Years ago when working at Amazon on shopping cart infrastructure and the precursor to DynamoDB, my co-founder and I realized that while distributed key value stores were useful for a few use-cases, we missed many of the benefits of relational databases: transactions, joins, and the power of the lingua franca of RDBMS’s: SQL. So we challenged ourselves to modernize the traditional relational database, to take a robust open source relational database and transform it into a distributed database. This talk is about my team’s journey to create a more modern relational database. I’ll talk about the distributed systems problems we had to solve in order to scale out the Postgres open source database, in order to achieve parallelism and a concomitant increase in performance. I'll describe the architecture of the distributed query planner; how we extend traditional relational algebra operators to plan distributed queries and scale reads. I’ll also describe distributed deadlock detection, and how that enabled us to scale out transactions spanning multiple machines.

1. Sumedh Pathak, Co-Founder & VP Engineering, Citus Data
QCon London 2018
The Future of Distributed
Databases is Relational
@moss_toss | @citusdata

2. About Me
• Co-Founder & VP
Engineering at Citus Data
• Amazon Shopping Cart
(former)
• Amazon Supply Chain &
Order Fulfillment (former)
• Stanford Computer Science
Citus Data co-founders, left-to-right
Ozgun Erdogan, Sumedh Pathak, & Umur Cubukcu
Photo cred: Willy Johnson, Monterey CA, Sep 2017
Sumedh Pathak | Citus Data | QCon London 2018

4. Why
RDBMS?
Sumedh Pathak | Citus Data | QCon London 2018

5. Streaming Storage
Map
Reduce
NoSQL
SQL
Database
SELECT ...
Application
Application
Sumedh Pathak | Citus Data | QCon London 2018
Because your architecture could be simpler

6. An RDBMS is a
general-purpose
data platform
Sumedh Pathak | Citus Data | QCon London 2018
Fast writes
Real-time & bulk
High throughput
High concurrency
Data consistency
Query optimizers

7. PostgreSQL
MySQL
MongoDB
SQL Server +
Oracle
Source: Hacker News, https://news.ycombinator.com
Startups Are Choosing Postgres% database job posts mentioning each database, across 20K+ job posts
Sumedh Pathak | Citus Data | QCon London 2018

8. but RDBMS’s
don’t scale, right?
Sumedh Pathak | Citus Data | QCon London 2018

9. but RDBMS’s don’t scale
RDBMS’s are hard to scale
Sumedh Pathak | Citus Data | QCon London 2018

10. What exactly needs to Scale?
- Tables (Data)
- Partitioning, Co-location, Reference Tables
- SQL (Reads)
- How do we express and optimize distributed SQL
- Transactions (Writes)
- Cross Shard updates/deletes, Global Atomic Transactions
Sumedh Pathak | Citus Data | QCon London 2018
1
2
3

11. Scaling
Tables
Sumedh Pathak | Citus Data | QCon London 2018

12. How to partition the data?
- Pick a column
- Date
- Id (customer_id, cart_id)
- Pick a method
- Hash
- Range

13. Partition data across nodes
R1
R2
R3
R4
R5
R6
R7
Coordinator
Node
Worker
Nodes
Shards
Sumedh Pathak | Citus Data | QCon London 2018

14. Worker → RDBMS, Shard → Table

15. Reference Tables
N1
N1
N1
N1
Copies of same table
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator
Node
Worker
Nodes

16. Co-Location
R1
R2
R3
R4
S1
S2
S3
S4
Explicit Co-Location API.
E.g. Partition by Tenant
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator
Node
Worker
Nodes

17. What about Foreign Keys?

18. The key to scaling tables...
- Use relational databases as a building block
- Understand semantics of application—to be smart about
partitioning
- Multi-tenant applications

19. Scaling
SQL

20. FROM table R
SELECT x Projectx
(R) → R’
WHERE f(x) Filterf(x)
(R) → R’
… JOIN … R × S → R’
SQL ↔ Relational Algebra
Sumedh Pathak | Citus Data | QCon London 2018

21. FROM sharded_table Collect(R1
,R2
,...) → R
Distributed Relational Algebra
R1
R2
C
Sumedh Pathak | Citus Data | QCon London 2018

22. Projectx
(Collect(R1
,R2
,...)) = Collect(Projectx
(R1
), Projectx
(R2
)...)
Commutative property
R1
R2
C
R1
R2
C
Px
Px
Px
Sumedh Pathak | Citus Data | QCon London 2018

