Scalable Systems Midterm 1

Cloud computing

On-demand network access to shared configurable resources that can be rapidly provisioned and released.

Provisioning vs. deployment

Provisioning creates/configures infrastructure; deployment places application code/configuration onto it.

On-demand self-service

Consumers can provision resources without a provider employee handling each request.

Broad network access

Cloud capabilities are available over networks through mechanisms usable by many client platforms.

Resource pooling

A provider dynamically assigns shared physical resources to isolated customers.

Rapid elasticity

Capacity can be acquired and released quickly as demand changes.

Measured service

The provider meters resource use, such as compute time, storage, requests, or transferred bytes.

Public cloud

Provider-owned infrastructure shared by many customers, such as AWS.

Private cloud

Cloud infrastructure reserved for one organization, even if some workloads are publicly reachable.

Hybrid cloud

A system that uses private/on-prem resources together with public-cloud resources.

Multicloud

Using resources from more than one public-cloud provider.

IaaS

Provider supplies infrastructure building blocks; customer manages the OS, software stack, and application.

PaaS

Provider operates the platform; customer deploys and remains responsible for its application code.

FaaS

A narrow PaaS model where customers provide event handlers and the provider supplies compute and scaling.

SaaS

Provider operates the application; customers mainly use and configure it.

Control plane

Operations that provision, configure, or delete resources.

Data plane

Normal steady-state work, such as an ALB forwarding a request to a backend.

Infrastructure as Code

Infrastructure described as versioned, reviewable, repeatable configuration rather than manual clicks.

AWS Console, CLI, and SDKs

Different clients that ultimately call the same AWS service APIs.

AWS Region

A geographic area containing multiple Availability Zones; most AWS resources belong to one Region.

Availability Zone

A failure-isolated group of datacenters inside one Region.

VPC

A virtual network that exists within one AWS Region.

Subnet

A range of VPC IP addresses that lies entirely within one Availability Zone.

What does choosing a subnet determine for an EC2 launch?

It determines the Availability Zone in which that instance launches.

Private IPv4 address

An address used within the VPC; it is not directly Internet-routable.

Public IPv4 address

An Internet-routable address AWS maps to an instance's private address.

Why is a public IPv4 address not enough for reachability?

A route must exist, security policy must allow traffic, and a process must be listening.

AMI

A reusable image that supplies an EC2 instance's starting software and storage configuration.

Does changing one running instance change its AMI?

No. Changes affect that instance's volume unless a new AMI is explicitly created.

Instance type

The EC2 configuration defining vCPUs, memory, architecture support, and network/storage performance limits.

Do more vCPUs automatically make one CPU-bound thread faster?

No. They allow more threads to run concurrently; one ready thread still runs on one vCPU at a time.

EBS volume

AWS-managed block storage attachable to an EC2 instance in the same Availability Zone.

Can an EBS volume attach directly across Availability Zones?

No. Create a snapshot and a new volume in the other AZ instead.

EBS snapshot

A point-in-time copy of an EBS volume that persists separately and can create new volumes.

AMI launch and root volumes

Each launch gets its own root volume, so later changes do not alter the AMI or another instance.

Process memory vs. EBS

Process memory is volatile; files on EBS can survive reboot and stop/start.

Security group

A set of stateful allow rules attached to a network interface.

What does 'stateful' mean for a security group?

If a connection is allowed in one direction, its response traffic is automatically allowed back.

What must exist for an EC2 connection to work?

An address, network path, network permission, access permission when needed, and a listening process.

Why can SSM replace direct SSH access?

SSM uses an outbound agent connection and AWS identity, so port 22 and distributed private SSH keys are unnecessary.

Instance role

An IAM role whose temporary credentials are made available to software running on an EC2 instance.

Instance role vs. administrator identity

The instance role controls the instance software's AWS calls; the administrator's identity controls their own AWS calls.

EC2 reboot

Restarts the OS; process memory is lost, but the instance, EBS volumes, NIC, private IP, and public IP remain.

EC2 stop/start

Loses process memory but keeps instance ID, EBS volumes, NIC, and private IP; an auto public IP changes.

EC2 termination

Permanently ends the instance; EBS volumes are deleted or retained according to DeleteOnTermination.

Why can a stopped EC2 instance still cost money?

Its compute charge stops, but retained EBS volumes and other allocated resources can still be billed.

P0 Region

us-west-2 (US West, Oregon).

P0 instance type

t3.nano.

P0 request path

Laptop -> SSM port forward -> Uvicorn -> FastAPI route -> in-memory dictionaries.

