Amazon DynamoDB
Serverless key-value at any scale: partition-key routing, quorum replication, single-digit ms.
DynamoDB is AWS managed key-value and document storage. Tables are hashed across partitions by a partition key, optionally ordered by a sort key; every item is replicated three times across availability zones with quorum agreement. Access is operations, not queries: GetItem, PutItem, Query by key — with GSIs for alternate access paths.
Partition-key routing
The partition key hashes to a physical storage partition; the sort key orders items inside it. Table design IS access design — the key schema determines what you can fetch in O(1).
GSIs for alternate paths
Global secondary indexes replicate the table under a different key schema, maintained asynchronously — extra access patterns at extra write cost.
Quorum consistency
Writes acknowledge at quorum; eventually consistent reads may see stale values, strongly consistent reads always read the quorum leader.
Streams and triggers
DynamoDB Streams publishes item-change events in order per key — feeding analytics, replicas in other tables, and event-driven architectures.
Access patterns are the schema
DynamoDB inverts relational design: you enumerate the queries first, then design tables so each query is a direct key lookup. The partition key spreads load across storage partitions — a hot partition key throttles, no matter the table size.
The single-table design pattern packs multiple entity types into one table using overlapping key schemas — powerful and famously easy to get wrong.
GetItem:
Table: orders
Key: { customer_id: "c-42", order_ts: 1699999999 }
→ one item, O(1), single-digit ms
Capacity, throttling, and cost
Provisioned or on-demand capacity units meter reads and writes; exceeding partition throughput throws throttling errors rather than queuing. Cost scales with throughput and storage — predictable if your access patterns are uniform, expensive if they spike.
Diagrams
[caption] The 2012 DynamoDB paper (Sivasubramanian) describes this design publicly.
A write routes by partition-key hash to one partition leader, replicates to two bystander replicas across AZs, and acknowledges at quorum into log-structured storage.
Base table partitions vs a global secondary index projection — the GSI duplicates chosen attributes under an alternate key.
[caption] Illustrative: 8 partitions, one GSI projecting an alternate access path. Hot partitions throttle independently of table size.
Base table partitions vs a global secondary index projection — the GSI duplicates chosen attributes under an alternate key.