23. Collect(R1
,R2
,...) x Collect(S1
,S2
,...) = Collect(R1
× S1
,R2
× S2,
...)
Distributive property
R1
R2
C
×
C
S1
S2
R1
R2
C
×
S1
S2
×
X = Join Operator

24. SUM(x)(Collect(R1
,R2
,...)) = SUM(Collect(SUM(R1
), SUM(R2
)...))
Associative property
R1
R2
C
R1
R2
C
Sumx
Sumx
Sumx
Sumedh Pathak | Citus Data | QCon London 2018
Sumx

25. SELECT sum(price) FROM orders, nation
WHERE orders.nation = nation.name AND
orders.date >= '2012-01-01' AND
nation.region = 'Asia';
Sumedh Pathak | Citus Data | QCon London 2018

26. Sumedh Pathak | Citus Data | QCon London 2018

27. Sumedh Pathak | Citus Data | QCon London 2018
Volcano style processing
Data flows
from
bottom
to top

28. Sumedh Pathak | Citus Data | QCon London 2018

29. Sumedh Pathak | Citus Data | QCon London 2018

30. Sumedh Pathak | Citus Data | QCon London 2018
Parallelize
Aggregate
Push Joins & Filters
below collect. Run in
parallel across all
nodes
Filters & Projections
done before Join

31. SELECT sum(intermediate_col)
FROM <concatenated results>;
SELECT sum(price)
FROM orders_2 JOIN nation_2
ON (orders_2.name = nation_2.name)
WHERE
orders_2.date >= '2017-01-01'
AND
nation_2.region = 'Asia';
SELECT sum(price)
FROM orders_2 JOIN nation_1
ON (orders_2.name = nation_1.name)
WHERE
orders_2.date >= '2017-01-01'
AND
nation_2.region = 'Asia';
Sumedh Pathak | Citus Data | QCon London 2018

32. Executing Distributed SQL
SQL database
orders_1
nation_1
orders
nation
SELECT sum(price)
FROM orders_2 o JOIN nation_1
ON (o.name = n.name)
WHERE o.date >= '2017-01-01'
AND n.region = 'Asia';
SELECT sum(price)
FROM orders_1 o JOIN nation_1
ON (o.name = n.name)
WHERE o.date >= '2017-01-01'
AND n.region = 'Asia';
SELECT sum(price) FROM <results>;
SQL database
orders_2
nation_1

33. The key to scaling SQL...
- New relational algebra operators for
distributed processing
- Relational Algebra Properties to optimize
tree: Commutativity, Associativity, &
Distributivity
- Map / Reduce operators

34. Scaling
Transactions

35. Money Transfer, as an example
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
Sumedh Pathak | Citus Data | QCon London 2018

36. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
Coordinator
Sumedh Pathak | Citus Data | QCon London 2018

37. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

38. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
COMMIT;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

39. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
Coordinator
BEGIN;
UPDATE accounts SET balance =
balance + 100 WHERE id = ‘BOB’;
Sumedh Pathak | Citus Data | QCon London 2018
BEGIN;
UPDATE accounts SET balance = balance
- 100 WHERE id = ‘ALICE’;

40. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
BEGIN;
UPDATE accounts SET balance =
balance + 100 WHERE id = ‘BOB’;
COMMIT
Sumedh Pathak | Citus Data | QCon London 2018
BEGIN;
UPDATE accounts SET balance = balance
- 100 WHERE id = ‘ALICE’;
Coordinator

41. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
PREPARE TRANSACTION ‘citus_...98’; PREPARE TRANSACTION ‘citus_...98’;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

42. What happens during PREPARE?
State of transaction stored
on a durable store
Locks are
maintained
&

43. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
Coordinator
PREPARE TRANSACTION ‘citus_...98’; ROLLBACK TRANSACTION ‘citus_...98’;
Sumedh Pathak | Citus Data | QCon London 2018

44. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
PREPARE TRANSACTION ‘citus_...98’; PREPARE TRANSACTION ‘citus_...98’;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

45. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
COMMIT PREPARED ‘citus_...98’; COMMIT PREPARED ‘citus_...98’;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

46. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
COMMIT PREPARED ‘citus_...98’; COMMIT PREPARED ‘citus_...98’;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

