Takeaway from Evolving for the Cloud

Original presentation:

http://www.infoq.com/presentations/Evolving-Programming-for-the-Cloud

Universal Scalability Law:

Throughput limited by two factors:

    1. contention - bottleneck on shared resources
    2. coherency - communication and coordination among multiple nodes

Programing in the cloud:

Parallelism

routing requests
aggregating services
queuing work
Async I/O and Futures
workload/data/processing partition
pipelines, map-reduce

Layering
strict layered system
common code in frameworks
 -- configuration
 -- instrumentation
 -- Failure management

Services
composition
dependency
addressing
persistence
deployment

State management
Stateless instances
Durable state in persistent storage

Data model
Distributed key-value stores
Constrained by design (no ER, limited schema)
plan to shard
plan to de-normalize
  - optimize for reads
  - pre-calculate joins/aggregations/sorts
  - use async queue to update
  - tolerate transient inconsistency

Failure handling
Expect failures and handle them
 - Graceful degradation
 - timeouts and retries
 - throttling and back-off

Testing
Test early and often (Test-driven and Test-first)
Test end-to-end

Operation in the cloud:

DevOps mindset
 - automate, automate, automate
 - If you do it more than twice, script it

Configuration Injection
 - Inject at boot time and run-time
 - Separate code and configuration

Instrumentation
 - Fully instrument all components
 - Remotely attach/profile/debug components
 - Logging is insufficient
 - Use a framework

Monitoring
 - monitor requests end-to-end
 - monitor activity and performance
 - metrics, metrics, metrics
 - understand your typical system behavior

Metering
 - Variable cost
 - Efficiency matters (inefficiency multiplied by many instances)