Building a Sustainable Data Platform on AWS
50 likes19,357 views
This document presents a case study on building a sustainable data platform using AWS at SmartNews, a news discovery app. It discusses the architecture, design principles, and technologies employed, such as Lambda architecture, Amazon EMR, and Presto, to handle large-scale data processing and analysis efficiently. The emphasis is on creating a reliable and scalable system that minimizes operational costs while enabling self-service for engineers and straightforward data querying.
![Forwards to S3
<source>
@type tail
format json
path /data/log/user_activity.log
pos_file /data/log/pos/user_activity.pos
tag smartnews.user_activity
time_key timestamp
</source>
<match smartnews.user_activity>
@type copy
<store>
@type relabel
@label @s3
</store>
<store>
@type forward
@label @forward
</store>
</match>
@include conf.d/s3.conf
@include conf.d/forward.conf
<label @s3>
<% node[:td_agent][:s3].each do |c| -%>
<match <%= c[:tag] %>>
@id s3.<%= c[:tag] %>
@type s3
...
path fluentd/<%= node[:env] %>/<%= node[:role] %>/<%= c[:tag] %>
time_slice_format dt=%Y-%m-%d/hh=%H
time_key timestamp
include_time_key
time_as_epoch
reduced_redundancy true
format json
utc
buffer_chunk_limit 2048m
</match>
<% end -%>
</label>
td-agent.conf conf.d/s3.conf](https://image.slidesharecdn.com/20160127awsnightv2-160127102625/85/Building-a-Sustainable-Data-Platform-on-AWS-22-320.jpg)
![JSON-SerDe
-- Define table with SERDE
CREATE TABLE json_table (
country string,
languages array<string>,
religions map<string,array<int>>
)
ROW FORMAT SERDE 'org.openx.data.jsonserde.JsonSerDe'
STORED AS TEXTFILE;
-- Result: 10
SELECT religions['catholic'][0] FROM json_table;](https://image.slidesharecdn.com/20160127awsnightv2-160127102625/85/Building-a-Sustainable-Data-Platform-on-AWS-32-320.jpg)
![cf. hive-ruby-scripting
-- Define your ruby (JRuby) script
SET rb.script=
require 'json'
def parse (json)
j = JSON.load(json)
j['profile']['attribute1']
end
;
-- Use the script in HQL
SELECT rb_exec('&parse', json) FROM user;
https://github.com/gree/hive-ruby-scripting](https://image.slidesharecdn.com/20160127awsnightv2-160127102625/85/Building-a-Sustainable-Data-Platform-on-AWS-33-320.jpg)
![Use case: A/B Test
-- Suppose that this table exists
DESC hive.default.user_activities;
user_id bigint
action varchar
abtest array<map<varchar, bigint>>
url varchar
-- Summarize page view per A/B Test identifier
-- for comparing two algorithms v1 & v2
SELECT
dt,
t['behaviorId'],
count(*) as pv
FROM hive.default.user_activities CROSS JOIN UNNEST(abtest) AS t (t)
WHERE dt like '2016-01-%' AND action = 'viewArticle'
AND t['definitionId'] = 163
GROUP BY dt, t['behaviorId'] ORDER BY dt
;
2015-12-01 | algorithm_v1 | 40000
2015-12-01 | algorithm_v2 | 62000](https://image.slidesharecdn.com/20160127awsnightv2-160127102625/85/Building-a-Sustainable-Data-Platform-on-AWS-57-320.jpg)
![Scheduled Auto Scaling
$ aws autoscaling describe-scheduled-actions
{
"ScheduledUpdateGroupActions": [
{
"DesiredCapacity": 2,
"AutoScalingGroupName": "presto-worker-prd",
"Recurrence": "59 14 * * *",
"ScheduledActionName": "scalein-2359-jst"
},
{
"DesiredCapacity": 20,
"AutoScalingGroupName": "presto-worker-prd",
"Recurrence": "45 0 * * 1-5",
"ScheduledActionName": "scaleout-0945-jst"
}
]
}](https://image.slidesharecdn.com/20160127awsnightv2-160127102625/85/Building-a-Sustainable-Data-Platform-on-AWS-59-320.jpg)
![Continuous View
-- Calculate unique users seen per media each day
-- Using only a constant amount of space (HyperLogLog)
CREATE CONTINUOUS VIEW uniques AS
SELECT
day(arrival_timestamp),
substring(url from '.*://([^/]*)') as hostname,
COUNT(DISTINCT user_id::integer)
FROM activity_stream GROUP BY day,hostname;
-- How many impressions have we served in the last five minutes?