47. A1
A2
BEGIN;
UPDATE accounts SET balance = balance -
100 WHERE id = ‘ALICE’;
UPDATE accounts SET balance = balance +
100 WHERE id = ‘BOB’;
COMMIT;
A3
A4
COMMIT PREPARED ‘citus_...98’; COMMIT PREPARED ‘citus_...98’;
Sumedh Pathak | Citus Data | QCon London 2018
Coordinator

48. But....
Deadlocks!

49. Example
// SESSION 1
BEGIN;
UPDATE accounts SET balance
= balance - 100 WHERE id =
‘ALICE’;
(LOCK on ROW with id
‘ALICE’)
// SESSION 2
Sumedh Pathak | Citus Data | QCon London 2018

50. Example
// SESSION 1
BEGIN;
UPDATE accounts SET balance = balance - 100 WHERE
id = ‘ALICE’;
(LOCK on ROW with id ‘ALICE’)
// SESSION 2
BEGIN;
UPDATE accounts SET balance
= balance + 100 WHERE id =
‘BOB’;
(LOCK on ROW with id ‘BOB’)
Sumedh Pathak | Citus Data | QCon London 2018

51. Example
// SESSION 1
BEGIN;
UPDATE accounts SET balance = balance - 100 WHERE
id = ‘ALICE’;
(LOCK on ROW with id ‘ALICE’)
UPDATE accounts SET balance
= balance + 100 WHERE id =
‘BOB’;
// SESSION 2
BEGIN;
UPDATE accounts SET balance = balance + 100 WHERE
id = ‘BOB’;
(LOCK on ROW with id ‘BOB’)
Sumedh Pathak | Citus Data | QCon London 2018

52. Example
// SESSION 1
BEGIN;
UPDATE accounts SET balance = balance - 100 WHERE
id = ‘ALICE’;
(LOCK on ROW with id ‘ALICE’)
UPDATE accounts SET balance = balance + 100
WHERE id = ‘BOB’;
// SESSION 2
BEGIN;
UPDATE accounts SET balance = balance + 100 WHERE
id = ‘BOB’;
(LOCK on ROW with id ‘BOB’)
UPDATE accounts SET balance
= balance - 100 WHERE id =
‘ALICE’
Sumedh Pathak | Citus Data | QCon London 2018

53. How do Relational DB’s solve this?
S1
S2
- Construct a
Directed Graph
- Each node is a
session/transaction
- Edge represents a
wait on a lock
Waiting
on ‘Bob’
Waiting on
‘Alice’

54. S1
S2
S1 S2
S2 S1
S2 Waits on S1 S1 Waits on S2
Sumedh Pathak | Citus Data | QCon London 2018

55. S1
S2
S1 S2
S2 S1
S2 Waits on S1 S1 Waits on S2
Distributed
Deadlock
Detector
S1
S2 S2
S1
Sumedh Pathak | Citus Data | QCon London 2018

56. keys to scaling transactions
- 2PC to ensure atomic transactions across nodes
- Deadlock Detection—to scale complex & concurrent
transaction workloads
- MVCC
- Failure Handling
4

57. It’s 2018. Distributed can be Relational
- Scale tables—via sharding
- Scale SQL—via distributed relational algebra
- Scale transactions—via 2PC & Deadlock Detection
Sumedh Pathak | Citus Data | QCon London 2018
Now, how do we implement all of this?

58. Postgres
Sumedh Pathak | Citus Data | QCon London 2018

59. PostgreSQL
Planner
Executor
Custom scan
Commit / abort
Extension (.so)
Access methods
Foreign tables
Functions
...
...
...
...
...
...
...
CREATE EXTENSION ...

60. PostgreSQL PostgreSQL
PostgreSQL
shardsshardsshard
shardsshardsshard
SELECT … FROM distributed_table …
SELECT … FROM shard… SELECT … FROM shard…
Citus

61. PostgreSQL
Citus
PostGIS PL/Python
JSONB
2PC
Replication...
Sequences
Indexes
Full-Text SearchTransactions
…
dblink
Sumedh Pathak | Citus Data | QCon London 2018
Rich
ecosystem of
tooling, data
types,
& extensions
in
PostgreSQL

62. PostgreSQL can be
extended into an all-purpose
distributed RDBMS

63. I believe the future of distributed
databases is relational.

64. Thank you!
Sumedh Pathak
sumedh@citusdata.com
@moss_toss | @citusdata | citusdata.com
QCon London 2018 | The Future of Distributed Databases is Relational