SlideShare a Scribd company logo
1 of 32
Download to read offline
Our Environment Challenges and Solutions Conclusions
PostgreSQL at 20TB and Beyond
Analytics at a Massive Scale
Chris Travers
Adjust GmbH
January 25, 2018
Our Environment Challenges and Solutions Conclusions
About Adjust
Adjust is a market leader in mobile
advertisement attribution. We basically
act as a referee in pay-per-install
advertising. We focus on fairness,
fraud-prevention and other ways of
ensuring that advertisers pay for the
services they receiving fairly and with
accountability.
Our Environment Challenges and Solutions Conclusions
Our Environment
About Adjust
Traffic Statistics
Analytics Environment
Challenges and Solutions
Staffing
Throughput
Autovacuum
Data Modeling
Backups and Operations
Conclusions
Our Environment Challenges and Solutions Conclusions
Basic Facts About Us
• We are a PostgreSQL/Kafka shop
• Around 200 employees worldwide
• Link advertisements to installs
• Delivering near-real-time analytics to software vendors
• Our Data is “Big Data”
Our Environment Challenges and Solutions Conclusions
Just How Big?
• Over 100k requests per second
• Over 2 trillion data points tracked in 2017
• Over 400 TB of data to analyze
• Very high data velocity
Our Environment Challenges and Solutions Conclusions
General Architecture
• Requests come from Internet
• Written to backends
• Materialized to analytics
shards
• Shown in dashboard
Our Environment Challenges and Solutions Conclusions
Common Elements of Infrastructure
• Bare metal
• Stripped down Gentoo
• Lots of A/B performance testing
• Approx 50% more throughput than stock Linux systems
• Standard PostgreSQL + extensions
Our Environment Challenges and Solutions Conclusions
Backend Servers
• Original point of entry
• Data distributed by load balancer
• No data-dependent routing.
• Data distributed more or less randomly
• Around 20TB per backend server gets stored.
• More than 20 backend servers
Our Environment Challenges and Solutions Conclusions
Materializer
• Aggregates new events
• Copies the aggregations to the shards
• Runs every few minutes
• New data only
Our Environment Challenges and Solutions Conclusions
Materializer and MapReduce
Our materializer aggregates data from many servers and transfers
it to many servers. It functions sort of like a mapreduce with the
added complication that it is a many server to many server
transformation.
Our Environment Challenges and Solutions Conclusions
Analytics Shards
• Each about 2TB each
• 16 shards currently, may grow
• Our own custom analytics software for managing and querying
• Custom sharding/locating software
• Paired for redundancy
Our Environment Challenges and Solutions Conclusions
Staffing: Challenges
• No Junior Database Folks
• Demanding environment
• Very little room for error
• Need people who are deeply grounded in both theory and
practice
Our Environment Challenges and Solutions Conclusions
Staffing: Solutions
• Look for people with enough relevant experience they can
learn
• Be picky with what we are willing to teach new folks
• Look for self-learners with enough knowledge to participate
• Expect people to grow into the role
• We also use code challenges
Our Environment Challenges and Solutions Conclusions
Throughput challenges
• Most data is new
• It is a lot of data
• Lots of btrees with lots of random inserts
• Ideally, every wrote inserted once and updated once
• Long-term retention
Our Environment Challenges and Solutions Conclusions
Throughput Solutions
Backend Servers
• Each point of entry server has its own database
• Transactional processing is separate.
• Careful attention to alignment issues
• We write our own data types in C to help
• Tables partitioned by event time
Our Environment Challenges and Solutions Conclusions
Throughput Solutions
Analytics Shards
• Pre-aggregated data for client-facing metrics
