distributed caches

7/24/2019 distributed caches

1/34

From distributed cachesto in-memory data grids

TechTalk by Max A. Alexejev

[email protected]


2/34

Memory Hierarchy

R

!ns

"!

#$ cycles%#!ns

"&

#!' cycles% #(ns

"(

#$& cycles% #!)ns*RAM

+,)ns

Flash * /0

H**

Ta1es% Remote systems% etc

Max A. A2

2ost

torae ter


3/34

o3t4are caches

5m1rove res1onse times by reducing data accesslatency

67oad 1ersistent storages

6nly 4ork 3or 56-bound a11lications8

Max A. A3


4/34

2aches and data location

Max A. A4

"ocal Remote

Hierarchical

*istributed

hared

2onsistency

1rotocol

*istribution

algorithm


5/34

6k% so ho4 do 4e gro4 beyond onenode9

*ata re1lication

Max A. A5


6/34

:ro;s and 2on;s o3 re1lication

0est read 1er3ormance limited by single machine si?e 5n case o3 master-master re1lication% reuires com1lex

consistency 1rotocol

2on

Max A. A6


7/34

6k% so ho4 do 4e gro4 beyond onenode9

*ata distribution

Max A. A7


8/34

:ro;s and 2on;s o3 datadistribution

2an scale hori?ontally beyond single machinesi?e

Reads and 4rites 1er3ormance scaleshori?ontally

:ro

o 3ault tolerance 3or cached data 5ncreased latency o3 reads


9/34

Bhat do high-load a11lications need3rom cache9

"o4latency

"inearhori?ont

alscalabili

ty

Max A. A9

*istributed cache


10/34

2ache access 1atterns>

2ache AsideFor reading data:1. Application ass !or so"e data !or a given e#

2. $%ec t%e cac%e

3. &! data is in t%e cac%e ret'rn it to t%e 'ser

4. &! data is not in t%e cac%e !etc% it !ro" t%e ()*p't it in t%e cac%e* ret'rn it to t%e 'ser.

For +riting data

5. Application +rites so"e ne+ data or 'pdatesexisting.

6. ,rite it to t%e cac%e

7. ,rite it to t%e ().

-verall:

&ncreases reads per!or"ance

-oads () reads

&ntrod'ces race conditions !or +rites

Max A. Alexejev1/

2lient

*0

2ache


11/34


Read ThroughFor reading data:

1. Application ass !or so"e data !or a givene#

2. $%ec t%e cac%e

3. &! data is in t%e cac%e ret'rn it to t%e 'ser

4. &! data is not in t%e cac%e 0 cac%e +illinvoe !etc%ing it !ro" t%e () # %i"sel!*

saving retrieved val'e and ret'rning it tot%e 'ser.

-verall:

ed'ces reads latenc#

-oads read load !ro" 'nderl#ing storage

Ma# %ave locing e%avior* t%'s %elping+it% dogpile eect

e'ires s"arter cac%e nodes

Max A. Alexejev11

2lient

*0

2ache


12/34


Brite ThroughFor +riting data

1. Application +rites so"e ne+data or 'pdates existing.


3. $ac%e +ill t%en s#nc%rono'sl#

+rite it to t%e ().

-verall:

lig%tl# increases +rites latenc#

rovides nat'ral invalidation

e"oves race conditions on

+rites

Max A. Alexejev12

2lient

*0

2ache


13/34


Brite 0ehindFor +riting data1. Application +rites so"e ne+ data or

'pdates existing.


3. $ac%e adds +rites re'est to itsinternal 'e'e.

4. ater* cac%e as#nc%rono'sl# ;'s%es

'e'e to () on a periodic asis and


14/34

A variety o3 1roducts on the marketC

Max A. A14

Memcached Ha?elcast

Terracotta

Dh2ache

6racle2oherence

Riak

Redis

Mongo*0

2assandra

Eiga1aces

5nnis1an

C


15/34

"ets sort em out8

Gcaches

Memcached

Dhcache

C

oI"

Redis

2assandra

Mongo*0

C

*aEri

6r2ohe

Eem

Eiga

Erid

Ha?

5nn

o"e prod'cts are reall#

%ard to sort 0 lie

>erracotta in ot% (-

and ?xpress "odes.

Max A. Alexejev15


16/34

Bhy don;t 4e have any distributedin-memory R*0M9

5s% i3 3act% an exam1le o3 re1lication Hel1s 4ith reads distribution% but does not hel1 4ith

4rites

*oes not scale beyond single master

Master J Multilaves conguration

Hel1s 4ith reads and 4rites 3or datasets 4ith good dataaKnity

*oes not 4ork nicely 4ith joins semantics


17/34

Gey-alue caches

Max A. A17

Memcached and DH2ache are good exam1les tolook at

Geys and values are arbitrary binary


18/34

Memcached (eveloped !or

ive@o'rnal in 2//3

as client liraries in* @ava* '#* #t%onand "an# ot%ers

Bodes are independentand donCt co""'nicate

+it% eac% ot%er

Max A. Alexejev18


19/34

DH2ache &nitiall# na"ed ?as# iernate

$ac%e

@avacentric* "at're prod'ct +it%openso'rce and co""ercialeditions

-penso'rce version provides onl#replication capailities* distri'tedcac%ing re'ires co""erciallicense !or ot% ?$ac%e and

>erracotta >A

Max A. Alexejev19


20/34

oI" ystems

A 4hole bunch o3 diLerent 1roducts 4ith both1ersistent and non-1ersistent storage o1tions."ets call them caches and storages% accordingly.

