Amazon Redshift – Comprehensive Lecture Review

Core Purpose

• Fully managed cloud data warehouse for OLAP workloads (analytics, BI, reporting)
• Executes complex SQL across $10^{2}$ GB–PB+ datasets

Key Architectural Pillars

• Massively Parallel Processing (MPP)
  • “Shared-nothing” cluster: Leader node + compute nodes → slices
  • Leader: parse, optimize, compile to C++, distribute, aggregate, cache
• Columnar storage engine
  • Stores values by column ⇒ scans only needed columns
  • High compression (auto or manual encodings such as AZ64, ZSTD)
  • Zone-maps on sort keys skip whole $1\,\text{MB}$ blocks

Node Families (Provisioned)

• RA3 (decoupled compute / Redshift Managed Storage)
• DC2 (fast SSD, <$5\times10^{2}$ GB data)
• DS2 (cost-optimized HDD, ≤$2$ PB)

Table Design Essentials

• Distribution styles
  • KEY (hash on DISTKEY) → collocated joins
  • ALL (replica on every node) → small dimensions
  • EVEN (round-robin) → default, avoids skew
• Sort keys
  • COMPOUND (prefix ordered)
  • INTERLEAVED (equal weight columns; higher maintenance)

Mandatory Maintenance (Provisioned)

• VACUUM ⇒ reclaim space & resort
• ANALYZE ⇒ refresh column statistics

Elastic & Modern Features

• Spectrum → SQL on S3 data lake without loading
• Concurrency Scaling → transient clusters for burst traffic (free credits each day)
• Data Sharing → zero-copy, live sharing between clusters/accounts/regions
• Serverless → pay-per-second RPUs, auto-provision & scale

Typical Use Cases

• Interactive BI dashboards & real-time metrics
• Batch big-data analytics / log processing
• In-warehouse ML with Redshift ML (SageMaker integration)

Comparison Snapshot

Strengths

• High performance on large scans/aggregations
• Multiple scaling options (RA3, Concurrency, Serverless)
• Cost-effective RIs & pay-per-use models
• Deep AWS integration & standard SQL interface

Limitations

• Manual tuning required for peak speed (keys, maintenance)
• Mis-design ⇒ skew, slow joins, queued queries
• No enforced primary/unique constraints
• Data loads fastest from S3; other sources need extra tooling