Summary: Dynamo

Summary

Dynamo is a highly available key-value store that Amazon built for its own internal services. It shows how a business requirement can shape a system design: Amazon chose higher availability over strong consistency, because that is what its business needed. The design also assumes that system and hardware failures happen, not that they might happen.

Dynamo is a peer-to-peer system, with no leader and no follower nodes. Every node is equal, with the same roles and responsibilities. That symmetry is also what removes any single point of failure.

Data gets spread across the cluster using consistent hashing, automatically and without a central authority. For fault tolerance, every write also replicates to other nodes. Dynamo does that through a sloppy quorum rather than a strict majority quorum, which keeps writes available even when some usual nodes are down.

Two more mechanisms round out the reliability story. Merkle trees, hash trees used to compare two replicas branch by branch, run anti-entropy in the background.

That catches and repairs drift between replicas without waiting for a read. Different storage engines can also be plugged into a node's local storage.

Nodes learn about each other through gossip protocol, exchanging state peer to peer instead of relying on a central directory. Dynamo also stays "always writeable" through hinted handoff: another node can accept a write temporarily when the one that should hold it is down.

That commitment to always allowing writes has a cost. Because different servers can accept writes to the same object at the same time, that object can end up with multiple versions. Dynamo tracks those versions with vector clocks and tries to reconcile them automatically.

When it cannot, it hands both versions to the client application, on the idea that the client understands the object well enough to merge them.

Pulling consistent hashing, peer-to-peer communication, gossip, vector clocks, and quorums into one working system is Dynamo's real achievement. And because Amazon built it strictly for internal use, security was never part of the design.

The table below summarizes the techniques Dynamo uses and what each one provides.

<table class="tg" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; border-spacing: 0px; border-color: black; --tw-border-opacity:1; margin: 2em auto; table-layout: fixed; width: 800px; color: rgb(61, 61, 78); font-family: &quot;Droid Serif&quot;, Georgia, serif; font-size: 18px;"> <tbody style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <tr style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <td class="tg-1yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Problem</td> <td class="tg-1yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Technique</td> <td class="tg-1yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Advantage</td> </tr> <tr style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <td class="tg-1rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Partitioning</td> <td class="tg-rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Consistent Hashing</td> <td class="tg-rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Incremental Scalability</td> </tr> <tr style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <td class="tg-1yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">High availability for writes</td> <td class="tg-yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Vector clocks with reconciliation during reads</td> <td class="tg-yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Version size is decoupled from update rates.</td> </tr> <tr style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <td class="tg-1rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Handling temporary failures</td> <td class="tg-rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Sloppy Quorum and Hinted Handoff</td> <td class="tg-rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Provides high availability and durability guarantee when some of the replicas are not available</td> </tr> <tr style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <td class="tg-1yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Recovering from permanent failures</td> <td class="tg-yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Anti-entropy using Merkle trees</td> <td class="tg-yw4l" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(103, 171, 159); vertical-align: top;">Synchronizes divergent replicas on the background</td> </tr> <tr style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent;"> <td class="tg-1rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; font-weight: bold; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Membership and failure detection</td> <td class="tg-rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Gossip-based membership protocol and failure detection</td> <td class="tg-rmb8" style="--tw-shadow:0 0 transparent; --tw-ring-inset:var(--tw-empty, ); --tw-ring-offset-width:0px; --tw-ring-offset-color:#fff; --tw-ring-color:rgba(59,130,246,0.5); --tw-ring-offset-shadow:0 0 transparent; --tw-ring-shadow:0 0 transparent; padding: 10px 5px; border-width: 1px; border-style: solid; border-color: black; --tw-border-opacity:1; word-break: normal; hyphens: auto; font-family: Arial, sans-serif; font-size: 17px; overflow: hidden; color: black; background-color: rgb(197, 214, 196); vertical-align: top;">Preserves symmetry and avoid centralized monitoring</td> </tr> </tbody> </table>

System design patterns

Here is a quick reference for the patterns behind Dynamo's design:

• Consistent hashing. Dynamo uses it to spread data across nodes.
• Quorum. A write only counts as successful once it reaches a set number of replica nodes, which keeps the data consistent.
• Gossip protocol. Each node uses it to keep track of state information about the other nodes in the cluster.
• Hinted handoff. Nodes use it to remember a write meant for a node that is currently down.
• Read repair. Dynamo uses it to push the latest version of the data to nodes still holding an older one.
• Vector clocks. Dynamo uses them to reconcile concurrent updates on the same object.
• Merkle trees. Dynamo uses them for anti-entropy, resolving conflicts in the background.

References and further reading

• Amazon's Dynamo
• Eventually Consistent
• Bigtable
• DynamoDB
• CRDT
• A Decade of Dynamo
• Riak
• Dynamo Architecture
• Dynamo: A flawed architecture