Skip to content

Latest commit

 

History

History
33 lines (28 loc) · 2.11 KB

OUTLINE.md

File metadata and controls

33 lines (28 loc) · 2.11 KB
Raw

Questions

  • Is it OK to assume basic Service Fabric knowledge? Fair assumption
  • Is it OK to assume basic ASB knowledge? Fair assumption
  • Basic messaging knowledge? Fair assumption

Outline

  • Quick ramp up on Service Fabric

  • Stateful Services focus

  • What is partitioning and why would I care?

  • Uniform hashing (maybe only relevant for Actors)

  • Business data defines partitioning / sharding (might be tricky if you want to uniformly distribute the data, you don't want to end up with imbalances on the cluster)

  • Mention: Repartitioning is hard in SF

  • Why not a single partition: Memory consumption, Performance, loosing data?

  • Visualize how data is sticky to a given partition

  • Service Fabric patterns to access partitioned data from withing the cluster and outside the cluster

  • Within: use Service Partition resolver to get to a service on a given partition

  • Outside: Either use Reverse Proxy which is built in or expose stateless service which maps to stateful service

  • Transactionality does not work cross service

  • Service needs to expose the data in the reliable collection over the communication listener

  • Brokered queues such as RabbitMQ or Azure Service Bus, with endpoint communication listener and scaled/partitioned service you can do the following:

  • Clients can be inside the cluster or outside the cluster

  • Competing consumer on the same queue

  • Take a message out of the queue on any instance, see if key matches local partition, if not put it back into the queue

  • Worst case a lot of back and forth

  • No internal rerouting is efficiently possible because instanes are not uniquely addressable

  • Uniquely addressable queue per instance

  • Reliable Queues (mention can't be used really)

  • Some messaging patterns do not require stickyness. For example when you just want a command being handled by any available worker. Complexity increases when you a worker wants to reply back to the command server instance because it uses a process manager to orchestrate the work process.

  • Routing and rerouting into the cluster and within the cluster