Management path

The route operators use to administer a system; it can differ from the application's client path.

systemd

Linux service manager that starts and supervises long-running processes such as the P0 URL shortener.

Uvicorn in P0

The HTTP application server process that accepts requests and invokes the FastAPI application.

127.0.0.1

Loopback: reachable only from the same machine.

0.0.0.0

Bind on all IPv4 interfaces, allowing connections through reachable non-loopback addresses too.

Why couldn't P0 be reached directly from the Internet?

Its security group allowed no inbound connection, and the app itself listened only on loopback.

SSM port forwarding

Maps a local laptop port through an authenticated SSM session to a port on the EC2 instance.

What does localhost mean?

The machine where the client is running; laptop localhost and EC2 localhost are different endpoints.

What does GET /health prove?

That the HTTP server and application can accept that request and return a successful response on the tested path.

Why can the wrong AWS Region make an instance look missing?

Regional resources have separate inventories; selecting another Region shows a different instance list.

Why does restarting the P0 process erase short links?

The application stores link state only in that process's memory.

Capability is not approval

A cloud API may technically permit an action even when policy or the assignment says you should not perform it.

Launch template

A versioned EC2 resource describing parameters for future instance launches.

What can a launch template include?

AMI, instance type, security groups, storage, instance role, user data, and other launch settings.

Does creating a launch template from an instance capture later filesystem changes?

No. It copies launch choices, not modifications made to the running instance afterward.

What does a new launch-template version change?

Future launches that use that version; it does not reconfigure already-running instances.

User data

Launch-time text, often a shell script, processed by cloud-init to configure a new instance.

cloud-init

Guest software that processes EC2 user data, normally during the instance's first boot.

Custom AMI vs. user-data setup

A custom AMI can start faster; user data keeps setup scripted but repeats work at every launch.

Vertical scaling

Change the capacity of one machine, such as moving to a larger EC2 instance type.

Horizontal scaling

Change the number of machines doing the work.

Why doesn't horizontal scaling always increase throughput?

A shared bottleneck, such as a database, may limit the whole system regardless of backend count.

Multiple instances in one AZ

They improve tolerance of individual-instance failure but not failure of that Availability Zone.

Why spread instances across AZs?

To reduce correlated failure risk if an entire Availability Zone becomes unavailable.

Reverse proxy

A server that accepts a client request and forwards it to another server.

Load balancer

A reverse proxy that chooses among multiple backend servers providing the same service.

Round-robin balancing

Rotate requests among eligible targets.

Least-outstanding-requests balancing

Prefer a target with fewer unfinished requests.

Health check

A probe used to decide whether a backend is eligible to receive traffic.

Application Load Balancer

AWS-managed HTTP/HTTPS reverse proxy that routes client requests to healthy targets.

ALB listener

The protocol and port on which an ALB accepts client connections.

ALB listener rule

Conditions plus an action, such as forwarding matching paths to a target group.

Target group

A named set of possible backends plus target port, health-check, and routing settings.

Target

A backend endpoint registered in a target group, such as an EC2 instance on port 8080.

Why did P1 bind Uvicorn to 0.0.0.0:8080?

The ALB and other VPC machines must reach the app through the instance's network interface, not only loopback.

Load balancer vs. Auto Scaling group

The load balancer distributes requests; the ASG changes how many EC2 instances exist.

Auto Scaling group

A resource that manages interchangeable EC2 instances using a launch template and selected subnets.

ASG minimum capacity

Lowest desired capacity the group may use.

ASG desired capacity

The current target number of instances the group tries to maintain.

ASG maximum capacity

Highest desired capacity the group may use.

Why isn't desired capacity the same as ready capacity?

Instances may still be launching, failing health checks, draining, or being replaced.

ASG scale-in and draining

The ASG deregisters a target, lets existing work drain, then terminates the instance.

ALB health vs. ASG replacement

ALB health controls routing; if enabled for the ASG, failed ALB health can also trigger replacement.

Manual scaling

An operator directly changes desired capacity.

Scheduled scaling

Capacity settings change at known future or recurring times.

Dynamic scaling

AWS changes desired capacity in response to metrics.

Target tracking

Dynamic scaling tries to keep a chosen metric near a target value.

Step scaling

Alarm ranges specify how much capacity to add or remove.

Stickiness

Routes a client's later requests to the same backend, but does not copy or preserve backend state.

Where should shared durable state live?

In an external database or service reachable by all replaceable application instances.

Main goal of P1a

Build a launch-template version that creates a working backend without manual repair.