Amit Ramesh, Qui Nguyen - Building Stream Processing Applications

Slide 1

Slide 1 text

Building Stream Processing Applications Amit Ramesh Qui Nguyen

Slide 2

Slide 2 text

Yelp’s Mission Connecting people with great local businesses.

Slide 3

Slide 3 text

I. Why stream processing? II. Putting an application together Example problem Components and data operations III. Design principles and tradeoffs Horizontal scalability Handling failures Idempotency Consistency versus availability

Slide 4

Slide 4 text

Slide 5

Slide 5 text

No content

Slide 6

Slide 6 text

Data processing measurements from a sensor clicking on ads

Slide 7

Slide 7 text

Data processing measurements from a sensor clicking on ads average value in the last minute total clicks on a day

Slide 8

Slide 8 text

Batch Finite chunk of data Operations defined over the entire input Data processing: Batch or stream 8

Slide 9

Slide 9 text

Batch Finite chunk of data Operations defined over the entire input Stream Unbounded stream of events flowing in Events are processed continuously (possibly with state) Data processing: Batch or stream 9

Slide 10

Slide 10 text

Why stream processing over batch? ● Lower latency on results ● Most data is unbounded, so streaming model is more flexible

Slide 11

Slide 11 text

Why stream processing over batch? ● Lower latency on results ● Most data is unbounded, so streaming model is more flexible Day 12 Day 13

Slide 12

Slide 12 text

Our evolution

Slide 13

Slide 13 text

Our evolution

Slide 14

Slide 14 text

Our evolution

Slide 15

Slide 15 text

Our evolution mrjob

Slide 16

Slide 16 text

Our evolution

Slide 17

Slide 17 text

Our evolution

Slide 18

Slide 18 text

Slide 19

Slide 19 text

Example problem: ad campaign metrics Ad Yelp

Slide 20

Slide 20 text

ad { id: 1200834, campaign_id: 2001, user_id: 9zkjacn81m, timestamp: 1490732147 } view { id: 1200834, timestamp: 1490732150 } click { id: 1200834, timestamp: 1490732168 }

Slide 21

Slide 21 text

Metrics (views, clicks) for each campaign over time Ad Yelp

Slide 22

Slide 22 text

Slide 23

Slide 23 text

Source of streaming data Stream processing pipelines Stream processing engine Storage Data sink

Slide 24

Slide 24 text

Stream processing pipelines Stream processing engine Storage Data sink Source of streaming data

Slide 25

Slide 25 text

Types of operations 1. Ingestion 2. Stateless transforms 3. Stateful transforms 4. Keyed stateful transforms 5. Publishing

Slide 26

Slide 26 text

Operations: 1. Ingestion Kafka Reader Operation Source

Slide 27

Slide 27 text

Operations: 1. Ingestion Kafka Reader Operation Source from pyspark.streaming.kafka import KafkaUtils ad_stream = KafkaUtils.createDirectStream( streaming_context, topics=[‘ad_events’], kafkaParams={...}, )

Slide 28

Slide 28 text

Operations: 2. Stateless transforms Operation Transform Operation

Slide 29

Slide 29 text

Operations: 2a. Stateless transforms Filter e.g., filtering

Slide 30

Slide 30 text

Operations: 2a. Stateless transforms Filter e.g., filtering def is_not_from_bot(event): return event[‘ip’] not in bot_ips filtered_stream = ad_stream.filter(is_not_from_bot)

Slide 31

Slide 31 text

Operations: 2b. Stateless transforms Project e.g., projection

Slide 32

Slide 32 text

Operations: 2b. Stateless transforms Project e.g., projection desired_fields = [‘ad_id’, ‘campaign_id’] def trim_event(event): return {key: event[key] for key in desired_fields} projected_stream = ad_stream.map(trim_event)

Slide 33

Slide 33 text

Operations: 3. Stateful transforms On windows of data Transform Sliding window

Slide 34

Slide 34 text

Operations: 3. Stateful transforms On windows of data Transform Sliding window Tumbling window

Slide 35

Slide 35 text

Operations: 3. Stateful transforms e.g., aggregation Sum 5 6 0 1 1 3 0 1 2

Slide 36

Slide 36 text

Operations: 3. Stateful transforms e.g., aggregation Sum 5 6 0 1 1 3 0 1 2 aggregated_stream = event_stream.reduceByWindow( func=operator.add, windowLength=4, slideInterval=3, )

