1.
Scaling :
What Went Wrong,
What Went Right
Ross Snyder
ross@etsy.com
@beamrider9 Sept. 30, 2011
1

2.
Etsy is the world’s handmade marketplace.
(vintage and supplies, too)
2

3.
Etsy was founded in mid-2005
and is constantly growing.
Gross Merchandise Sales ($MM)
3

4.
From humble beginnings...
June Four employees, one web*,
2005: one db, founder’s apartment
* until getting slashdotted by a link
from Boing Boing in Aug. 2005
4

5.
... to today’s handmade juggernaut.
Sept. 250+ employees, multiple
2011: offices, billions of pageviews
(NYC Mayor Mike Bloomberg visited
Etsy in June 2011)
5

6.
How’d we get here?
6

7.
Answer: with some difficulty.
“There is no education like adversity.” - Benjamin Disraeli
7

8.
A few disclaimers
8

9.
Hindsight is 20/20
9

10.
“History is written by the victors”
10

11.
Etsy thrives today
because of what
its early employees
accomplished
11

12.
Your narrator wasn’t present for most
of the events covered in this talk
12

13.
Etsy Architecture: 2007
13

14.
Etsy Architecture: 2007
Operating System:
Database:
Webserver:
Languages:
14

15.
Etsy Architecture: 2007
Most business logic in
Postgres stored procedures
15

16.
Etsy Architecture: 2007
Front end / database interaction = stored
procedure calls wrapped with PHP functions
16

17.
Etsy Architecture: 2007
Some database partitioning by feature,
but still with a large central DB
17

18.
Etsy Architecture: 2007
Site uptime = not great
18

19.
Etsy Architecture: 2007
“How do we scale?”
19

20.
Etsy Architecture: 2007
“Let’s write some middleware!”
(runners up: “Let’s rewrite the site in Java!”
and “Let’s rewrite the site in Python!”)
20

21.
Conway’s Law:
“Any organization that designs
a system (defined broadly) will
produce a design whose
structure is a copy of the
organization's communication
structure.”
- Melvin Conway, 1968
21

22.
Etsy Engineering: 2007
Dev DBA Ops
22

23.
Etsy Engineering: 2007
Dev DBA Ops
Devs write code
23

24.
Etsy Engineering: 2007
Dev DBA Ops
DBAs write SQL
24

25.
Etsy Engineering: 2007
Dev DBA Ops
Ops deploys code & touches prod
25

26.
SILOS
26

27.
Etsy’s big bet: “Sprouter”
(the Stored Procedure Router)
27

28.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
Runs on each webserver,
listens on port 8010
28

29.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
Maps name/arguments to a Postgres
stored procedure, calls it, returns results
29

30.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
Caches things
30

31.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
Supports sharding (in theory)
31

32.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
Devs write PHP, DBAs write SQL,
meet somewhere in the middle
32

33.
SILOS
33

34.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
The hope: easier to scale Sprouter
than to scale the database itself
34

35.
Sprouter
Web Sprouter DB
(PHP) (Python) (Postgres)
(scaling the db when everything’s in
stored procedures = somewhere between
hard and impossible)
35

36.
Sprouter: Timeline
Fall ’07: Idea first discussed
Spring ’08: Alpha version debuts
Fall ’08: Released in production
36

37.
Sprouter: Timeline
Fall ’07: Idea first discussed
Spring ’08: Alpha version debuts
Fall ’08: Released in production
Spring ’09: Sprouter deprecated
37

38.
What happened?
38

39.
Sprouter: “Good” Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Forcibly centralizes database access
39

40.
Sprouter: “Good” Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Hides data store implementation
from caller
40

41.
Sprouter: “Good” Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Opens the door for
“clever” automatic caching
41

42.
Sprouter: “Good” Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Prevents developers from writing SQL (?)
42

43.
43

44.
Sprouter: Not-As-Good Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Creates substantial developer friction
44

45.
Sprouter: Not-As-Good Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Homegrown daemon + dependencies
for Ops to maintain
45

46.
Sprouter: Not-As-Good Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Lack of community support / provability
46

47.
Sprouter: Not-As-Good Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Complex synchronization required to deploy
(due to tight coupling with Postgres)
47

48.
Sprouter: Not-As-Good Parts
Web Sprouter DB
(PHP) (Python) (Postgres)
Database remains single point of failure
(sharding features never fully formed)
48

49.
Sprouter: Summary
Extra barriers to development
49

50.
Sprouter: Summary
Extra barriers to development
+ Negligible (negative?) effect on site reliability
50

51.
Sprouter: Summary
Extra barriers to development
+ Negligible (negative?) effect on site reliability
+ Deploys even more painful
51

52.
Sprouter: Summary
Extra barriers to development
+ Negligible (negative?) effect on site reliability
+ Deploys even more painful
+ Requires extra Ops/Dev resources
52

53.
Sprouter: Summary
Extra barriers to development
+ Negligible (negative?) effect on site reliability
+ Deploys even more painful
+ Requires extra Ops/Dev resources
=
53

