high-performing-architectures on Alperen Doc

s3-storage-and-performance

Mon, 01 Jan 0001 00:00:00 +0000

S3 (Simple Storage Service) Storage Classes and Performance#

Two things get tested: which storage class for which data (mostly a cost question - see s3-cost - but with performance implications) and how to move big data in and out fast.

Storage classes at a glance#

Class	Durability / AZs (Availability Zone)	Access pattern	Min storage
S3 Standard	99.99% / 3+ AZs	Frequent, low latency	-
Standard-IA	99.99% / 3+ AZs	Infrequent but instant	30 days
One Zone-IA	99.5% / 1 AZ	Infrequent, re-creatable	30 days
Intelligent-Tiering	99.99% / 3+ AZs	Unknown/changing patterns	None (small monitoring fee/object)
Glacier Instant Retrieval	99.99% / 3+ AZs	Archival, ms retrieval	90 days
Glacier Flexible Retrieval	99.99% / 3+ AZs	Archive, minutes-hours	90 days
Glacier Deep Archive	99.99% / 3+ AZs	Archive, 12-48h retrieval	180 days

Rules of thumb:

block-and-file-storage-ebs-efs-fsx

Mon, 01 Jan 0001 00:00:00 +0000

Block and File Storage: EBS, EFS and Amazon FSx Deep Dive#

S3 (Simple Storage Service) handles objects, but applications running on EC2 (Elastic Compute Cloud) or containers often need block storage (like a local hard drive) or shared file storage (NFS/SMB). Choosing the right storage type, volume tier, or FSx flavor is one of the most heavily tested topics on the SAA-C03 exam.

Block vs File vs Object at a glance#

flowchart TB
 s{Storage Requirement?}
 s -->|Single EC2 / low-latency DB disk| ebs[EBS / Instance Store]
 s -->|Shared Linux POSIX across AZs| efs[Amazon EFS]
 s -->|Windows SMB / AD share| fsxw[FSx for Windows]
 s -->|HPC / ML + S3 integration| fsxl[FSx for Lustre]
 s -->|Enterprise NetApp / multi-protocol| fsxn[FSx for NetApp ONTAP]
 s -->|HTTP REST / static / archive| s3[Amazon S3]

Amazon EBS (Elastic Block Store)#

EBS volumes are network-attached block disks locked to a single Availability Zone (AZ). To move a volume to another AZ, you snapshot it and restore in the target AZ.

caching-elasticache-cloudfront

Mon, 01 Jan 0001 00:00:00 +0000

Caching: ElastiCache, CloudFront and DAX (DynamoDB Accelerator)#

The fastest query is the one you never make. Caching questions on the exam come down to: what layer (CDN, application, database) and which engine (Redis vs Memcached vs DAX).

Cache placement - pick the layer#

Latency source	Fix
Static content far from users	CloudFront (edge caching)
Repeated DB queries / session data	ElastiCache in front of the DB
DynamoDB reads need microseconds	DAX
API responses repeat	API Gateway caching

ElastiCache: Redis vs Memcached#

	Redis	Memcached
Data structures	Rich (lists, sets, sorted sets, streams)	Simple key-value strings
Replication / failover	Yes (Multi-AZ with auto failover)	No
Persistence	Yes (snapshots, AOF)	No
Clustering	Yes (cluster mode, sharding)	Multi-threaded scaling
Pub/Sub, Lua, transactions	Yes	No
Typical use	Session store, leaderboards, chat, queues	Simple object caching

Exam shortcut: “failover / durability / data structures / sorted leaderboard” → Redis. “Pure cache, simplest and fastest horizontal scaling” → Memcached.

database-selection-matrix

Mon, 01 Jan 0001 00:00:00 +0000

Database Selection Matrix#

The most common question format in Domain 3: “the application needs X - which database?” There are ~10 AWS database services and exactly one right answer per workload. Learn the workload → service mapping.