• Sharded at roughly 2TB per shard
• 16 shards currently
• Custom sharding framework optimized to reduce network
usage
• Goal is to have dashboards load fast.
• We know where data is on these shards.
Our Environment Challenges and Solutions Conclusions
Throughput Solutions
Materializer
• Two phases
• First phase runs on original entry servers
• Aggregates and copies data to analytics shards
• Second phase runs on analytics shards
• further aggregates and copies.
Our Environment Challenges and Solutions Conclusions
Materializer: A Special Problem
• Works great when you only have one data center
• Foreign data wrapper bulk writes are very slow across data
centers
• This is a known issue with the Postgres FDW
• This is a blocking issue.
Our Environment Challenges and Solutions Conclusions
Materializer: Solution
• C extension using COPY
• Acts as libpq client
• Wrote a global transaction manager
• Throughput restored.
Our Environment Challenges and Solutions Conclusions
Introducing Autovacuum
• Queries have to provide consistent snapshots
• All updates in PostgreSQL are copy-on-write
• In our case, we write once and then update once.
• Have to clean up old data at some point
• By default, 50 rows plus 20% of table being “dead” triggers
autovacuum
Our Environment Challenges and Solutions Conclusions
Autovacuum problems
• For small tables, great but we have tables with 200M rows
• 20% of 200M rows is 40 million dead tuples....
• Autovacuum does nothing and then undertakes a heavy
task....
• performance suffers and tables bloat.
Our Environment Challenges and Solutions Conclusions
Autovacuum Solutions
• Change to 150k rows plus 0%
• Tuning requires a lot of hand-holding
• Roll out change to servers gradually to avoid overloading
system.
Our Environment Challenges and Solutions Conclusions
Why it Matters
• Under heavy load, painful to change
• Want to avoid rewriting tables
• Want to minimize disk usage
• Want to maximize alignment to pages
• Lots of little details really matter
Our Environment Challenges and Solutions Conclusions
Custom 1-byte Enums
• Country
• Language
• OS Name
• Device Type
Our Environment Challenges and Solutions Conclusions
IStore
Like HStore but for Integers
• Like HStore but for integers
• Supports addition, etc, between values of same key
• Useful for time series and other modelling problems
• Supports GIN indexing among others
Our Environment Challenges and Solutions Conclusions
The Elephant in the Room
How do we aggregate that much data?
• Basically incremental Map Reduce
• Map and first phase aggregation on backends
• Reduce and second phase aggregation on shards
• Further reduction and aggregation possible on demand
Our Environment Challenges and Solutions Conclusions
Operations Tools
• Sqitch
• Rex
• Our own custom tools
Our Environment Challenges and Solutions Conclusions
Backups
• Home grown system
• Base backup plus WAL
• Runs as a Rex task
• We can also do logical backups (but...)
Our Environment Challenges and Solutions Conclusions
Ongoing Distributed Challenges
• Major Upgrades
• Storage Space
• Multi-datacenter challenges
• Making it all fast
Our Environment Challenges and Solutions Conclusions
Overview
This environment is all about careful attention to detail and being
willing to write C code when needed. Space savings, better
alignment, and other seemingly small gains add up over tens of
billions of rows.
Our Environment Challenges and Solutions Conclusions
Major Points of Interest
• We are using PostgreSQL as a big data platform.
• We expect this architecture to scale very far.
• Provides near-real-time analytics on user actions.
Our Environment Challenges and Solutions Conclusions
PostgreSQL makes all this Possible
In buiding our 400TB analytics environment we have yet to
outgrow PostgreSQL. In fact, this is one of the few pieces of our
infrastructure we are perfectly confident in scaling.

