Transactions at Scale

The shift to the cloud required developers and IT professionals to re-think the way they build their applications. Cloud native applications have certain unique properties that make them well suited for running on modern cloud infrastructure, including elasticity and self-healing capabilities that help to minimize downtime when faced with changes in load or other issues.
Despite these advantages, some applications can still be challenging to scale effectively in a cloud environment, particularly relational databases. Fortunately, Cloud Spanner is a platform that helps businesses overcome these challenges and we will shed some light on real use cases.

Transcript

2. 2 Customer Balanc e Vourcers Used time Arbitrary Metadata …. 99282 1231412 { “id” : “123”, } 12312421321L foobert Customers Vouchers List of Records Book virtual currency and vouchers for customers on a global scale Use Case
3. 3 Bird’s-Eye View
4. Non-Functional Requirements Response time < 100 ms High Availability Zero Downtime 1000 req/s Mio Customers Strong Consistency Scalability
5. 5 Dynamic Load
6. 6 Infrastructure Overview
7. What Data Store? Existing SQL Knowledge Need for TRANSACTIONS Relational Semantics
8. 8 ACID Transactions & Database Constraints External Consistency Completed operation will be visible to all (globally) Globally consistent order of operation Atomic Writes Consistency, what?
10. • SCALING CHALLENGES Scaling DBMS
11. • SCALING CHALLENGES Vertical Scaling Scaling DBMS
12. • SCALING CHALLENGES Vertical Scaling Scaling DBMS
13. • SCALING CHALLENGES Vertical Scaling Scaling DBMS
14. • SCALING CHALLENGES Vertical Scaling Read bottleneck? Add read replica! Scaling DBMS
15. • SCALING CHALLENGES Vertical Scaling Read bottleneck? Add read replica! Write Bottleneck? Data Sharding ! Consistency? Scaling DBMS
16. • SCALING CHALLENGES Vertical Scaling Read bottleneck? Add read replica! Write Bottleneck? Data Sharding ! Consistency? Scaling DBMS
17. • SCALING CHALLENGES Vertical Scaling Read bottleneck? Add read replica! Horizontal Scaling Write Bottleneck? Data Sharding ! Consistency? Scaling DBMS
18. 18 SQL SQL New SQL
20. 20 • Algorithm used to achieve consensus among a distributed set of computers that communicate in an asynchronous network • 2 phase commit protocol • Majority wins • Google implemented own version of PAXOS tailored to their high availability network! BUILDING BLOCKS II – TRUE TIME API (TT) [Andr ] Consensus
21. 21 Concept of Real Time (wall clock) for global ordering BUILDING BLOCKS II – TRUE TIME API (TT) [Andr ] Method Returns TT.now() TTinterval:[earliest,latest] TT.after(t) true if t has definitely passed TT.before(t) true if t has definitely not arrived Order Bounded error interval (Tmin, Tmax) Using Atomic Oscillators and GPS Antennas on Racks to synchronize clock with data center (minimize clock skew < 1ms)
22. 22 Google runs its own private global network only over Google-controlled routers and links three independent fibers connecting in each data center BUILDING BLOCKS III – THE NETWORK Network Limiting „Blast Radius“ for software configuration and updates A regional setup is defined to have a round trip time of 2ms
23. 23 DISTRIBUTED MULTI VERSION KEY-VALUE STORE [JAMC] Replacement Driver ID first_name Second_name info updated Singer(1) „Sam“ „Smith“ <Bytes> 2019-10- 12T14:00:00 Singer(2) „Adele“ <Bytes> 2019-10- 18T15:00:00 Singer(3) „Britney“ „Spears“ <Bytes> 2019-11- 12T13:00:00 Singer(4) „Toni“ „Polster“ <Bytes> 2019-09- 12T14:00:00 Singer(5) „Marilyn“ „Manson“ <Bytes> 2018-01- 10T14:00:00 Span Server 1 Span Server 2 Span Server 3 Span Server 4 Span Server 5 [GOOG] Potential split boundary Real split boundary Internals
24. 24 SQL and PostgreSQL Interface Client Libraries for many languages available (e.g. ORM Mapper, JDBC Drivers, Spring Data, etc.) Open Source Emulator for local testing Operate with IaC (e.g. Terraform, CNRMs, etc.) Auto Scaling of Spanner nodes based on node metrics Advanced Tracing Tools Well intergrated with other Google services, e.g. AutoML and BigQuery Auto Backup But How Developer Experience
25. 25 REQUEST LATENCIES Request Latency Request Rate
26. 26 Relational Models are not dead Cloud Spanner is a fully managed DBAS product Brings dynamical horizontal scalability to relational databases (build for the future) Affordable even for small businesses or startups (total cost of ownership) It has strong global consistency High peformant on OLTP workloads, Small, short-scoped and indexed lookups Very high Availability 99,999 in multi-region setup (4 9’s in single zone) Data Center Locations in the EU (GDPR) Enterprise Grade Security But How Takeaway