CREATE CONTINUOUS VIEW imps WITH (max_age = '5 minutes') AS
SELECT COUNT(*) FROM imps_stream;
-- What are the 90th, 95th, 99th percentiles of request latency?
CREATE CONTINUOUS VIEW latency AS
SELECT
percentile_cont(array[90, 95, 99])
WITHIN GROUP (ORDER BY latency::integer)
FROM latency_stream;](https://image.slidesharecdn.com/20160127awsnightv2-160127102625/85/Building-a-Sustainable-Data-Platform-on-AWS-77-320.jpg)
Building a Sustainable Data Platform on AWS
- 1. Building a Sustainable Data Platform on AWS Takumi Sakamoto 2016.01.27
- 2. Takumi Sakamoto @takus 😍 = ⚽ ✈ 📷
- 7. What is SmartNews? • News Discovery App • Launched in 2012 • 15M+ Downloads in World Wide https://www.smartnews.com/en/
- 8. Our Mission the world's quality information? the people who need it? How?
- 9. Machine Learning URLs Found Structure Analysis Semantics Analysis Importance Estimation Diversification Internet 100,000+ /day 1000+ /day Feedback Deliver Trending Stories
- 10. Data Platform Use Cases • Product development • track KPI such as DAU and MAU • A/B test for new feature, on-boarding, etc... • ad-hoc analysis • Provide data to applications • realtime re-ranking news articles • CTR prediction of Ads system • dashboard service for media partners
- 11. Data & Its Numbers • User activities • ~100 GBs per day (compressed) • 60+ record types • User demographics or configurations etc... • 15M+ records • Articles metadata • 100K+ records per day
- 13. Sustainable Data Platform • Provide a reliable and scalable "Lambda Architecture" • Minimize both operation & running cost • Be open to uncertain future
- 15. Why Sustainable? • Do a lot with a few engineers • no one is a full-time maintainer • avoid to waste too much time • Empower brilliant engineers in SmartNews • everything should be as self-serve as possible • don't ask for permission, beg for forgiveness
- 16. System Design
- 17. λ Architecture at SmartNews Input Batch Serving Speed Output
- 18. Design Principles • Decoupled "Computation" and "Storage" layers • multiple consumers can use the same data • run consumers on Spot Instances • prevent serious data lost with minimum effort • Use the right tool for the job • leverage AWS managed service as possible • fill in the missing pieces by Presto & PipelineDB
- 19. An Example Amazon EMR AMI 3.x Amazon S3 Amazon EMR Hive General Users Application Engineer I wanna upgrade hive Ad Engineer I wanna combine news data with ad data Amazon EMR AMI 4.x Amazon EMR Spark We’re satisfied with current version Data Scientist I wanna test my algorithm with the latest spark Batch Layer Run multiple EMR clusters for each usages Kinesis Stream Spark on EMR AWS Lambda Data Scientist I wanna consume streaming data by Spark Application Engineer I wanna add a streaming monitor by Lambda Speed Layer Consume the same data for each usages • AWS managed services • Replicated data into Multiple AZs • High availability
- 20. Input Data
- 21. Collect Events by Fluentd • Forwarder (running on each instances) • store JSON events to S3 • forward events to aggregators • collect metrics and post them to Datadog • Aggregator • input events into Kinesis & PipelineDB • other reporting tasks (not mentioned today)
- 22. Forwards to S3 <source> @type tail format json path /data/log/user_activity.log pos_file /data/log/pos/user_activity.pos tag smartnews.user_activity time_key timestamp </source> <match smartnews.user_activity> @type copy <store> @type relabel @label @s3 </store> <store> @type forward @label @forward </store> </match> @include conf.d/s3.conf @include conf.d/forward.conf <label @s3> <% node[:td_agent][:s3].each do |c| -%> <match <%= c[:tag] %>> @id s3.<%= c[:tag] %> @type s3 ... path fluentd/<%= node[:env] %>/<%= node[:role] %>/<%= c[:tag] %> time_slice_format dt=%Y-%m-%d/hh=%H time_key timestamp include_time_key time_as_epoch reduced_redundancy true format json utc buffer_chunk_limit 2048m </match> <% end -%> </label> td-agent.conf conf.d/s3.conf
- 23. Capture DynamoDB Streams <source> type dynamodb_streams stream_arn YOUR_DDB_STREAMS_ARN pos_file /path/to/table.pos fetch_interval 1 fetch_size 100 </source> https://github.com/takus/fluent-plugin-dynamodb-streams DynamoDB DynamoDB Streams Amazon S3 AWS Lambda Fluentd
- 24. Recommended Practices • Make configuration simple as possible • fluentd can cover everything, but shouldn't • keep stateless • Use v0.12 or later • "Filter" : better performance • "Label": eliminate 'output_tag' configuration
- 25. Monitor Fluentd Status • Monitor traffic volume & retry count by Datadog • Datadog's fluentd integration • fluent-plugin-flowcounter • fluent-plugin-dogstatsd
- 26. Archive to Amazon S3 • I have 2 recommended settings • versioning • enable to recover from human error • lifecycle policy • minify storage cost Archives to IA or Gracier xx days after the creation date Keep previous versions xx days Save you in the future!!