More Related Content

What's hot

What's hot (20)

MariaDB Galera Cluster presentation
MariaDB Galera Cluster presentationMariaDB Galera Cluster presentation
MariaDB Galera Cluster presentation
 
Running MariaDB in multiple data centers
Running MariaDB in multiple data centersRunning MariaDB in multiple data centers
Running MariaDB in multiple data centers
 
The Full MySQL and MariaDB Parallel Replication Tutorial
The Full MySQL and MariaDB Parallel Replication TutorialThe Full MySQL and MariaDB Parallel Replication Tutorial
The Full MySQL and MariaDB Parallel Replication Tutorial
 
Maximizing performance via tuning and optimization
Maximizing performance via tuning and optimizationMaximizing performance via tuning and optimization
Maximizing performance via tuning and optimization
 
MariaDB Server Performance Tuning & Optimization
MariaDB Server Performance Tuning & OptimizationMariaDB Server Performance Tuning & Optimization
MariaDB Server Performance Tuning & Optimization
 
Linux performance tuning & stabilization tips (mysqlconf2010)
Linux performance tuning & stabilization tips (mysqlconf2010)Linux performance tuning & stabilization tips (mysqlconf2010)
Linux performance tuning & stabilization tips (mysqlconf2010)
 
MySQL Performance Tuning. Part 1: MySQL Configuration (includes MySQL 5.7)
MySQL Performance Tuning. Part 1: MySQL Configuration (includes MySQL 5.7)MySQL Performance Tuning. Part 1: MySQL Configuration (includes MySQL 5.7)
MySQL Performance Tuning. Part 1: MySQL Configuration (includes MySQL 5.7)
 
MySQL Administrator 2021 - 네오클로바
MySQL Administrator 2021 - 네오클로바MySQL Administrator 2021 - 네오클로바
MySQL Administrator 2021 - 네오클로바
 
Introduction to Galera Cluster
Introduction to Galera ClusterIntroduction to Galera Cluster
Introduction to Galera Cluster
 
Migrating from InnoDB and HBase to MyRocks at Facebook
Migrating from InnoDB and HBase to MyRocks at FacebookMigrating from InnoDB and HBase to MyRocks at Facebook
Migrating from InnoDB and HBase to MyRocks at Facebook
 
Optimizing RocksDB for Open-Channel SSDs
Optimizing RocksDB for Open-Channel SSDsOptimizing RocksDB for Open-Channel SSDs
Optimizing RocksDB for Open-Channel SSDs
 
SSD Deployment Strategies for MySQL
SSD Deployment Strategies for MySQLSSD Deployment Strategies for MySQL
SSD Deployment Strategies for MySQL
 
Optimizing MariaDB for maximum performance
Optimizing MariaDB for maximum performanceOptimizing MariaDB for maximum performance
Optimizing MariaDB for maximum performance
 
RocksDB compaction
RocksDB compactionRocksDB compaction
RocksDB compaction
 
Mastering PostgreSQL Administration
Mastering PostgreSQL AdministrationMastering PostgreSQL Administration
Mastering PostgreSQL Administration
 
Best practices for MySQL High Availability Tutorial
Best practices for MySQL High Availability TutorialBest practices for MySQL High Availability Tutorial
Best practices for MySQL High Availability Tutorial
 
MariaDB Optimization
MariaDB OptimizationMariaDB Optimization
MariaDB Optimization
 
Oracle RAC 19c: Best Practices and Secret Internals
Oracle RAC 19c: Best Practices and Secret InternalsOracle RAC 19c: Best Practices and Secret Internals
Oracle RAC 19c: Best Practices and Secret Internals
 
MySQL 상태 메시지 분석 및 활용
MySQL 상태 메시지 분석 및 활용MySQL 상태 메시지 분석 및 활용
MySQL 상태 메시지 분석 및 활용
 
Percona XtraDB Cluster
Percona XtraDB ClusterPercona XtraDB Cluster
Percona XtraDB Cluster
 

Similar to PostgreSQL at 20TB and Beyond

The challenges of live events scalability
The challenges of live events scalabilityThe challenges of live events scalability
The challenges of live events scalability
Guy Tomer
 

Similar to PostgreSQL at 20TB and Beyond (20)

PostgreSQL as a Big Data Platform
PostgreSQL as a Big Data Platform PostgreSQL as a Big Data Platform
PostgreSQL as a Big Data Platform
 
No stress with state
No stress with stateNo stress with state
No stress with state
 
Why retail companies can't afford database downtime
Why retail companies can't afford database downtimeWhy retail companies can't afford database downtime
Why retail companies can't afford database downtime
 
Large Scale Architecture -- The Unreasonable Effectiveness of Simplicity
Large Scale Architecture -- The Unreasonable Effectiveness of SimplicityLarge Scale Architecture -- The Unreasonable Effectiveness of Simplicity
Large Scale Architecture -- The Unreasonable Effectiveness of Simplicity
 
MongoDB: What, why, when
MongoDB: What, why, whenMongoDB: What, why, when
MongoDB: What, why, when
 
DNN-Connect 2019: DNN Horror Stories
DNN-Connect 2019: DNN Horror StoriesDNN-Connect 2019: DNN Horror Stories
DNN-Connect 2019: DNN Horror Stories
 
