dynobase-icon
Dynobase

DynamoDB Architecture and How It Works

Rafal Wilinski

Written by Rafal Wilinski

Published: November 20th, 2026

Investigating throttling or capacity? Dynobase shows a table's CloudWatch metrics next to its data, tails DynamoDB Streams and can refresh results automatically. Metrics and streams in Dynobase

DynamoDB hashes an item's partition key to locate its data and routes the request to a replica of the partition that holds it. Within an AWS Region, DynamoDB replicates each partition across Availability Zones. The partition leader handles writes and strongly consistent reads, while eventually consistent reads can use any storage replica in the group. A table can scale across many partitions, but a busy key range can still throttle while the rest of the table has spare capacity. DynamoDB paper, section 3 · AWS partition-key guidance

How a DynamoDB request reaches storage

When your application calls a regional DynamoDB endpoint with GetItem or PutItem, a request router authenticates the caller and checks authorization. It uses routing metadata for the table and the target key range to locate the replication group, then sends the operation to a storage node that hosts a replica of that partition. Each storage node can host replicas of many different partitions. The paper calls the distributed metadata store MemDS and describes how routers cache routing information and refresh it as placement changes. DynamoDB paper, sections 3, 6.5, and 6.6

This is the regional path for one item. The metadata lookup tells the router where to send the operation. The storage replica handles the item data. Control-plane calls such as CreateTable take a different route to the autoadmin service, which also handles scaling and fleet health. The DynamoDB paper guide covers the service's history and the paper's durability and failover details. AWS DynamoDB API · DynamoDB paper, section 3

How keys map to physical partitions

A physical partition owns a range of the table's internal key space. DynamoDB hashes the partition key value to find where to look. Several different partition key values can share a physical partition. For a composite primary key, the hash locates the item's collection and the sort key locates the item within that collection. If an item collection spans physical partitions, the sort key also helps determine which range holds the item. AWS partitions and data distribution · AWS split-for-heat explanation

For a tenant audit log, you could use the tenant ID as the partition key and a timestamp plus event ID as the sort key. GetItem uses the complete key to read one event. Query uses the tenant ID and can add a sort-key condition to read just the events in a time range. A tenant's events form one logical item collection even if DynamoDB splits that collection across physical partitions. The partition key guide covers how to pick keys for your access patterns. AWS key conditions for Query · AWS partitions and data distribution

What happens when DynamoDB writes or reads an item?

DynamoDB replicates data across three Availability Zones in a Region. The 2022 paper describes an elected leader in each partition's replication group. The group uses Multi-Paxos to agree on the order of writes. The leader records each write in a write-ahead log and sends the entry to its peers. DynamoDB acknowledges the write after a quorum has persisted the entry. In the three-replica setup described in the paper, that quorum is two replicas. Different partitions have independent leaders, so writes to different ranges can proceed in parallel. AWS DynamoDB resilience · DynamoDB paper, sections 3 and 6.1

  • Strongly consistent reads go to the partition leader in the regional architecture described in the paper. Set ConsistentRead to true on a supported table or local secondary index read when you need it to reflect a successful prior write.
  • Eventually consistent reads can go to the leader or a follower that stores item data. This is the default for table reads. A follower may still be catching up, so the read can return an older committed value or miss a newly created item.

For the audit log, the detail view can use a strong read when the user opens a just-written event by its full table key. An activity feed that can tolerate a brief delay can use eventual reads at half the read-unit cost. See DynamoDB read consistency for API and index behavior. DynamoDB paper, section 3 · AWS read consistency

How partitions split and handle uneven traffic

DynamoDB adds physical partitions as data grows or when increased provisioned throughput requires them. In both capacity modes, design for a limit of 3,000 read units and 1,000 write units per second per physical partition. Table or index capacity can constrain aggregate throughput sooner. Read and write units depend on item size and, for reads, consistency. AWS partitions and data distribution · AWS partition-key guidance

  • Burst capacity lets DynamoDB use some unused throughput for brief spikes. AWS documents retaining up to five minutes of unused read and write capacity, some of which DynamoDB can spend on background work.
  • Adaptive capacity automatically lets a disproportionately busy partition use more of the available capacity, up to the applicable table and partition limits.
  • Split for heat divides a busy partition when its sustained load can be distributed across two key ranges. Without a local secondary index, DynamoDB can split an item collection at a sort-key boundary. With a local secondary index, the split has to fall between item collections.

