

1 Able John 123 First New York NY 2128675309

2 Baker Kris 234 Second

UNKNOWN UNKNOWN

UNKNOWN

3 Charlie Larry 345 Third Seattle WA 4258675309

4 Delta Moe 456 Fourth Austin TX 7378675309

+ ACID

+ Familiarity

- Requires pre-defined schema

- Changes to schema can be traumatic, limiting dynamic application development.

- Poor durability on SSD

Structured Databases

Unstructured DatabasesUnstructured databases do not have a pre-defined schema and bins may exist in some records, but not in others. Different kinds of records may be mixed in sets.

Id lname fname

address city state

Phone Size

1 Able John 123 First New York

NY +81 2128 6753 909

45 bytes

2 Baker Kris 234 Second

20 bytes

3 Charlie

8 bytes

4 Delta Moe 456 Fourth

Austin TX 7378675309 47 bytes

+ No predefined schema

+ Addition of new bins can be done from client

+ Addition of new sets (like tables) can be done from client

+ Makes most of sequential write speed of disks

- Difficult to predict object size

- Updates to a record require an entire record re-write (AS solution is LDTs)

Aerospike

What Do You Want From A Distributed DB?• Hide the complexity of distribution.• Linear scalability.• Better service availability.

Smart Partition Architecture

Cluster creates a map of how data is distributed, called a partition map.

Combine features from other architectures to create a map.

Smart Partitioning

• Every key is hashed using the RIPEMD160 hash function

• The creates a fixed 160 bits (20 bytes) string.

• 12 bits of this hash are used to identify the partition id

• There are 4096 partitions

• Are distributed among the nodes

PaikPaik

182023kh15hh3kahdjsh182023kh15hh3kahdjsh

PartitionID

Master node

Replica node

… 1 4

1820 2 3

1821 3 2

4096 4 1

Aerospike uses a partition table

Smart Partitioning

For simplicity, let’s take a 3 node cluster with only 9 partitions and a replication factor of 2.

Database HierarchyTerm Definition NotesCluster An Aerospike cluster services a

single database service. While a company may deploy multiple clusters, applications will only connect to a single cluster.

Node A single instance of an Aerospike database.

For production deployments, a host should only have a single node. For development, you may place more than one node on a host.

Namespace An area of storage related to the media. Can be either RAM or SSD based.

Similar to a “database” or “tablespaces” in relational databases.

Set An unstructured grouping of data that have some commonality.

Similar to “tables” in a relational database, but do not require a schema.

Record A key and all data related to that key.

Similar to a “row” in a relational database.

Bin One part of data related to a key. Bins in Aerospike are typed, but the same bin in different records can have different types. Bins are not required. Single bin optimizations are allowed.

(Large Data Type) LDT LDTs provide functions for storing arbitrarily large amounts of data without requiring the database to read the entire record.

Most commonly the data stored in LDTs will be time series data, but this is not a requirement. This feature is still in development.

Data HierarchyCluster

Node 1 Node 2 Node 3

Namespace

Record

Record BinBin

Cluster

➤ Will be distributed on different nodes. ➤ Management of cluster is automated,

so no manual rebalancing or reconfiguration is necessary.

➤ Will contain one or more namespaces. Adding/removing namespaces requires a cluster-wide restart.

➤ Each node is assumed to be identical.

➤ Data (and their associated traffic) will be evenly balanced across the nodes.

➤ Big differences between nodes imply a problem.

➤ Node capacity should take into account node failure patterns.

Namespaces➤ Are associated with the storage media:

Hybrid (ram for index and SSD for data) RAM + disk for persistence only RAM only

➤ Each can be configured with their own: replication factor (change requires a cluster-wide

restart) RAM and disk configuration settings for high-watermark default TTL (if you have data that must never be

automatically deleted, you must set this to “0”)

➤ Similar to “tables” in relational databases.

➤ Sets are optional.➤ Schema does not have to be pre-

defined.➤ In order to request a record, you

must know its set.➤ Scans can be done across a set

Records

➤ Similar to a row in a relational database.

➤ All data for a record will be stored on the same node. This is true even for LDTs.

➤ Any change to a record will result in a complete write of the entire record, unless using LDTs.

Bins➤ Values Are typed. Current types are:

Simple (integer, string, blob [language specific]) Complex (list, map) Large Data Types (LDTs)

➤ A single bin may be updated by the client. Increment Replacement User Defined Function (UDF)

Data HierarchyCluster

Node 1 Node 2 Node 3

Namespace

Record

Record BinBin

Data Access Patterns Read Write Update

Accessing An Object In AerospikeReading A Standard Data Type With SSDs

128 KB Blocks

Master Node

SSD (DATA)

ClientRAM (Index)

1) Client finds Master Node from partition map.

2) Client makes read request to Master Node.

3) Master Node finds data location from index in RAM.

4) Master Node reads entire object from SSD. This is true even if only reading bin.

5) Master Node returns value.