The decision table#

Workload	Service	Why
Relational OLTP (online transaction processing), joins, transactions, known schema	RDS (Relational Database Service) (or Aurora for cloud-native scale)	Managed SQL, Multi-AZ, replicas
Massive-scale key-value, any traffic shape, single-digit ms	DynamoDB	Serverless NoSQL, unlimited scale
Hot data, microsecond reads, session store	ElastiCache (Redis/Memcached)	In-memory
Data warehouse, OLAP (online analytical processing), columnar analytics	Redshift	Petabyte analytics
Interactive SQL on S3 (Simple Storage Service) data lake	Athena	Serverless, pay per query
Search, autocomplete, log analytics	OpenSearch	Full-text search
Document workloads, MongoDB-compatible	DocumentDB	Mongo API without Mongo ops
Graph relationships (social, fraud, routes)	Neptune	Graph DB
Time series (IoT, metrics)	Timestream	Built for timestamps
Immutable ledger / audit	Aurora PostgreSQL / S3 Object Lock (Note: AWS deprecated QLDB* on July 31, 2025)*	Cryptographic verifiability / WORM
In-memory cache with SQL on RDS-adjacent	Redis	Microsecond latency
Query exabytes in S3 directly from a data warehouse	Redshift Spectrum	Joins Redshift tables with S3 Parquet files without loading

RDS vs DynamoDB - the heavyweight bout#

Pick RDS/Aurora when:

dynamodb-deep-dive

Mon, 01 Jan 0001 00:00:00 +0000

DynamoDB Deep Dive#

DynamoDB shows up in ~a third of all “which database” and “how do I scale reads” questions. The exam-level knowledge is: partitioning, capacity modes, indexes, DAX (DynamoDB Accelerator) and Global Tables.

Data model basics#

Table → items (rows) → attributes (columns) - schemaless except the key
Partition key (required) - hashed to decide storage partition; good partition keys spread load evenly
Composite key = partition key + sort key - enables range queries (WHERE orderId = X AND date BETWEEN...)
Item max size 400 KB
Access is via Query (key condition) or Scan (everything - avoid it)

Capacity modes#

	On-demand	Provisioned
Pricing	Per request (pay per use)	Per hour on WCU (write capacity unit)/RCU (read capacity unit) + autoscaling
Best for	Unknown, spiky traffic	Steady, predictable traffic
Scaling	Instant, no planning	Auto scaling policy on utilization

1 WCU = 1 KB write per second; 1 RCU = 4 KB strongly consistent read per second (2x for eventually consistent)
Reads: eventually consistent by default (duplicates at read time possible), strongly consistent = 2x RCU

Indexes#

	LSI (local secondary index)	GSI (global secondary index)
Key	Same partition key, new sort key	Any key (new partition + sort)
Consistency	Can be strongly consistent	Always eventually consistent
Limits	5 per table	20 per table
Availability	Same partition as item	“Global” = separate index

Sparse index trick: items missing the indexed attribute are simply not in the index - filter by putting an attribute only on some items.

compute-performance-ec2-options

Mon, 01 Jan 0001 00:00:00 +0000

Compute Performance: EC2 (Elastic Compute Cloud) Options#

“Which instance type / placement / purchase option” questions test whether you know the instance family alphabet and the three placement group shapes. (Pricing is covered in pricing-models.)

Instance families#

Family	Optimized for	Mnemonic
T (t3, t4g)	Burstable general purpose (baseline CPU + credits)	“t” = tiny/test
M (m5, m6i)	Balanced general purpose	“m” = middle
C (c5, c7g)	Compute intensive (app servers, game servers, batch)	“c” = compute
R (r5, x2idn) / X / Z	Memory intensive (in-memory DB, caches, big data)	“r” = RAM
P / G / VT / Inf	GPU/ML training and inference	“p/g” = graphics
Hpc / D	High performance compute, high disk throughput
Mac	macOS build farms

Burstable (T family) detail: instances earn CPU credits when idle and spend them under load; “unlimited” mode can pay for extra credits. “Variable low CPU baseline” → T family.

serverless-and-containers-lambda-ecs-eks

Mon, 01 Jan 0001 00:00:00 +0000

Serverless & Containers: Lambda, RDS Proxy, API Gateway, ECS and EKS#

Modern architecture questions on SAA-C03 revolve around eliminating server management while respecting service limits, networking rules, and database connection pools.

AWS Lambda - limits, networking and scaling#

Hands-on: aws-lambda, lambda-api-gateway.