Slide 37

Slide 37 text

Operations: 4. Keyed stateful transforms Shuffle Group events by key (shuffle) within each window before transform Transform

Slide 38

Slide 38 text

Operations: 4a. Keyed stateful transforms c_id: 1 views: 1 c_id: 2 views: 2 c_id: 1 views: 1 c_id: 2 views: 1 c_id: 2 views: 1 sum views by c_id e.g., aggregate views by campaign_id

Slide 39

Slide 39 text

Operations: 4a. Keyed stateful transforms e.g., aggregate views by campaign_id aggregated_views = view_stream.reduceByKeyAndWindow( func=operator.add, windowLength=3, slideInterval=3, ) c_id: 1 views: 1 c_id: 2 views: 2 c_id: 1 views: 1 c_id: 2 views: 1 c_id: 2 views: 1 sum views by c_id

Slide 40

Slide 40 text

Operations: 4b. Keyed stateful transforms Can also be on more than one stream, e.g., join by id Shuffle Join

Slide 41

Slide 41 text

Operations: 4b. Keyed stateful transforms e.g., join by ad_id Join by ad_id Ad ad_id: 11 c_id: 1 ad_id: 22 c_id: 2 ad_id: 22 time: 5 ad_id: 11 time: 7 ad_id: 11 ad: { c_id: 1 }, view: { time: 7 } ad_id: 22 ad: { c_id: 2 }, view: { time: 5 }

Slide 42

Slide 42 text

Operations: 4b. Keyed stateful transforms windowed_ad_stream = ad_stream.window( windowLength=2, slideInterval=2, ) windowed_view_stream = view_stream.window( windowLength=2, slideInterval=2, ) joined_stream = windowed_ad_stream.join( windowed_view_stream, ) e.g., join by ad_id

Slide 43

Slide 43 text

Operations: 5. Publishing Sink File writer Operation

Slide 44

Slide 44 text

Operations: 5. Publishing results_stream.saveAsTextFiles(‘s3://my.bucket/results/’) File writer Operation Sink

Slide 45

Slide 45 text

Operations: Summary 1. Ingestion 2. Stateless transforms: on single events a. Filtering b. Projections 3. Stateful transforms: on windows of events 4. Keyed stateful transforms a. On single streams, transform by key b. Join events from several streams by key 5. Publishing

Slide 46

Slide 46 text

Putting it together: campaign metrics Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project

Slide 47

Slide 47 text

read Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project { ip: bot_id, ... } { ip: OK_id, ... }

Slide 48

Slide 48 text

filter Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project { ip: bot_id, ... } { ip: OK_id, ... }

Slide 49

Slide 49 text

project Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project { ip: OK_id, scoring: { ... }, ... }

Slide 50

Slide 50 text

project Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project { ip: OK_id, scoring: { ... }, ... }

Slide 51

Slide 51 text

join by ad id er join by ad id transform write sum by campaign project transform write er project er project { ad_id: 1, ad_data: ... } { ad_id: 1, view_data: ... }

Slide 52

Slide 52 text

join by ad id er join by ad id transform write sum by campaign project transform write er project er project { ad_id: 1, ad_data: ..., view_data: ..., }

Slide 53

Slide 53 text

transform er join by ad id transform write sum by campaign project transform write er project er project { ad_id: 1, campaign_id: 7, view: true, click: false }

Slide 54

Slide 54 text

sum by campaign oin by ad id transform write sum by campaign transform write { ad_id: 1, campaign_id: 7, view: true, click: false } { ad_id: 23, campaign_id: 7, view: true, click: false }

Slide 55

Slide 55 text

sum by campaign oin by ad id transform write sum by campaign transform write { campaign_id: 7, views: 2, clicks: 0 }

Slide 56

Slide 56 text

write db.write( campaign_id=7, views=2, clicks=0, ) m write sum by campaign m write

Slide 57

Slide 57 text

Ad campaign metrics pipeline Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project

Slide 58

Slide 58 text

Slide 59

Slide 59 text

Horizontal scalability: Basic idea

Slide 60

Slide 60 text

Horizontal scalability: Basic idea

Slide 61

Slide 61 text

Horizontal scalability: Basic idea

Slide 62

Slide 62 text

Horizontal scalability: Basic idea

Slide 63

Slide 63 text

Horizontal scalability: Why?

Slide 64

Slide 64 text

Horizontal scalability: Why?