0uilt to 1rovide good hori?ontal scalability

Try to ll the 3eature ga1 bet4een 1ure G and3ull-blo4n R*0M

Max A. A2/


21/34

2ase study> Redis

Britten in 2% su11orted byMBare

2lient libraries 3or 2% 2% Nav

cala% :H:% Drlang% etc ingle-threaded async im1l

Has congurable 1ersistence

Borks 4ith G- 1airs% 4here a string and may be either

number% string or 6bject strings% hashes% sorted lists%sets% sorted sets

u11orts transactions

%set 'sers:go'

po+erlevel 9/// %get

'sers:go' po+erlevel

Max A. Alexejev21


22/34

/se cases> Redis

Eood 3or xed lists% tagging% ratings% counters%analytics and ueues


23/34

2ase study>2assandra

Britten in Nava% develo1ein Facebook.

5ns1ired by Ama?on

*ynamo re1licationmechanics% but usescolumn-based data mode

Eood 3or logs 1rocessingindex storage% voting% job

storage etc.

0ad 3or transactional1rocessing.

Bant to kno4 more9 AskAlexey8

Max A. Alexejev23


24/34

5n-Memory *ata Erids

Max A. A24

Be+ generation o! cac%ing prod'cts* tr#ing to co"ine

eneDts o! replicated and distri'ted sc%e"es.


25/34

5M*E> Dvolution

Max A. A25

Modern

5M*E

*ata Erids Reliable storage

and live databalancing amonggrid nodes

2om1utationalErids Reliable jobs

execution%scheduling andload balancing


26/34

5M*E> 2aching conce1ts

5m1lements G cache inter3ace

:rovides indexed search by values

:rovides reliable distributed locks inter3ace

2aching scheme J 1artitioned or distributed% may be s1ecied 1er cache or cache service

:rovides events subscri1tion 3or entries


27/34

5M*E> /nder the

hood All data is split in a n'"er o!sections* called partitions.

artition* rat%er t%en entr#* is anato"ic 'nit o! data "igration +%engrid realances. B'"er o! partitionsis Dxed !or cl'ster li!eti"e.

&ndexes are distri'ted a"ong gridnodes.

$lients "a# or "a# not e part o! t%egrid cl'ster.

Max A. Alexejev27


28/34

5M*E /nder the hood>

Reuests routingFor getE and p'tE re'ests:

1. $l'ster "e"er* t%at "aes a re'est* calc'latese# %as% code.

2. artition n'"er is calc'lated 'sing t%is %as%code.

3. Bode is identiDed # partition n'"er.

4. e'est is t%en ro'ted to identiDed node*exec'ted* and res'lts are sent ac to t%e client

"e"er +%o initiated re'est.

For Dlter 'eries:

5. $l'ster "e"er initiating re'ests sends it to allstorage enaled nodes in t%e cl'ster.

6. G'er# is exec'ted on ever# node 'sing distri'tedindexes and partial res'lts are sent to t%ere'esting "e"er.

7. e'esting "e"er "erges partial res'lts locall#.

8. Final res'lt set is ret'rned !ro" Dlter "et%od.

Max A. Alexejev28


29/34

5M*E> Advanced use-cases

Messaging

Ma1-Reduce calculations

2luster-4ide singleton

And moreC

Max A. A29


30/34

E2 tuning 3or large grid nodes

An easy 4ay to go> rolling restarts or storage-enabledcluster nodes. 2an not be used in any 1roject.

A com1lex 4ay to go> ne-tune 2M collector toensure that it 4ill al4ays kee1 u1 cleaning garbage

concurrently under normal 1roduction 4orkload.

An ex1ensive 4ay to go> use 6LHea1 storages1rovided by some vendors


31/34

5M*E> Market 1layers

6racle 2oherence> commercial% 3ree 3or evaluationuse.

Eiga1aces> commercial.

EridEain> commercial.

Ha?elcast> o1en-source.

5nnis1an> o1en-source.

Max A. A31


32/34

Terracotta

Max A. A32

A co"pan# e%ind ?$ac%e* G'art= and >erracotta erver Arra#

Ac'ired # o!t+are AH.


33/34

Terracotta erver Array

Max A. A33

All data is split in a n'"er o! sections* called stripes.

tripes consist o! 2 or "ore >erracotta nodes. -ne o! t%e" is Active node* ot%ers %ave assive stat's.

All data is distri'ted a"ong stripes and replicated inside stripes.

-pen o'rce li"itation: onl# one stripe. 'c% set'p +ill s'pport A* 't +ill not distri'te cac%e data. &.e.* it is not %ori=ontall# scala


34/34

IA essionAnd thank you 3or coming8

distributed caches

Documents