展开查看详情

1. HBase/PH ENIX @ Scale

A study of Salesforce’s use of HBase and Phoenix

Lars Hofhansl

Vice President and Principal Architect at Salesforce

Apache HBase, Apache Phoenix committer and PMC member

2.Two Years Ago, I showed you this

3.

4.Zookeeper?

5. Zookeeper?

HBase?

6. Zookeeper?

HBase?

HDFS?

7. Zookeeper?

HBase?

HDFS?

Commodity

Hardware?

8. Zookeeper?

HBase?

Unstructured HDFS?

Data? Commodity

Hardware?

9.HBase/Phoenix are BIG* at Salesforce

* Numbers are from some time in the past and do not reflect the current scale

10.Heavy users of relational and semi-structure data

11.Mix of customer* and internal data

* Through the Salesforce Platform Offering

12.Typical Use Cases

● Samples of Customer Data:

○ Login data to track anomalies in real-time

○ Archiving, historical data moved from operational, relational storage to HBase

○ Denormalized feed views for Salesforce Chatter

○ Chatter @mention low-latency relevancy queries

○ Storage of user activity on marketing campaigns for reporting and AI/ML

● Samples of Internal usage:

○ Periodic thread dumps from all AppServers

○ Machine metrics from all machines

13.Customer Facing

14.> 100 clusters of varying size

15.~4bn write requests / day

16.~80TB written / day

17.That’s about ~8 gbit/s, sustained

18.~600m read requests / day

19.~500GB read / day

20.Central Metrics Database

21.Central Metrics Database

22.Collecting data from > 80.000 machines

23.11.4 trillion metrics stored and growing

24.2.8 tn metrics in 6 months and growing

25.210 bn reads in 6 months