Slide 65

Slide 65 text

Horizontal scalability: How? Random partitioning Partitioning

Slide 66

Slide 66 text

Horizontal scalability: How? Ad read read read filter filter filter project project project read read read filter filter filter project project project Partitioning Random partitioning

Slide 67

Slide 67 text

project project project join by ad id Horizontal scalability: How? Partitioning

Slide 68

Slide 68 text

project project project join by ad id Horizontal scalability: How? Partitioning Keyed partitioning

Slide 69

Slide 69 text

Horizontal scalability: watch out! Hot spots / data skew transform sum by campaign transform

Slide 70

Slide 70 text

Horizontal scalability: watch out! Hot spots / data skew Keyed partitioning transform sum by campaign transform

Slide 71

Slide 71 text

Horizontal scalability: Summary ● Random partitioning for stateless transforms ● Keyed partitioning for keyed transformations ● Watch out for hot spots, and use appropriate mitigation strategy

Slide 72

Slide 72 text

Slide 73

Slide 73 text

Idempotency

Slide 74

Slide 74 text

Idempotency An idempotent operation can be applied more than once and have the same effect.

Slide 75

Slide 75 text

Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project

Slide 76

Slide 76 text

Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project project write

Slide 77

Slide 77 text

What operations are idempotent? Transforms: filters, projections, etc No side effects! Stateful operations

Slide 78

Slide 78 text

Ad filter read join by ad id transform write sum by campaign project transform write filter read project filter read project project write

Slide 79

Slide 79 text

Idempotent writes with unique keys campaign_id = 7, minute = 20, views = 2 campaign _id minute views 7 20 2 campaign_id = 7, minute = 20, views = 2

Slide 80

Slide 80 text

Writes that aren’t idempotent campaign _id hour views 7 2 0

Slide 81

Slide 81 text

Writes that aren’t idempotent campaign_id = 7, hour = 2, views += 1 campaign _id hour views 7 2 1

Slide 82

Slide 82 text

Writes that aren’t idempotent campaign_id = 7, hour = 2, views += 1 campaign _id hour views 7 2 2 campaign_id = 7, hour = 2, views += 1

Slide 83

Slide 83 text

Support for idempotency campaign_id = 7, hour = 2, views += 1, version = 1 campaign _id hour views 7 2 1 campaign_id = 7, hour = 2, views += 1 version = 1

Slide 84

Slide 84 text

Idempotency in streaming pipelines Both in output to data sink and in local state (joining, aggregation) Re-processing of events - Some frameworks provide exactly once guarantees

Slide 85

Slide 85 text

Consistency vs. availability

Slide 86

Slide 86 text

Always a tradeoff between consistency and availability when handling failures

Slide 87

Slide 87 text

Consistency Every read sees a current view of the data. Availability Capacity to serve requests

Slide 88

Slide 88 text

A = 9 A = 9

Slide 89

Slide 89 text

A = 3 A = 3 A = 3 A = 3

Slide 90

Slide 90 text

A = 9 A = 9

Slide 91

Slide 91 text

A = 9 A = 9 Consistency > availability A = 3 A = 3

Slide 92

Slide 92 text

A = 9 A = 9 Consistency > availability A = 3 Error: write unavailable

Slide 93

Slide 93 text

A = 9 A = 9 Availability > consistency A = 3 A = 3

Slide 94

Slide 94 text

A = 9 A = 3 Availability > consistency Not consistent: 3 != 9

Slide 95

Slide 95 text

Prioritizing consistency or availability Applies to systems for both your data source and data sink Source Stream processing engine Data sink Storage

Slide 96

Slide 96 text

Prioritizing consistency or availability Applies to systems for both your data source and data sink ● Some systems pick one, be aware ● Others let you choose ○ ex. Cassandra - how many replicas respond to write? Streaming applications run continuously

Slide 97

Slide 97 text

Prioritizing consistency or availability Depends on the needs of your application Metrics (views, clicks) for each campaign over time

Slide 98

Slide 98 text

Prioritizing consistency or availability More consistency Metrics (views, clicks) for each campaign over time

Slide 99

Slide 99 text

Prioritizing consistency or availability More availability Internal graphs Metrics (views, clicks) for each campaign over time

Slide 100

Slide 100 text

Conclusion ● Stream processing: data processing with operations on events or windows of events ● Horizontal scalability, as data will grow and change over time ● Handle failures appropriately ○ Keep operations idempotent, for retries ○ Tradeoff between availability and consistency