54.
How did attitudes change so quickly?
54

55.
Sprouter: Timeline
Fall ’07: Idea first discussed
Spring ’08: Alpha version debuts
Fall ’08: Released in production
Spring ’09: Sprouter deprecated
55

56.
The Great Etsy Culture Shift
56

57.
The Great Etsy Culture Shift
Just as Sprouter went live, many of its
strongest proponents departed Etsy
57

58.
The Great Etsy Culture Shift
Taking with them...
58

59.
The Great Etsy Culture Shift
Devotion to Postgres stored
procedures / types
59

60.
The Great Etsy Culture Shift
Fear of developers writing SQL
60

61.
The Great Etsy Culture Shift
Fear of developers touching prod
61

62.
The Great Etsy Culture Shift
Infrequent / large deploys to production
62

63.
The Great Etsy Culture Shift
“Not developed here”
63

64.
The Great Etsy Culture Shift
Then Now
Fall
’08
64

65.
DevOps
65

66.
DevOps
Silos = bad
66

67.
DevOps
Trust, cooperation, transparency,
shared responsibility = good
67

68.
DevOps
“We’re all in this together”
68

69.
The Way Forward: Part 1
Stabilize the site
69

70.
The Way Forward: Part 1
Stabilize the site
Improve metrics & monitoring
70

71.
The Way Forward: Part 1
Stabilize the site
StatsD
http://github.com/etsy/statsd
71

72.
The Way Forward: Part 1
Stabilize the site
Upgrade database hardware
vertically as far as possible
72

73.
The Way Forward: Part 1
Stabilize the site
Give developers production access to
help troubleshoot problems
73

74.
The Way Forward: Part 2
Continuous Deployment
74

75.
The Way Forward: Part 2
Continuous Deployment
Any engineer can deploy to prod
(generally happens 25+ times per day)
75

76.
The Way Forward: Part 2
Continuous Deployment
Deployinator
http://github.com/etsy/deployinator
76

77.
The Way Forward: Part 2
Continuous Deployment
One button that deploys the site
77

78.
The Way Forward: Part 2
Continuous Deployment
Small changesets, deployed frequently
78

79.
The Way Forward: Part 2
Continuous Deployment
Requires solid tests,
good communication
79

80.
The Way Forward: Part 2
Continuous Deployment
Distributed developer-driven QA
80

81.
The Way Forward: Part 3
Circumvent Sprouter
81

82.
The Way Forward: Part 3
Circumvent Sprouter
Object-Relational Mapping (ORM)
82

83.
The Way Forward: Part 3
Circumvent Sprouter
aka “The Vietnam of Computer Science”
(Google it)
83

84.
The Way Forward: Part 3
Circumvent Sprouter
Front-end PHP talks directly to database
via ORM (also written in PHP)
84

85.
The Way Forward: Part 3
Circumvent Sprouter
ORM can cache where appropriate
(as can front end)
85

86.
The Way Forward: Part 4
Database Sharding
86

87.
The Way Forward: Part 4
Database Sharding
Etsy has a lot of DNA from flickr -
including their DB sharding scheme
87

88.
The Way Forward: Part 4
Database Sharding
Based on MySQL
88

89.
The Way Forward: Part 4
Database Sharding
Battle-tested, well-known
89

90.
The Way Forward: Part 4
Database Sharding
Scales horizontally to infinity
(or close enough)
90

91.
The Way Forward: Part 4
Database Sharding
No single points of failure
(master-master replication)
91

92.
The Way Forward: Part 4
Database Sharding
Gradually phase out Sprouter,
phase in ORM / sharded data
92

93.
Sprouter: Timeline
Fall ’07: Idea first discussed
Spring ’08: Alpha version debuts
Fall ’08: Released in production
Spring ’09: Sprouter deprecated
93

94.
Sprouter: Timeline
Fall ’07: Idea first discussed
Spring ’08: Alpha version debuts
Fall ’08: Released in production
Spring ’09: Sprouter deprecated
Spring ’11: Sprouter turned off
94

95.
95

96.
Lessons Learned
96

97.
Etsy Architecture: 2007
Operating System:
Database:
Webserver:
Languages:
97

98.
Etsy Architecture: 2011
Operating System:
Database:
Webserver:
Languages:
98

99.
Open & trusting > closed & afraid
(DevOps DevOps DevOps)
99

100.
Front end/database interaction is too critical
to take chances on novel/untested solutions
100

101.
Side corollary: If you’re doing something
“clever”, you’re probably doing it wrong
101

102.
The architectural decisions you make today
will have large impact long after you’re gone
102

103.
No architectural hole is so deep that proven
scaling strategies don’t exist for digging out
103

104.
Acknowledgement
We are probably making decisions today that
will be the subject of a similar talk in 2015
104

105.
Learn More:
http://codeascraft.etsy.com/
@codeascraft
105

106.
Etsy is hiring!
http://www.etsy.com/careers
@etsy
106