AWS DynamoDB (Design Patterns)

Write Sharding (Table)

• append known random value to the key
  • e.g. CandidateId + rand(0,n)
• Scatter (write) + Gather (read)
• Use case
  • Low key cardinality (e.g. voting application for 2 people)

Write Sharding (GSI)

• create additional attribute "GSIKey" - rand(0,n)
• create GSI
  • Primary Key = GSIKey
  • Sort Key = timestamp
• Use case
  • Avoids full table scan on "timestamp"

Time Series Data

• Partition by time (e.g. April_2015, March_2015)
• Precreate tables
• Writes go to current table
• Lower RCU/WCU as tables age-out (no traffic to them)

Sparse Index

• Comparable to  "FilteredIndex" in SQL Server
• LSI
  • Items with no SortKey attribute are not indexed
  • Example
    
    • Table: CustomerId, OrderId, OrderOpenDate
    • LSI: CustomerId, OrderOpenDate
      • When the order is closed DELETE the "OrderOpenDate" attribute
• GSI
  • Items with no index key value are not indexed
  • Example
    • Table: UserId, Champ attribute (for some items)
    • GSI: PK = Champ, SortKey = UserId

Geo-Hashing

• Split the map into big squares (1,2,3,4)
  • Each square contains smaller squares (1:11,12,13,14)
    • Up to desired level of granularity
• Use SortKey to do comparison (e.g. "show me neighbours")
• Client library available

Query Filters

• Filter may be expensive (less data "on the wire" but reads still happen)
  • RCU consumed
• Concatenate attributes to form useful composite range key
  • e.g. Status + date: "PENDING_2015-12-25"
  • Query "begins with" PENDING
• Use cases
  • filtering data

Vertical Partitioning

• Split item into multiple to limit RCU consumption
• Use GSI to model M:N relationship
• Use case
  • Message bodies  separated out to a different table

Adjacency Lists

• Partition on nodeId
• Model relationships as edges
  • Default (self-edge) to describe node itself
• Use Sharded GSI to query on "data"

References

• https://www.youtube.com/watch?v=jzeKPKpucS0