Compute¶

EC2¶

Pricing:
- On demand: by hour (recent change: Linux instances by seconds)
- Reserved: 1 or 3 year commitment
  - Standard
  - Convertible (changeable, as long as higher priced)
  - Scheduled: only during a time window
- Spot: bid prices
  - If you terminate, pay the full hour, if Amazon terminates, hour is free
- Dedicated hosts: (can be purchased on-demand, reserved)
Types: General purpose or optimized for Compute, Storage, Memory
Root volume types:
- EBS -- permanent
- Instance store -- Ephemeral cannot stop instance, only reboot/terminate
Instance Status check:
- System status: The underlying h/w and hypervisor are okay
- Instance status: The guest OS is reachable, not hung
Default AMI can’t encrypt root volume
Metadata URL: http://169.254.169.264/latest/meta-data/
Placement group: For low latency, high network throughput, eg Hadoop cluster
- Must be within single AZ
AMIs are regional

load balance across AZ
uses launch configuration as template to launch additional instances
have life-cycle hooks to take action before and after instance is put in the fleet
Will receive health check data from ELB when selected
Policies
- Simple: increase/decrease instances based on a metric alarm
- Stepped: same as simple, but multiple thresholds
- Target tracking: dynamically auto scale based on a metric (eg CPU utilization)

Application, Network and Gateway
route traffic to single AZ or multiple AZ
health check: ELB routes traffic only to healthy instance, containers, IP addresses, microservices, lambda and appliances
- Unhealthy: response time exceed specified number of consecutive times
- Needs a health-check URL
supports sticky session, i.e. requests from the same client go to the same target
Gateway LB allow third-party virtual appliances to be deployed and scale
- e.g. DDoS protection service, intrusion detection and prevention devices
Application LB
- Features: WebSocket, sticky sessions, TLS offloading, User authentication (using Amazon Cognito)
- content based routing: path, header, http method, query string parameter, source IP (CIDR based)
- can route to outpost, lambda functions, containers running on single instance via multiple ports
Network
- static IP, very high performance
Classic LB: legacy, for applications built with EC2 classic network

feature	Application	Network	Gateway
Layer	7	4	3 (Gateway) 4 (LB)
Target	IP, instance, Lambda	IP, Instance	IP, instance
Proxy	Yes	Yes	No
Protocol	HTTP(S), gRPC	TCP, UDP, TLS	IP
Reachable Via	VIP	VIP	Route table entry

Can cache results for TTL period
Can scale automatically, can be throttled to prevents attacks
Can log results to CloudWatch
CORS: When using Javascript or AJAX that use multiple domains (eg S3 images, fonts etc to be referenced from external website) enable CORS
Lambda can run up to 15 minutes, but API Gateway has a limit of 30 seconds

is a managed cluster for BigData processing. Cluster consists of
- One master node to control the cluster and distribute data to other nodes
- One or more core node runs application data and access HDFS
- One or more task node that can run application but does not access HDFS (typically spot instances)
can run in Cluster mode, which is long running, or Step execution mode which allows you to add steps to run when you create a cluster and cluster terminates after the steps run.
can use multiple instance groups, for example, to take advantage of spot instances
storage:
- HDFS: ephemeral storage that is reclaimed when cluster is terminated. Ideal for a lot of random IO for transient storage
- EMRFS: S3 that works as HDFS. Ideal for persistent starting input and final output storage
- Local: instance level ephemeral storage
YARN is the default resource manager
lifecycle of a cluster
- Starting provisions EC2 instances with either the default or a custom AMI
- bootstrapping install custom applications via bootstrapping
- install native applications on each instance. e.g. pig, hive, spark etc
- Running running the work that is specified at cluster creation. Can add more steps that get queued behind current work
- Waiting/shutting down after the curren work has been processed, if auto terminate is not set/set.
- completed after the cluster has been terminated