Slide 1

Slide 1 text

stabilizing the registry

Slide 2

Slide 2 text

C J Silverio director of engineering, npm @ceejbot

Slide 3

Slide 3 text

This is the story of a plucky package registry named npm

Slide 4

Slide 4 text

scaling problem manifesting itself as a stability problem

Slide 5

Slide 5 text

"scaling" capacity to meet growing demands

Slide 6

Slide 6 text

"At scale" huge demand & lots of data

Slide 7

Slide 7 text

"stability" not falling over under normal demand

Slide 8

Slide 8 text

What's normal demand?

Slide 9

Slide 9 text

129K packages 239 GB package tarballs 40 million pkg dls/day 1500 req/sec, peak 3200

Slide 10

Slide 10 text

"Legacy" Anything you've put into production

Slide 11

Slide 11 text

this is the story of a legacy system becoming more flexible

Slide 12

Slide 12 text

No content

Slide 13

Slide 13 text

January 2013 20K packages .5 million dls/day

Slide 14

Slide 14 text

Oct 2013 44K packages 108 million dls/month 3.6 million dls/day

Slide 15

Slide 15 text

No content

Slide 16

Slide 16 text

our plucky little registry had to change

Slide 17

Slide 17 text

step 1: CDN Put Fastly.com in front of the registry

Slide 18

Slide 18 text

cache rules everything around me

Slide 19

Slide 19 text

step 2: tarballs get them out of couchdb

Slide 20

Slide 20 text

tarballs are huge! couch runs better without them base64 decoding is work.

Slide 21

Slide 21 text

No content

Slide 22

Slide 22 text

January 2014 60K packages 6+ million dls/day

Slide 23

Slide 23 text

step 3: visibility are things going wrong? what's going wrong?

Slide 24

Slide 24 text

reactive monitoring monitor deeply fix things quickly

Slide 25

Slide 25 text

proactive monitoring self-healing (also things don't break)

Slide 26

Slide 26 text

monitoring is unit testing Add monitoring after every outage

Slide 27

Slide 27 text

visibility is a prerequisite but not a solution

Slide 28

Slide 28 text

act on what monitoring and metrics reveal

Slide 29

Slide 29 text

step 4: redundancy several CouchDBs! reads, writes, & replication

Slide 30

Slide 30 text

fewer responsibilities for each piece isolates errors

Slide 31

Slide 31 text

step 5: automation ansible no server is special

Slide 32

Slide 32 text

June 2014 Superficially similar.

Slide 33

Slide 33 text

June 2014 80K packages 10 million dls/day

Slide 34

Slide 34 text

step 6: simplification now that it's not on fire we can modify at leisure

Slide 35

Slide 35 text

No content

Slide 36

Slide 36 text

Nov 2014 105K packages 28 million dls/day peak

Slide 37

Slide 37 text

50/50 AWS region split no AWS-specific magic Ubuntu 14.04 Trusty

Slide 38

Slide 38 text

Fastly: geoloc + varnish haproxy + CouchDB nginx + a filesystem

Slide 39

Slide 39 text

where's the node?

Slide 40

Slide 40 text

registry 2 electric boogaloo with 500% more node

Slide 41

Slide 41 text

No content

Slide 42

Slide 42 text

haproxy + node services couchdb ➜ postgres redis for caching nginx + filesystem

Slide 43

Slide 43 text

more complicated more flexible & redundant more scaling dials to turn

Slide 44

Slide 44 text

excited about postgres ad-hoc queries are fun

Slide 45

Slide 45 text

scaling node is exactly like scaling everything else

Slide 46

Slide 46 text

Understand system get visibility cool down hot spots add redundancy

Slide 47

Slide 47 text

npm client <3 npm install -g npm@latest

Slide 48

Slide 48 text

npm loves you