Building Scalable Stateful Services

The Stateless Service design principle has become ubiquitous in the tech industry for creating horizontally scalable services. However our applications do have state, we just have moved all of it to caches and databases. Today as applications are becoming more data intensive and request latencies are expected to be incredibly low, we’d like the benefits of stateful services, like data locality and sticky consistency. In this talk I will address the benefits of stateful services, how to build them so that they scale, and discuss projects from Halo and Twitter of highly distributed and scalable services that implement these techniques successfully.

Transcript

1. Building Scalable Stateful Services StrangeLoop 2015
2. Caitie McCaffrey Distributed Systems Engineer Tech Lead Observability @ Twitter @caitie caitiem.com
3. Stateless Services Service
4. Stateless Services Service
5. Stateless Services Service Service
6. Stateless Services Service Service
7. Stateless Services Service Service
8. Stateless Services Service Service Service
9. Stateless Services Service Service Service
10. Stateless Services Service Service Service
11. Stateless Services Service Service Service Service
12. Stateless Services Service Service Service Service
13. Stateless Services Service Service Service Service
14. Stateless Services Service Service Service Service Service
15. Stateless Services Service Service Service Service Service
16. Data Shipping Paradigm Service Service Service
17. Data Shipping Paradigm Service Service Service
18. Data Shipping Paradigm Service Service Service
19. Data Shipping Paradigm Service Service Service
20. Data Shipping Paradigm Service Service Service
21. Data Shipping Paradigm Service Service Service
22. Data Shipping Paradigm Service Service Service
23. Data Shipping Paradigm Service Service Service
24. Data Shipping Paradigm Service Service Service
25. Data Shipping Paradigm Service Service Service
26. Data Shipping Paradigm Service Service Service
27. Data Shipping Paradigm Service Service Service
28. Data Shipping Paradigm Service Service Service
29. Data Shipping Paradigm Service Service Service
30. Data Shipping Paradigm Service Service Service
31. Overview Benefits Building Real World Caution
32. Data Locality For Low Latency & Data Intensive Services
33. Function Shipping Paradigm Service Service Service
34. Function Shipping Paradigm Service Service Service
35. Function Shipping Paradigm Service Service Service
36. Function Shipping Paradigm Service Service Service
37. Function Shipping Paradigm Service Service Service
38. Function Shipping Paradigm Service Service Service
39. Function Shipping Paradigm Service Service Service
40. Function Shipping Paradigm Service Service Service
41. Function Shipping Paradigm Service Service Service
42. Function Shipping Paradigm Service Service Service
43. Function Shipping Paradigm Service Service Service
44. Sticky Connections & Consistency Additional Available Consistency Models
45. Linearizable Sequential Causal Pipelined Random Access Memory Read Your Write Monotonic Read Monotonic Write Write From Read Consistency Models CP Consistency AP Consistency
46. Linearizable Sequential Causal Pipelined Random Access Memory Read Your Write Monotonic Read Monotonic Write Write From Read CP Consistency AP Consistency AP Consistency w/ Sticky Connections Consistency Models
47. - Werner Vogel 2007 “Whether or not read-your-write, session and monotonic consistency can be achieved depends in general on the "stickiness" of clients to the server that executes the distributed protocol for them… Using sessions, which are sticky, makes this explicit and provides an exposure level that clients can reason about.”
48. Building Sticky Connections For Low Latency & Data Intensive Services
49. Building Sticky Connections
50. Building Sticky Connections
51. Persistent Connections Load Balancing Problems No Stickiness Once Connection Breaks Problems
52. Persistent Connections Load Balancing Problems No Stickiness Once Connection Breaks Problems
53. Routing Logic • Cluster Membership • Work Distribution Problems to Solve
54. Routing Logic • Cluster Membership • Work Distribution Problems to Solve
55. Routing Logic • Cluster Membership • Work Distribution Problems to Solve
56. Static Cluster Membership
57. Static Cluster Membership
58. Static Cluster Membership
59. Dynamic Cluster Membership
60. Dynamic Cluster Membership
61. Dynamic Cluster Membership
62. Dynamic Cluster Membership