PostgreSQL as a Strategic Tool
PostgreSQL as a Strategic ToolPostgreSQL as a Strategic Tool
PostgreSQL as a Strategic Tool
 
A data analyst view of Bigdata
A data analyst view of Bigdata A data analyst view of Bigdata
A data analyst view of Bigdata
 
Scaling apps for the big time
Scaling apps for the big timeScaling apps for the big time
Scaling apps for the big time
 
Lessons Learned Replatforming A Large Machine Learning Application To Apache ...
Lessons Learned Replatforming A Large Machine Learning Application To Apache ...Lessons Learned Replatforming A Large Machine Learning Application To Apache ...
Lessons Learned Replatforming A Large Machine Learning Application To Apache ...
 
CD presentation march 12th, 2018
CD presentation march 12th, 2018CD presentation march 12th, 2018
CD presentation march 12th, 2018
 
Voldemort Nosql
Voldemort NosqlVoldemort Nosql
Voldemort Nosql
 
Mixing d ps building architecture on the cross cutting example
Mixing d ps building architecture on the cross cutting exampleMixing d ps building architecture on the cross cutting example
Mixing d ps building architecture on the cross cutting example
 
Intro to Big Data
Intro to Big DataIntro to Big Data
Intro to Big Data
 
Big Data and Hadoop
Big Data and HadoopBig Data and Hadoop
Big Data and Hadoop
 
Scaling Systems: Architectures that grow
Scaling Systems: Architectures that growScaling Systems: Architectures that grow
Scaling Systems: Architectures that grow
 
In (database) automation we trust
In (database) automation we trustIn (database) automation we trust
In (database) automation we trust
 
The challenges of live events scalability
The challenges of live events scalabilityThe challenges of live events scalability
The challenges of live events scalability
 
AOEcon17: Searchperience - The journey from PHP and Solr to Scala and Elastic...
AOEcon17: Searchperience - The journey from PHP and Solr to Scala and Elastic...AOEcon17: Searchperience - The journey from PHP and Solr to Scala and Elastic...
AOEcon17: Searchperience - The journey from PHP and Solr to Scala and Elastic...
 
DBmaestro's State of the Database Continuous Delivery Survey- Findings Revealed
DBmaestro's State of the Database Continuous Delivery Survey- Findings RevealedDBmaestro's State of the Database Continuous Delivery Survey- Findings Revealed
DBmaestro's State of the Database Continuous Delivery Survey- Findings Revealed
 

Recently uploaded

Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night StandCall Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
amitlee9823
 
notes on Evolution Of Analytic Scalability.ppt
notes on Evolution Of Analytic Scalability.pptnotes on Evolution Of Analytic Scalability.ppt
notes on Evolution Of Analytic Scalability.ppt
MsecMca
 
AKTU Computer Networks notes --- Unit 3.pdf
AKTU Computer Networks notes ---  Unit 3.pdfAKTU Computer Networks notes ---  Unit 3.pdf
AKTU Computer Networks notes --- Unit 3.pdf
ankushspencer015
 
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Christo Ananth
 

Recently uploaded (20)

Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night StandCall Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
 
Water Industry Process Automation & Control Monthly - April 2024
Water Industry Process Automation & Control Monthly - April 2024Water Industry Process Automation & Control Monthly - April 2024
Water Industry Process Automation & Control Monthly - April 2024
 
(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7
(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7
(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7
 
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
 
Generative AI or GenAI technology based PPT
Generative AI or GenAI technology based PPTGenerative AI or GenAI technology based PPT
Generative AI or GenAI technology based PPT
 
UNIT - IV - Air Compressors and its Performance
UNIT - IV - Air Compressors and its PerformanceUNIT - IV - Air Compressors and its Performance
UNIT - IV - Air Compressors and its Performance
 
Double Revolving field theory-how the rotor develops torque
Double Revolving field theory-how the rotor develops torqueDouble Revolving field theory-how the rotor develops torque
Double Revolving field theory-how the rotor develops torque
 
Online banking management system project.pdf
Online banking management system project.pdfOnline banking management system project.pdf
Online banking management system project.pdf
 
Thermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.pptThermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.ppt
 
chapter 5.pptx: drainage and irrigation engineering
chapter 5.pptx: drainage and irrigation engineeringchapter 5.pptx: drainage and irrigation engineering
chapter 5.pptx: drainage and irrigation engineering
 
