Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
LXJS 2013: backpack — scalable photo storage
Search
Ivan Babrou
October 02, 2013
Programming
2
150
LXJS 2013: backpack — scalable photo storage
http://bobrik.name/talks/lxjs2013.pdf
— slides with notes.
http://youtu.be/T4DgxvS9Xho
— video.
Ivan Babrou
October 02, 2013
Tweet
Share
More Decks by Ivan Babrou
See All by Ivan Babrou
node.js for millions of images
bobrik
7
1.4k
Other Decks in Programming
See All in Programming
複雑なドメインに挑む.pdf
yukisakai1225
5
1.2k
詳解!defer panic recover のしくみ / Understanding defer, panic, and recover
convto
0
240
Design Foundational Data Engineering Observability
sucitw
3
200
How Android Uses Data Structures Behind The Scenes
l2hyunwoo
0
460
1から理解するWeb Push
dora1998
7
1.9k
ファインディ株式会社におけるMCP活用とサービス開発
starfish719
0
1.6k
Navigating Dependency Injection with Metro
zacsweers
3
960
ぬるぬる動かせ! Riveでアニメーション実装🐾
kno3a87
1
220
AIを活用し、今後に備えるための技術知識 / Basic Knowledge to Utilize AI
kishida
22
5.8k
もうちょっといいRubyプロファイラを作りたい (2025)
osyoyu
1
440
AIでLINEスタンプを作ってみた
eycjur
1
230
RDoc meets YARD
okuramasafumi
4
170
Featured
See All Featured
Building a Modern Day E-commerce SEO Strategy
aleyda
43
7.6k
Side Projects
sachag
455
43k
The Invisible Side of Design
smashingmag
301
51k
Building Adaptive Systems
keathley
43
2.7k
Optimising Largest Contentful Paint
csswizardry
37
3.4k
Building Flexible Design Systems
yeseniaperezcruz
328
39k
The World Runs on Bad Software
bkeepers
PRO
70
11k
Building Applications with DynamoDB
mza
96
6.6k
StorybookのUI Testing Handbookを読んだ
zakiyama
31
6.1k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
285
13k
YesSQL, Process and Tooling at Scale
rocio
173
14k
jQuery: Nuts, Bolts and Bling
dougneiner
64
7.9k
Transcript
HI THERE, LXJS
% whoami Ian Babrou, Topface.com
60+ million users 100+ million photos
16 photos on main page up to 200 in feed
many small “previews”
This talk is about PHOTOS
Let’s look at some more numbers
12 storage nodes 70TB total space 44TB used
250 TB per month 1.6 Gbps peak 850 Mbps average
powered by node.js & nginx! open-source FTW
ARCHITECTURE aka part 1
frontend cache resizer storage
of course there are frontends, probably more than one frontend
frontend frontend frontend
round-robin dns + ipvs, probably frontend frontend frontend frontend
NGINX ngx_http_upstream_hash_module is your friend
NGINX ngx_http_upstream_hash_module because you need more than one cache, right?
#protip don’t cache anything twice
don’t do: frontend cache cache cache file#3 file#1 file#2 file#1
file#2 file#2 file#3 file#3 file#1
do: frontend cache cache cache file#3 file#1 file#2
NGINX + SSD is just great for caching, forget about
tmpfs
#protip overallocate caches
RESIZING resizing on the fly saves disk, but eats cpu
NGINX ngx_http_image_filter is your friend
BACKPACK aka part 2
first try nginx
okay for 1k files
okay for 10k files
okay for 50k files
okay until you fit in memory or have ssd
RANDOM ACCESS
DISKS ARE SPINNING
node.js to the rescue!
... and redis
... and zookeeper
simple idea: no extra fseek(3)
inspired by haystack from facebook
concatenate small files into bigger
always keep index in memory
REALIZATION ON DISK
3.5 gb files as many as you need
index for each name:offset:length name:offset:length name:offset:length
but.. no worries, this is only needed if redis goes
crazy
REALIZATION IN MEMORY
keys for files name -> file:offset:length name -> file:offset:length name
-> file:offset:length
redis 3.5gb data + index 3.5gb data + index 3.5gb
data + index memory disk all together:
POWERED BY node.js looks like webdav
PUT: 1. write data 2. write index 3. write redis
key
GET: 1. read redis key 2. read data data files
are always open!
let’s read 100K files! 0 37,5 75 112,5 150 backpack
nginx
LESS SEEKS LEAD TO BIGGER THROUGHPUT
BONUS! linearized access for processing
BUT WHAT ABOUT FUTURE?
NO MORE MEMORY vs DISK* Probably, someday.
MANAGEMENT aka part 3
1. adding servers 2. replication 3. failover
COORDINATOR
COORDINATOR combines servers into shards
COORDINATOR that’s where we need zookeeper
I KNOW let’s use DHT! like dynamo!
Rebalancing on capacity change
NO!
NO. THANK YOU!
LET’S MAKE IT SIMPLE
SHARDS (aka buckets) backpack #1 backpack #2 backpack #3 backpack
#4 backpack #5 backpack #6 shard #1 (50%) shard #2 (50%) 1:lol.jpg 2:wtf.jpg
ADDING SHARD backpack #1 backpack #3 backpack #3 backpack #4
backpack #5 backpack #6 1:lol.jpg 1:wtf.jpg shard #1 (50%) shard #2 (50%) backpack #7 backpack #8 backpack #9 50% chance shard #3 (0%)
COORDINATOR knows how to handle next file
NO REBALANCING SIMPLE
REPLICATOR
WHAT IF METEORITE WILL HIT YOUR NODE?
IT HAPPENS. YOU NEED TO ACCEPT THAT.
REPLICATOR to the rescue!
make multi-node SHARDS
DISTRIBUTE SHARDS ACROSS SERVERS
backpack #1 shard #1 lol.jpg backpack #1 lol.jpg backpack #1
lol.jpg server #1 server #1 server #1 coordinator replicator
REPLICATOR EVENTUALLY MAKES COPIES
THE WHOLE THING IS BULLET-PROOF IF YOU NEED IT
backpack #1 backpack #4 backpack #2 backpack #5 backpack #3
backpack #6 backpack #7 backpack #8 backpack #9 server #1 server #2 server #3 zookeeper #1 zookeeper #2 zookeeper #3 redis-queue #1 redis-queue #2 redis-queue #3 coordinator #1 coordinator #2 coordinator #3 replicator #1 replicator #2 replicator #3
GET THE CODE /Topface/backpack npm install backpack{,-coordinator,-replicator}
That’s it! bobrik ibobrik