- 27. Batch Layer
- 28. Various ETL Tasks • Extract • dump MySQL records by Embulk • make files on S3 readable to Hive • Transform • transform text files into columnar files (RCFile, ORC) • generate features for machine learning • aggregate records (by country, by channel) • Load • load aggregated metrics into Amazon Aurora
- 29. Hive • Most popular project on Hadoop ecosystem • famous for its lovely logo :) • HiveQL and MapReduce • convert SQL-like query into MR jobs • Not adopt Tez engine yet • Amazon EMR doesn't support now • limited improvement to our queries
- 30. How to process JSON? A. Transform into columnar table periodically • required converting job • better performance B. Use JSON-SerDe for temporary analysis • easy way for querying raw json text files • required to "drop table" for change schema • performance is not good
- 31. Transform Tables -- Make S3 files readable by Hive ALTER TABLE raw_activities ADD IF NOT EXISTS PARTITION (dt='${DATE}', hh='${HOUR}'); -- Transform text files into columnar files (Flatten JSON) INSERT OVERWRITE TABLE activities PARTITION (dt='${DATE}', action) SELECT user_id, timestamp, os, country, data, action FROM raw_activities LATERAL VIEW json_tuple( raw_activities.json, 'userId','timestamp','platform','country','action','data' ) a as user_id, timestamp, os, country, action, data WHERE dt = '${DATE}' CLUSTER BY os, country, action, user_id ;
- 32. JSON-SerDe -- Define table with SERDE CREATE TABLE json_table ( country string, languages array<string>, religions map<string,array<int>> ) ROW FORMAT SERDE 'org.openx.data.jsonserde.JsonSerDe' STORED AS TEXTFILE; -- Result: 10 SELECT religions['catholic'][0] FROM json_table;
- 33. cf. hive-ruby-scripting -- Define your ruby (JRuby) script SET rb.script= require 'json' def parse (json) j = JSON.load(json) j['profile']['attribute1'] end ; -- Use the script in HQL SELECT rb_exec('&parse', json) FROM user; https://github.com/gree/hive-ruby-scripting
- 35. Self-Serve via AWS CLI # Create EMR clusters that runs Hive & Spark & Ganglia aws emr create-cluster --name "My Cluster" --release-label emr-4.2.0 --applications Name=Hive Name=Spark Name=GANGLIA --ec2-attributes KeyName=myKey --instance-type c3.4xlarge --instance-count 4 --use-default-roles
- 36. Minimize expenses • Use Spot Instances as possible • typically discount 50-90% • select instance type with stable price • C3 families spike often :( • Dynamic cluster resizing • x2 capacity during daily batch job • 1/2 capacity during midnight
- 38. Typical Anti-Pattern 5 * * * * app hive -f query_1.hql 15 * * * * app hive -f query_2.hql 30 * * * * app hive -f query_3.hql 0 * * * * app hive -f query_4.hql 1 * * * * app hive -f query_5.hql
- 39. Workflow Management • Define dependencies • task E is executed after finishing task C and task D • Scheduling • task A is kicked after 09:00 AM • throttle concurrent running of the same task • Monitoring • notification in failure • task C must finish before 01:00 PM (SLA) cf. http://www.slideshare.net/taroleo/workflow-hacks-1-dots-tokyo
- 40. Airflow • A workflow management systems • define workflow by Python • built in shiny UI & CLI • pluggable architecture http://nerds.airbnb.com/airflow/
- 41. Define Tasks dag = DAG('tutorial', default_args=default_args) t1 = BashOperator( task_id='print_date', bash_command='date', dag=dag) t2 = BashOperator( task_id='sleep', bash_command='sleep 5', retries=3, dag=dag) t3 = BashOperator( task_id='templated', bash_command=""" {% for i in range(5) %} echo "{{ ds }}" echo "{{ macros.ds_add(ds, 7)}}" echo "{{ params.my_param }}" {% endfor %} """, params={'my_param': 'Parameter I passed in'}, dag=dag) t2.set_upstream(t1) t3.set_upstream(t1) Task Dependencies Python code DAG
- 42. Workflow as Code Deploy codes automatically after merging into master
- 44. What is done or not?