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance BookingCall Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
 
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank  Design by Working Stress - IS Method.pdfIntze Overhead Water Tank  Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
 
notes on Evolution Of Analytic Scalability.ppt
notes on Evolution Of Analytic Scalability.pptnotes on Evolution Of Analytic Scalability.ppt
notes on Evolution Of Analytic Scalability.ppt
 
AKTU Computer Networks notes --- Unit 3.pdf
AKTU Computer Networks notes ---  Unit 3.pdfAKTU Computer Networks notes ---  Unit 3.pdf
AKTU Computer Networks notes --- Unit 3.pdf
 
Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)
 
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
 
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
 
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance BookingCall Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
 
(INDIRA) Call Girl Aurangabad Call Now 8617697112 Aurangabad Escorts 24x7
(INDIRA) Call Girl Aurangabad Call Now 8617697112 Aurangabad Escorts 24x7(INDIRA) Call Girl Aurangabad Call Now 8617697112 Aurangabad Escorts 24x7
(INDIRA) Call Girl Aurangabad Call Now 8617697112 Aurangabad Escorts 24x7
 
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptxBSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
 

PostgreSQL at 20TB and Beyond

  • 1. Our Environment Challenges and Solutions Conclusions PostgreSQL at 20TB and Beyond Analytics at a Massive Scale Chris Travers Adjust GmbH January 25, 2018
  • 2. Our Environment Challenges and Solutions Conclusions About Adjust Adjust is a market leader in mobile advertisement attribution. We basically act as a referee in pay-per-install advertising. We focus on fairness, fraud-prevention and other ways of ensuring that advertisers pay for the services they receiving fairly and with accountability.
  • 3. Our Environment Challenges and Solutions Conclusions Our Environment About Adjust Traffic Statistics Analytics Environment Challenges and Solutions Staffing Throughput Autovacuum Data Modeling Backups and Operations Conclusions
  • 4. Our Environment Challenges and Solutions Conclusions Basic Facts About Us • We are a PostgreSQL/Kafka shop • Around 200 employees worldwide • Link advertisements to installs • Delivering near-real-time analytics to software vendors • Our Data is “Big Data”
  • 5. Our Environment Challenges and Solutions Conclusions Just How Big? • Over 100k requests per second • Over 2 trillion data points tracked in 2017 • Over 400 TB of data to analyze • Very high data velocity
  • 6. Our Environment Challenges and Solutions Conclusions General Architecture • Requests come from Internet • Written to backends • Materialized to analytics shards • Shown in dashboard
  • 7. Our Environment Challenges and Solutions Conclusions Common Elements of Infrastructure • Bare metal • Stripped down Gentoo • Lots of A/B performance testing • Approx 50% more throughput than stock Linux systems • Standard PostgreSQL + extensions
  • 8. Our Environment Challenges and Solutions Conclusions Backend Servers • Original point of entry • Data distributed by load balancer • No data-dependent routing. • Data distributed more or less randomly • Around 20TB per backend server gets stored. • More than 20 backend servers
  • 9. Our Environment Challenges and Solutions Conclusions Materializer • Aggregates new events • Copies the aggregations to the shards • Runs every few minutes • New data only
  • 10. Our Environment Challenges and Solutions Conclusions Materializer and MapReduce Our materializer aggregates data from many servers and transfers it to many servers. It functions sort of like a mapreduce with the added complication that it is a many server to many server transformation.
  • 11. Our Environment Challenges and Solutions Conclusions Analytics Shards • Each about 2TB each • 16 shards currently, may grow • Our own custom analytics software for managing and querying • Custom sharding/locating software • Paired for redundancy
  • 12. Our Environment Challenges and Solutions Conclusions Staffing: Challenges • No Junior Database Folks • Demanding environment • Very little room for error • Need people who are deeply grounded in both theory and practice
  • 13. Our Environment Challenges and Solutions Conclusions Staffing: Solutions • Look for people with enough relevant experience they can learn • Be picky with what we are willing to teach new folks • Look for self-learners with enough knowledge to participate • Expect people to grow into the role • We also use code challenges
