Slide 1

Slide 1 text

Development was the easy part

Slide 2

Slide 2 text

André Arko @indirect

Slide 3

Slide 3 text

No content

Slide 4

Slide 4 text

No content

Slide 5

Slide 5 text

Development is very different

Slide 6

Slide 6 text

from Production

Slide 7

Slide 7 text

you rn →

Slide 8

Slide 8 text

you later →

Slide 9

Slide 9 text

Metrics

Slide 10

Slide 10 text

Metrics are important

Slide 11

Slide 11 text

Metrics tell you what is happening

Slide 12

Slide 12 text

Metrics convince you you understand

Slide 13

Slide 13 text

Averages convince you you understand

Slide 14

Slide 14 text

Averages are lie-candy for your brain

Slide 15

Slide 15 text

Averages 5 -5 -4 -3 -2 -1 0 1 2 3 4 0 0.1 0.2 0.3 0.4

Slide 16

Slide 16 text

Averages 5 -5 -4 -3 -2 -1 0 1 2 3 4 0 0.1 0.2 0.3 0.4

Slide 17

Slide 17 text

No content

Slide 18

Slide 18 text

No content

Slide 19

Slide 19 text

just heard “we have a great average” →

Slide 20

Slide 20 text

Averages mask problems

Slide 21

Slide 21 text

10 0 1 2 3 4 5 6 7 8 9 250 0 50 100 150 200

Slide 22

Slide 22 text

Graph the median

Slide 23

Slide 23 text

10 0 1 2 3 4 5 6 7 8 9 250 0 50 100 150 200

Slide 24

Slide 24 text

Graph 95th percentile

Slide 25

Slide 25 text

10 0 1 2 3 4 5 6 7 8 9 250 0 50 100 150 200

Slide 26

Slide 26 text

Graph 99th percentile

Slide 27

Slide 27 text

10 0 1 2 3 4 5 6 7 8 9 1000 0 250 500 750

Slide 28

Slide 28 text

Aggregate graphs another kind of average

Slide 29

Slide 29 text

No content

Slide 30

Slide 30 text

Breakout graphs see individual variations

Slide 31

Slide 31 text

No content

Slide 32

Slide 32 text

No content

Slide 33

Slide 33 text

Aggregate alerts more dead servers than alive servers

Slide 34

Slide 34 text

site’s up if any servers are up!

Slide 35

Slide 35 text

Breakout alerts first dead server not all the servers

Slide 36

Slide 36 text

Servers

Slide 37

Slide 37 text

Servers you have no idea what is going on

Slide 38

Slide 38 text

really.

Slide 39

Slide 39 text

Routing

Slide 40

Slide 40 text

Routing your app has this

Slide 41

Slide 41 text

Routing how does it work?

Slide 42

Slide 42 text

Development App You

Slide 43

Slide 43 text

Production People Router Server App App Router Server App App Router

Slide 44

Slide 44 text

Routing how slow is it?

Slide 45

Slide 45 text

Routing does it back up?

Slide 46

Slide 46 text

Request time

Slide 47

Slide 47 text

Request time not the time you measure

Slide 48

Slide 48 text

Request time wall-clock time from real clients

Slide 49

Slide 49 text

Request time make requests from around the world

Slide 50

Slide 50 text

Request time graph them

Slide 51

Slide 51 text

Request time graph them alert on them

Slide 52

Slide 52 text

Request time graph them alert on them thank me later

Slide 53

Slide 53 text

VM lag

Slide 54

Slide 54 text

VM lag do you have it?

Slide 55

Slide 55 text

VM lag do you check for it?

Slide 56

Slide 56 text

VM lag do you know how to check for it?

Slide 57

Slide 57 text

Runtime lag

Slide 58

Slide 58 text

Runtime lag how do you tell you lost consciousness?

Slide 59

Slide 59 text

Runtime lag do you have it?

Slide 60

Slide 60 text

Runtime lag do you have it? you have it.

Slide 61

Slide 61 text

Runtime lag do you have it? you have it. how bad is it?

Slide 62

Slide 62 text

Data stores

Slide 63

Slide 63 text

Data stores in production

Slide 64

Slide 64 text

Data stores in production are distributed

Slide 65

Slide 65 text

what does that mean?

Slide 66

Slide 66 text

your experience (so far) is wrong

Slide 67

Slide 67 text

Saving data

Slide 68

Slide 68 text

Saving data tries to save your data

Slide 69

Slide 69 text

Saving data might save your data

Slide 70

Slide 70 text

Replication

Slide 71

Slide 71 text

Replication is not data- saving magic

Slide 72

Slide 72 text

Replication tries to save your data…

Slide 73

Slide 73 text

Replication tries to save your data… repeatedly

Slide 74

Slide 74 text

Postgres

Slide 75

Slide 75 text

Postgres totally safe, right?

Slide 76

Slide 76 text

Postgres async replication

Slide 77

Slide 77 text

Postgres network failures lose “saved” data

Slide 78

Slide 78 text

Redis

Slide 79

Slide 79 text

Redis is single-threaded

Slide 80

Slide 80 text

Redis has no failover

Slide 81

Slide 81 text

Redis-sentinel elects a new leader

Slide 82

Slide 82 text

Redis-sentinel throws away non- winners’ writes

Slide 83

Slide 83 text

Mongo (gem < 1.8) returns before the first write

Slide 84

Slide 84 text

Mongo (gem < 1.8) your data is on zero disks so far

Slide 85

Slide 85 text

Mongo replication sets default to one write

Slide 86

Slide 86 text

Mongo demand N copies survive N-1 failures

Slide 87

Slide 87 text

trust no one

Slide 88

Slide 88 text

if you didn’t try it you are guessing

Slide 89

Slide 89 text

test it yourself

Slide 90

Slide 90 text

So, in the end what did we learn?

Slide 91

Slide 91 text

Production is fundamentally

Slide 92

Slide 92 text

Production is fundamentally systemically

Slide 93

Slide 93 text

Production is fundamentally systemically different

Slide 94

Slide 94 text

Failures will happen

Slide 95

Slide 95 text

Failures can be resisted

Slide 96

Slide 96 text

Failures should not result in one-off patches

Slide 97

Slide 97 text

Survival requires systematic trials & testing

Slide 98

Slide 98 text

Development is not like production