Hard limits the exam tests#

Max execution timeout: 15 minutes (900 seconds). If a job takes 25 minutes → use AWS Fargate or AWS Batch, never Lambda!
Memory: 128 MB to 10 GB — CPU and network bandwidth scale proportionally with memory. “A CPU-bound Lambda function is too slow” → increase its memory configuration.
Storage:
- /tmp ephemeral storage: 512 MB default, configurable up to 10 GB (shared across warm invocations in the same execution environment — great for caching static reference files).
- Need persistent shared storage across invocations or big ML models? Mount Amazon EFS (Elastic File System) (requires Lambda in a VPC).
Deployment package: 50 MB zipped (250 MB unzipped including Lambda Layers), or up to 10 GB as a container image from ECR (Elastic Container Registry).

Concurrency: Reserved vs Provisioned#

Default account concurrency limit: 1,000 concurrent executions per region (soft limit).
Reserved Concurrency: caps the max concurrency for one function so it can’t starve other functions (and guarantees that slice of capacity). Costs nothing extra; setting it to 0 throttles/disables the function.
Provisioned Concurrency: keeps N environments pre-initialized and warm → eliminates cold starts for latency-sensitive APIs (costs money).

Lambda in a VPC & the RDS Proxy pattern#

flowchart LR
 apigw[API Gateway] --> lambda[Lambda in Private Subnet]
 lambda -->|pooled connections| proxy[Amazon RDS Proxy]
 proxy --> rds[(RDS / Aurora Multi-AZ)]
 lambda -->|outbound internet via| nat[NAT Gateway]

Lambda in a VPC has NO internet access by default (even in a “public” subnet, Lambda ENIs (Elastic Network Interface) never get public IPs!). To call an external public API from a VPC-connected Lambda, put Lambda in a Private Subnet with a route to a NAT Gateway.
The RDS connection storm problem: Lambda scales out to 1,000 concurrent instances in seconds; each opens a DB connection and crashes RDS (Too many connections).
- The fix: Amazon RDS Proxy (placed in the VPC between Lambda and RDS/Aurora).
- Pools and shares DB connections, enforces IAM (Identity and Access Management) authentication with credentials stored in Secrets Manager, and reduces Multi-AZ failover time by up to 66% by preserving client connections during failover.

Sınav Tuzağı (Exam Trap): “Lambda fonksiyonları trafik patlamasında RDS veritabanında bağlantı hatası (Too many connections / connection exhaustion) veriyor” sorusunun cevabı Read Replica veya daha büyük RDS değil, Amazon RDS Proxy‘dir!

data-transfer-and-hybrid-services

Mon, 01 Jan 0001 00:00:00 +0000

Data Transfer and Hybrid Services#

Moving data is its own question family: “we have 500 TB on-premises, get it to S3 (Simple Storage Service)” or “users must still write to a local file server”. Each answer is a specific service - and mixing them up is easy, so learn the one job per service.

The transfer service lineup#

Service	Job	Scenario keyword
DataSync	Scheduled bulk transfer (NFS/SMB/HDFS/object → S3/EFS (Elastic File System)/FSx)	“migrate file shares”, “agent, scheduled, fast”
Transfer Family	Managed SFTP / FTPS / FTP endpoints backed by S3/EFS	“partners upload via SFTP”
Storage Gateway	Hybrid: local file/volume/tape interface → S3/EBS-backed	“on-prem apps keep using local protocols”
Snow Family	Physically ship petabytes	“no bandwidth / 100 TB+”
AWS Backup	Centralized backup plans	“tag-based backup across services”

Storage Gateway modes (classic exam trio)#

File Gateway - NFS/SMB share backed by S3 objects (“local file share, objects in S3”)
Volume Gateway - iSCSI block storage; cached (hot data local, everything in S3) vs stored (full copy local, async backup to EBS (Elastic Block Store) snapshots)
Tape Gateway - virtual tape library (VTL) for legacy backup software

Snow Family#

Device	Scale	Notes
Snowcone	~8-14 TB, rugged edge device	Edge computing + transfer
Snowball Edge	80-210 TB	The standard “ship a box” answer
Snowmobile	100 PB per truck	Exabyte-scale, literally a shipping container

“Petabytes of data, limited bandwidth” → Snowball. “Ongoing streaming” → Kinesis / DataSync instead.