[AWS Certificate]-DynamoDB Accelerator (DAX)

DAX(DynamoDB Accelerator)란

 

• In-Memory Caching, microsecond latency
• Sits between DynamoDB and Client Application (acts a proxy)
• Saves costs due to reduced read load on DynamoDB
• Helps prevent hot partitions
• Minimal code changes required to add DAX to your existing DynamoDB app
• Supports only eventual consistency (strong consistency requests pass-through to DynamoDB)
• Not for write-heavy applications
• Runs inside the VPC
• Multi AZ (3 nodes minimum recommended for production)
• Secure (Encryption at rest with KMS, VPC, IAM, CloudTrail ...)

 

---

 

DAX Architecture

• DAX has two types of caches (internally)
  • Item Cache
  • Query Cache
• Item cache stores results of index reads (=GetItem and BatchGetItem)
  • Default TTL of 5 min (specified while creating DAX cluster)
  • When cache becomes full, older and less popular items get removed
• Query cache stores results of Query and Scan operations 
  • Default TTL of 5 min
  • Updates to the Item cache or to the underlying DynamoDB table do not invalidate the query cache. So, TTL value of the query cache should be chosen accordingly.

---

DAX Operations

 

 

• Only for item level operations
• Table level operations must be sent directly to DynamoDB

 

---

 

DAX Operations

 

• Only for item level operations
• Table level operations must be sent directly to DynamoDB
• Write Operations use write-through approach
• Data is first written to DynamoDB and then to DAX, snd write operation is considered as successful only if both writes are successful
• You can use write-around approach to bypass DAX, e.g. for writing large amount of data, you can write directly to DynamoDB (Item cache goes out of sync)

 

---

 

DAX Operations

• Only for item level operations
• Table level operations must be sent directly to DynamoDB
• Write Operations use write-through approach
• Data is first written to DynamoDB and then to DAX, and write operation is considered as successful only if both writes are successful
• You can use write-around approach to bypass DAX, e.g. for writing large amount of data, you can write directly to DynamoDB (Item cache goes out of sync)
• For reads, if DAX has the data (=Cache hit), it's simply returned without going through DynamoDB

 

---

DAX Operations

 

• Only for item level operations
• Table level operations must be sent directly to DynamoDB
• Write Operations use write-through approach
• Data is first written to DynamoDB and then to DAX, and write operation is considered as successful only if both writes are successful
• You can use write-around approach to bypass DAX, e.g. for writing large amount of data, you can write directly to DynamoDB (Item cache goes out of sync)
• For reads, if DAX has the data (=Cache hit), it's simply returned without going through DynamoDB
• If DAX doesn't have the data (=Cache miss), it's returned from DynamoDB and updated in DAX on the master node

---

DAX Operations

• Only for item level operations
• Table level operations must be sent directly to DynamoDB
• Write Operations use write-thorugh appraoch
• Data is first written to DynamoDB and then to DAX, and write operation is consider as successful only if both writes are successful
• You can use write-around approach to bypass DAX, e.g. for writing large amount of data, you can w rite directly to DynamoDB (Item cache goes out of sync)
• For reads, if DAX has the data (=Cache hit), it's simply returned without going through DynamoDB
• If DAX doesn't have the d ata (=Cache miss), it's returned from DynamoDB and updated in DAX on the master node
• Strongly consistent reads are served directly from DynamoDB and will not be updated in DAX

---

 

DynamoDB - DAX vs ElastiCache

 

 

---

 

Implementing DAX

 

• To implement DAX, we create a DAX Cluster
• DAX Cluster consists of one or more nodes (up to 10 nodes per cluster)
• Each node is an instance of DAX
• One node is the master node or primary node 
• Remaining nodes act as read replicas
• DAX internally handles load balancing between these nodes
• 3 nodes minimum recommended for production

---

Demo

---