Index reference

Accessing An Object In AerospikeWriting A New Standard Data Type Record With SSDs

128 KB Blocks

Master Node

SSD (DATA)

ClientRAM (Index)

2) Client makes write request to Master Node.

3) Master Node make an entry indo index (in RAM) and queues write in temporary write buffer.

4) Master Node coordinates write with replica nodes (not shown).

5) Master Node returns success to client.

6) Master Node asynchronously writes data in 128 KB blocks.

7) Index in RAM points to location on SSD.

Asynchronous write

Accessing An Object In AerospikeUpdating A Standard Data Type Record With SSDs

128 KB Blocks

Master Node

SSD (DATA)

ClientRAM (Index)

2) Client makes update request to Master Node.

3) Master Node reads the existing record (if using multiple bins)

4) Master Node queues write of updated record in a temporary write buffer

5) Master Node coordinates write with replica nodes (not shown).

6) Master Node returns success to client.

7) Master Node asynchronously writes data in 128 KB blocks.

8) Index in RAM points to new location on SSD.

Asynchronous write

Accessing An Object In AerospikeKeeping It Efficient

128 KB Blocks

Master Node

SSD (DATA)

ClientRAM (Index)

Index reference

Minimize the

number of

network round trips

Minimize the

number of

network round trips

Minimize the

network bandwidt

Minimize the

network bandwidt

hMinimize

SSD reads/writ

Minimize SSD

reads/writes

Issues With Standard Data Types➤ Record size is limited by block size

(128 KB by default).➤ Even a small update to a record

results in a complete record re-write.

Example Use Case

To compare different systems, let’s take a look at a standard task.

➤Find out if an object has some value➤If it does, update the record and return a value

Example: Simple KVS MethodValue is one large string JSON object.

Example record:➤Key=user_id➤Value={“name” : “john”,

“dob” : “08-20-1970” ,“gender” : “male” ,“likes” : “cars,computers,goats”}

Business logic is that if the person is older than 18 years old, put them into campaign “bluesky”.

1.Client will request entire value from the node 2.Node reads entire value from disk3.Node sends entire value to client4.Client parses data and check logic on age5.Client updates record with new value

Value={“name” : “john”,

“dob” : “08-20-1970” ,“gender” : “male” ,“likes” : “cars,computers,goats” ,“campaigns” : “bluesky”}

6.Node writes entire value to disk

Client

Node Storage

Read (all)Read (all)

Read (all)

Write (all)

Write (all)Return status

Example: KVS with BinsValues are stored in bins

Example record:➤Key=user_id➤Value= “name” = “john”

“dob” = “08-20-1970” “gender” = “male” “likes” = “cars,computers,goats”

1.Client will request dob and campaign bins from the node 2.Node reads entire value from storage3.Node sends only dob and campaigns to client4.Client checks logic on age5.Client updates record with new bin

1.Node writes entire value to disk. Node must read value first.

Client

Node Storage

Read (bin)Read (all)

Read (all)

Read (bin)

Write (bin)

Write (all)

Read (all)

Return status

Example: Using UDFsValues are stored in bins

Example record:➤Key=user_id➤Value= “name” = “john”

“dob” = “08-20-1970” “gender” = “male” “likes” = “cars,computers,goats”

1.Client makes UDF request 2.Node reads entire value from storage3.Node applies UDF on returned data4.Nodes writes data5.Node returns status

Client

Node Storage

UDF Read (all)

Read (all)

Return status

Write (all)

Example: Connecting to a cluster

Policy contains operational defaults like timeout

Seed hostSeed hostSeed portSeed port

Do some workDo some work

Disconnect from the clusterDisconnect from the cluster

List of hostsList of hosts

Example: Get/Put operations

Setup some preliminary valuesSetup some preliminary values

Write a record with two bin valuesWrite a record with two bin values

Read a record with all bin valuesRead a record with all bin values

Example: Increment/Decrement operation

Add operation – avoids the read-add-write cycleAdd operation – avoids the read-add-write cycle

Example: Touch operation

Write a record with a 2 second expiryWrite a record with a 2 second expiry

Change it to a 5 second expiryChange it to a 5 second expiry

Aerospike: Key Value Data Access

aerospike database

node cluster

aerospike cluster services

aerospike aer

data hierarchy cluster

data ram

single node

different records

Technology

Aerospike Hybrid Memory Architecture

Hybrid Memory Aerospike @ PayPal€¦ · AEROSPIKE USER...

plug truncation in aerospike nozzle

Scaling up with Aerospike!

Aerospike Multi-site Clustering: Globally Distributed...

Aerospike Rapid Rebalance

Aerospike on IDCF Cloud

GoogleBenchmark Aerospike

Aerospike Architecture

Design of axisymmetric aerospike nozzle based on...

Aerospike Technical Overview · CEP Engine with all...

Brian Bulkowski. Aerospike

aerospike ppt

Introduction to Aerospike

aerospike Documentation

Bcache and Aerospike