- 45. Alerting to Slack • SLA Violation • task A should be done till 00:00 PM • other team's task K has dependency into task A • Output validation failure • stop the following tasks if the output is doubtful
- 46. Retry from Web UI Once clear histories, airflow scheduler back fill the histories
- 47. Retry from CLI // Clear some histories from 2016-01-01 airflow clear etl_smartnews --task_regex user_ --downstream --start_date 2016-01-01 // Backfill uncompleted tasks airflow backfill etl_smartnews --start_date 2016-01-01
- 49. How Long Each Tasks?
- 50. Pluggable Architecture • Built-in plugins • operator: bash, hive, preto, mysql • transfer: hive_to_mysql • sensor: wait_hive_partition, wait_s3_file • Written our own plugin • mysql_partition
- 51. Examples user_sensor = S3KeySensor( task_id='wait_user', bucket_name='smartnews', bucket_key='user/dt={{ ds }}/dump.csv', ) etl = HiveOperator( task_id="task1", hql="INSERT OVERWRITE INTO...." ) etl.set_upstream(user_sensor) import = HiveToMySqlTransfer( task_id=name, mysql_preoperator="DELETE FROM %s WHERE date = '{{ ds }}'" % table, sql="SELECT country, count(*) FROM %s" % table, mysql_table=table ) import.set_upstream(etl) Wait a S3 file creation After the file is created, Run ETL Query After that, Import into MySQL
- 52. Serving Layer
- 53. Provides batch views in low-latency and ad-hoc way
- 54. Presto • A distributed SQL query engine • join multiple data sources (Hive + MySQL) • support standard ANSI SQL • designed to handle TBs or PBs scale data cf. http://www.slideshare.net/frsyuki/presto-hadoop-conference-japan-2014
- 55. Presto Architecture Amazon S3 Kinesis Stream Amazon RDS Amazon Aurora Presto Worker Presto Worker Presto Worker Presto Worker Presto Worker Presto Worker Presto Coordinator Client 1. Query with Standard SQL 4. Scan data concurrently 5. Aggregate data without disk I/O 6. Return result to client 2. Generate execution plan 3. Dispatch tasks into multiple workers Amazon EMR (Hive Metastore) Provides Hive table metadata (S3 access only) ※ https://github.com/qubole/presto-kinesis ※
- 56. Why Presto? • Join multiple data sources • skip large parts of ETL process • enable to merge Hive/MySQL/Kinesis/PipelineDB • Low latency • ~30s to scan billions records in S3 • Low maintenance cost • stateless, and easy to integrate with Auto Scaling
- 57. Use case: A/B Test -- Suppose that this table exists DESC hive.default.user_activities; user_id bigint action varchar abtest array<map<varchar, bigint>> url varchar -- Summarize page view per A/B Test identifier -- for comparing two algorithms v1 & v2 SELECT dt, t['behaviorId'], count(*) as pv FROM hive.default.user_activities CROSS JOIN UNNEST(abtest) AS t (t) WHERE dt like '2016-01-%' AND action = 'viewArticle' AND t['definitionId'] = 163 GROUP BY dt, t['behaviorId'] ORDER BY dt ; 2015-12-01 | algorithm_v1 | 40000 2015-12-01 | algorithm_v2 | 62000
- 58. Use case: Troubleshoot -- Store access logs to S3, and query to them -- Summarize access & 95pct response time by SQL SELECT from_unixtime(timestamp), count(*) as access, approx_percentile(reqtime, 0.95) as pct95_reqtime FROM hive.default.access_log WHERE dt = '2015-11-04' AND hh = '13' AND role = 'xxx' GROUP BY timestamp ORDER BY timestamp ; 2015-11-04 22:00:00.000 | 6377 | 0.522 2015-11-04 22:00:01.000 | 3580 | 0.422
- 59. Scheduled Auto Scaling $ aws autoscaling describe-scheduled-actions { "ScheduledUpdateGroupActions": [ { "DesiredCapacity": 2, "AutoScalingGroupName": "presto-worker-prd", "Recurrence": "59 14 * * *", "ScheduledActionName": "scalein-2359-jst" }, { "DesiredCapacity": 20, "AutoScalingGroupName": "presto-worker-prd", "Recurrence": "45 0 * * 1-5", "ScheduledActionName": "scaleout-0945-jst" } ] }
- 60. Presto Covers Everything? No! • Fixed system on Amazon Aurora (or other RDB) • provides KPI for products & business • require high availability & low latency • has no flexibility • Ad-hoc system on Presto • provides access to all dataset on data platform • require high scalability • has flexibility (join various data sources)