63. Dynamic Cluster Membership
64. Dynamic Cluster Membership Gossip Protocols Consensus Systems Availability vs Consistency
65. Work Distribution Consistent Hashing Distributed Hash Tables Random Placement
66. Random Placement Write Anywhere Read from Everywhere
67. Random Placement Write Anywhere Read from Everywhere
68. Random Placement Write Anywhere Read from Everywhere
69. Random Placement Write Anywhere Read from Everywhere
70. Random Placement Write Anywhere Read from Everywhere
71. Random Placement Write Anywhere Read from Everywhere
72. Random Placement Write Anywhere Read from Everywhere
73. Consistent Hashing Deterministic Placement Node A Node B Node C NodeD Consistent Hashing & Random Trees: Distributed caching protocols for relieving hot spots on the World Wide Web
74. Distributed Hash Table Non- Deterministic Placement Node B Node A Node C
75. Distributed Hash Table Non- Deterministic Placement Node B Node A Node C
76. Distributed Hash Table Non- Deterministic Placement Node B Node A Node C
77. Distributed Hash Table Non- Deterministic Placement Node B Node A Node C
78. Distributed Hash Table Non- Deterministic Placement Node B Node A Node C
79. Stateful Services In the Real World
80. Scuba is a fast, scalable, distributed, in-memory database built at Facebook. It is the workhorse behind code regression analysis & bug report, revenue, and performance debugging Fan-out request to all machines in the cluster Compose Results Return Results and Completeness
81. Scuba is a fast, scalable, distributed, in-memory database built at Facebook. It is the workhorse behind code regression analysis & bug report, revenue, and performance debugging Fan-out request to all machines in the cluster Compose Results Return Results and Completeness
82. Scuba is a fast, scalable, distributed, in-memory database built at Facebook. It is the workhorse behind code regression analysis & bug report, revenue, and performance debugging Fan-out request to all machines in the cluster Compose Results Return Results and Completeness
83. Uber Ringpop is an open- source Node.js library that brings application-layer sharding to many of their dispatching platform services. Swim Gossip Protocol Consistent Hashing +
84. Uber Ringpop is an open- source Node.js library that brings application-layer sharding to many of their dispatching platform services. Swim Gossip Protocol Consistent Hashing +
85. Uber Ringpop is an open- source Node.js library that brings application-layer sharding to many of their dispatching platform services. Swim Gossip Protocol Consistent Hashing +
86. Orleans Cluster Orleans is a runtime and Programming model for building distributed systems based on the Actor Model from the eXtreme Computing Group at MSR Gossip Protocol Consistent Hashing + + Distributed Hash Table Actor Actor Actor Actor Actor
87. Orleans Cluster Orleans is a runtime and Programming model for building distributed systems based on the Actor Model from the eXtreme Computing Group at MSR Gossip Protocol Consistent Hashing + + Distributed Hash Table Actor Actor Actor Actor Actor
88. Orleans Cluster Orleans is a runtime and Programming model for building distributed systems based on the Actor Model from the eXtreme Computing Group at MSR Gossip Protocol Consistent Hashing + + Distributed Hash Table Actor Actor Actor Actor Actor
89. Orleans Cluster Orleans Distributed Hash Table
90. Orleans Cluster Orleans Distributed Hash Table
91. Orleans Cluster Orleans Distributed Hash Table
92. Orleans Cluster Orleans Distributed Hash Table
93. Orleans Cluster Orleans Distributed Hash Table
94. Orleans Cluster Orleans Distributed Hash Table
95. Orleans Cluster Orleans Distributed Hash Table
96. Orleans Cluster Orleans Distributed Hash Table
97. Caution Stateful Services Ahead
98. Unbounded Data Structures Implicit Assumptions are the Killer of Distributed Systems
99. Memory Management Get Ready to Make Friends with the Garbage Collector Profiler
100. Reloading State • First Connection • Recovering From Crashes • Deploying New Code
101. Fast Restarts at Facebook “Our Key Observation is that we can decouple the memory lifetime from the process lifetime. When we shutdown a server for a planned upgrade.”
102. Conclusion Data Locality & Available Consistency Cluster Membership & Work Distribution Successful Statefull Real World Systems Caution: Some New Challenges
103. Should I Read Papers?
104. Should I Read Papers? YES!
105. Thank You Kyle Kingsbury Chris Meiklejohn Tom Santero Ines Sombra @Caitie