  • 14. Our Environment Challenges and Solutions Conclusions Throughput challenges • Most data is new • It is a lot of data • Lots of btrees with lots of random inserts • Ideally, every wrote inserted once and updated once • Long-term retention
  • 15. Our Environment Challenges and Solutions Conclusions Throughput Solutions Backend Servers • Each point of entry server has its own database • Transactional processing is separate. • Careful attention to alignment issues • We write our own data types in C to help • Tables partitioned by event time
  • 16. Our Environment Challenges and Solutions Conclusions Throughput Solutions Analytics Shards • Pre-aggregated data for client-facing metrics • Sharded at roughly 2TB per shard • 16 shards currently • Custom sharding framework optimized to reduce network usage • Goal is to have dashboards load fast. • We know where data is on these shards.
  • 17. Our Environment Challenges and Solutions Conclusions Throughput Solutions Materializer • Two phases • First phase runs on original entry servers • Aggregates and copies data to analytics shards • Second phase runs on analytics shards • further aggregates and copies.
  • 18. Our Environment Challenges and Solutions Conclusions Materializer: A Special Problem • Works great when you only have one data center • Foreign data wrapper bulk writes are very slow across data centers • This is a known issue with the Postgres FDW • This is a blocking issue.
  • 19. Our Environment Challenges and Solutions Conclusions Materializer: Solution • C extension using COPY • Acts as libpq client • Wrote a global transaction manager • Throughput restored.
  • 20. Our Environment Challenges and Solutions Conclusions Introducing Autovacuum • Queries have to provide consistent snapshots • All updates in PostgreSQL are copy-on-write • In our case, we write once and then update once. • Have to clean up old data at some point • By default, 50 rows plus 20% of table being “dead” triggers autovacuum
  • 21. Our Environment Challenges and Solutions Conclusions Autovacuum problems • For small tables, great but we have tables with 200M rows • 20% of 200M rows is 40 million dead tuples.... • Autovacuum does nothing and then undertakes a heavy task.... • performance suffers and tables bloat.
  • 22. Our Environment Challenges and Solutions Conclusions Autovacuum Solutions • Change to 150k rows plus 0% • Tuning requires a lot of hand-holding • Roll out change to servers gradually to avoid overloading system.
  • 23. Our Environment Challenges and Solutions Conclusions Why it Matters • Under heavy load, painful to change • Want to avoid rewriting tables • Want to minimize disk usage • Want to maximize alignment to pages • Lots of little details really matter
  • 24. Our Environment Challenges and Solutions Conclusions Custom 1-byte Enums • Country • Language • OS Name • Device Type
  • 25. Our Environment Challenges and Solutions Conclusions IStore Like HStore but for Integers • Like HStore but for integers • Supports addition, etc, between values of same key • Useful for time series and other modelling problems • Supports GIN indexing among others
  • 26. Our Environment Challenges and Solutions Conclusions The Elephant in the Room How do we aggregate that much data? • Basically incremental Map Reduce • Map and first phase aggregation on backends • Reduce and second phase aggregation on shards • Further reduction and aggregation possible on demand
  • 27. Our Environment Challenges and Solutions Conclusions Operations Tools • Sqitch • Rex • Our own custom tools
  • 28. Our Environment Challenges and Solutions Conclusions Backups • Home grown system • Base backup plus WAL • Runs as a Rex task • We can also do logical backups (but...)
  • 29. Our Environment Challenges and Solutions Conclusions Ongoing Distributed Challenges • Major Upgrades • Storage Space • Multi-datacenter challenges • Making it all fast
  • 30. Our Environment Challenges and Solutions Conclusions Overview This environment is all about careful attention to detail and being willing to write C code when needed. Space savings, better alignment, and other seemingly small gains add up over tens of billions of rows.
  • 31. Our Environment Challenges and Solutions Conclusions Major Points of Interest • We are using PostgreSQL as a big data platform. • We expect this architecture to scale very far. • Provides near-real-time analytics on user actions.
  • 32. Our Environment Challenges and Solutions Conclusions PostgreSQL makes all this Possible In buiding our 400TB analytics environment we have yet to outgrow PostgreSQL. In fact, this is one of the few pieces of our infrastructure we are perfectly confident in scaling.