- 61. Why Fixed vs Ad-hoc? • Difficulties on the Ad-hoc only solution • difficult to prevent heavy queries • large distinct count exhausts computing resources • decrease presto maintainability
- 62. Output Data
- 63. Chartio • Dashboard as A Service • helps businesses analyze and track their critical data • one of AWS partners (※) • Combine multiple data sources at one dashboard • Presto, MySQL, Redshift, BigQuery, Elasticsearch ... • enable to join BigQuery + MySQL internally • Easy to use for every one • everyone can make their own dashboard • write SQL directly / generate query by drag & drop ※ http://www.aws-partner-directory.com/PartnerDirectory/PartnerDetail?id=8959
- 64. Creating dashboard 1. Building query (Drag&Drop / SQL) 2. Add step (filter、sort、modify) 3. Select visualize way (table、graph)
- 65. Examples
- 66. Why Chartio? • Chartio saves a lot of engineering resources • before • maintain in-house dashboard written by rails • everyone got tired to maintain it • after • everyone can build their own dashboard easily • Chartio's UI is cool • very important factor for dashboard tool
- 67. Missing Pieces of Chartio • No programable API provides • need to edit dashboard / chart manually • No rollback feature • all changes are recorded, but not rollback to the previous state • work around : clone => edit => rename
- 68. Speed Layer
- 69. Why Speed is Matter?
- 70. Today’s News is Wrapping Tomorrow’s Fish and Chips
- 73. Use cases • Re-rank news articles by user feedback • track user's positive/negative signal • consider gender, age, location, interests • Realtime article monitoring • detect high bounce rate (may be broken?) • make realtime reporting dashboard for A/B test
- 74. Realtime Re-Ranking ref. Stream 処理 (Spark Streaming + Kinesis) と Offline 処理 (Hive) の統合 www.slideshare.net/smartnews/stremspark-streaming-kinesisofflinehive Amazon CloudSearch Search API API Gateway Kinesis Stream Amazon S3 Amazon EMR Amazon S3 Amazon EMR DynamoDB Realtime Feedback Re-rank Articles Article Metadata User Interests User Behaviors Offline Procees by Hive / Spark
- 75. Realtime Monitoring API Gateway Stream Continuous View Continuous View Continuous View Discard raw record soon after consumed by Continuous View Incrementally updated in realtime PipelineDB Chartio AWS Lambda Slack Access Continuous View by PostgreSQL Client Record ※1 ※1 Use cron on 26 Feb. 2016 Migrate it soon after supporting VPC
- 76. PipelineDB • OSS & enterprise streaming SQL database • PostgreSQL compatible • connect to Chartio 😍 • join stream to normal PostgreSQL table • Support probabilistic data structures • e.g. HyperLogLog https://www.pipelinedb.com/ http://developer.smartnews.com/blog/2015/09/09/20150907pipelinedb/
- 77. Continuous View -- Calculate unique users seen per media each day -- Using only a constant amount of space (HyperLogLog) CREATE CONTINUOUS VIEW uniques AS SELECT day(arrival_timestamp), substring(url from '.*://([^/]*)') as hostname, COUNT(DISTINCT user_id::integer) FROM activity_stream GROUP BY day,hostname; -- How many impressions have we served in the last five minutes? CREATE CONTINUOUS VIEW imps WITH (max_age = '5 minutes') AS SELECT COUNT(*) FROM imps_stream; -- What are the 90th, 95th, 99th percentiles of request latency? CREATE CONTINUOUS VIEW latency AS SELECT percentile_cont(array[90, 95, 99]) WITHIN GROUP (ORDER BY latency::integer) FROM latency_stream;
- 78. Summary
- 79. Sustainable Data Platform • build a reliable and scalable lambda architecture • minimize operation & running cost • be open to uncertain future
- 80. My Wishlist to AWS • Support Reduced Redundancy Storage (RRS) on EMR • Faster EMR Launch • Set TTL to DynamoDB records • Auto-scale Kinesis Stream • Launch Kinesis Analytics in Tokyo region
- 81. Thank you!!