One item always belongs to one physical partition at a time. A tenant that writes exclusively to ever-increasing timestamp sort keys can also stay hot because new writes would continue to land on the newest side of a split. If a busy tenant repeatedly exceeds the capacity of its active range, see the guide to diagnosing DynamoDB hot partitions. AWS burst and adaptive capacity · AWS split-for-heat explanation

How secondary indexes fit into the write path

A global secondary index (GSI) has its own key schema and physical partitions. DynamoDB maintains it asynchronously from base-table changes. A successful table write may take time to appear in a GSI query, and GSI reads are eventually consistent. If the index runs short of write capacity, including on one hot index partition, DynamoDB can throttle base-table writes to keep the index from falling further behind. Design the distribution of your GSI keys independently of the table keys. AWS global secondary indexes · AWS GSI back-pressure

A local secondary index (LSI) shares the base table's partition key, physical partition placement, and throughput budget. It provides an alternate sort key and supports strong reads. Each item collection has a 10 GB limit that includes its base table items and all related LSI entries. For the audit log, an LSI can serve another ordering within a tenant, provided that the tenant's collection fits that limit. AWS local secondary indexes · AWS burst and adaptive capacity

Where Streams, global tables, backups, and DAX fit

FeatureWhere it sits and what the application sees
DynamoDB StreamsAfter item changes, an enabled stream exposes change records to consumers such as Lambda. Ordering is per item and retention is up to 24 hours. AWS Streams documentation
Global tablesReplicate between AWS Regions. Multi-Region eventual consistency (MREC) propagates writes asynchronously. Multi-Region strong consistency (MRSC) replicates a write to at least one other Region before acknowledging it. Strong reads on any MRSC replica return the latest item version. AWS global tables documentation
Backups and point-in-time recovery (PITR)On-demand backups capture a table snapshot. When enabled, PITR continuously backs up table data and lets you restore to a new table within the configured recovery window. Use this to recover from changes such as an accidental delete that would also reach live replicas. AWS backup and restore documentation
DynamoDB Accelerator (DAX)A separate, optional cache between your application and DynamoDB. It can serve cache hits for eventual reads and forwards strong reads to DynamoDB without caching their results. AWS DAX consistency documentation

What the architecture means for table design

Choose keys by active traffic. A tenant key spreads writes when many tenants are active. Test the busiest tenant separately. If it can overwhelm an active range, a calculated shard suffix based on event ID gives DynamoDB more partition key values to distribute. The tradeoff is that reading a complete tenant timeline now requires querying the shards and merging their results. AWS write sharding

Calculate capacity from item size. For a nontransactional single-item write, DynamoDB rounds up to 1 KB increments. A 3 KB event needs three write units, so a partition with a 1,000-unit write budget can sustain about 333 such writes per second before other traffic or capacity constraints. A strongly consistent GetItem rounds up to 4 KB increments. Returning just a few attributes does not reduce the capacity charge for reading that item. Store large, rarely used event details separately if the common read path only needs a summary. AWS read and write operations

Put frequent lookups into key conditions. If operators often need failed events for one tenant in a time range, filtering the tenant's timeline by outcome still consumes read capacity for the events evaluated before filtering. A GSI keyed by tenant and outcome, with event time in its sort key, lets the query target that range directly. That saves read work at the cost of index writes and eventual visibility. Choose the base-table key or a suitable LSI for a lookup that must see a newly written result immediately. AWS Query filters · AWS global secondary indexes · AWS read consistency

Frequently Asked Questions

How does DynamoDB find an item?

It hashes the partition key and uses routing metadata to find the responsible partition. With a composite primary key, the sort key identifies the item within its collection and can also identify the range when that collection spans partitions.

Does a DynamoDB table have one write leader?

Each physical partition has its own leader in the regional architecture described in the DynamoDB paper. That leader orders writes and acknowledges them after a quorum of replicas has persisted the log entry.

Which replica serves a strongly consistent DynamoDB read?

In the regional architecture described in the DynamoDB paper, the partition leader serves strong reads. Any storage replica in that partition's group can serve an eventually consistent read.

Can one DynamoDB partition key span multiple physical partitions?

Yes. On a table without a local secondary index, DynamoDB can split the collection of items that share a partition key at a sort-key boundary. An individual item stays on one physical partition at a time.

Why is an item visible in the table before it appears in a GSI?

DynamoDB propagates changes to global secondary indexes asynchronously. If you know the full base-table key, a strongly consistent read on the table can see a successful local write before it reaches the index.

Dynobase is a Professional GUI Client for DynamoDB

Start your 7-day free trial today

Product Features

Download
/
Changelog
/
Pricing
/
Member Portal
/
Privacy
/
EULA
/
Twitter
© 2026 Dynobase