Question

A payments company is building a new order-settlement platform. Each account's transactions must be processed in strict chronological order, and duplicate transaction submissions from retrying upstream systems must not be double-processed. Throughput is expected to reach roughly 2,000 transactions per second at peak, spread across hundreds of thousands of distinct account IDs. Multiple independent downstream services (fraud scoring, ledger update, notification) must each receive every transaction. Which architecture best meets ordering, deduplication, and fan-out requirements while scaling elastically?

Accepted Answer

Publish transactions to an SNS FIFO topic using the account ID as the message group ID and the transaction ID as the deduplication ID, fanning out to an SQS FIFO queue per downstream service.

Answer

Publish transactions to a standard SNS topic fanning out to standard SQS queues, and have each consumer sort messages by timestamp and check a DynamoDB table for previously seen transaction IDs.

Answer

Write transactions to a single Kinesis Data Stream with one shard and have each downstream service run its own consumer application reading from the beginning of the stream.

Answer

Send transactions to a single SQS FIFO queue shared by all three downstream services, using the account ID as the message group ID and enabling content-based deduplication.