Architecture Overview

Uniflow is a fully serverless CDP deployed to your AWS account. Every component auto-scales and runs at near-zero cost when idle.

Data flow

Client SDK
  └─▶ API Gateway (HTTP API)
        └─▶ Lambda (validate + enrich)
              ├─▶ Kinesis Data Stream ──▶ Firehose ──▶ S3 (raw, GZIP)
              └─▶ Kinesis Data Stream
                    └─▶ Lambda (processor)
                          ├─▶ DynamoDB  (identity graph + profile upsert)
                          └─▶ SQS       (destination fan-out)
                                └─▶ Lambda (connector) ──▶ External system

EventBridge Scheduler (hourly)
  └─▶ Glue PySpark Job (audience-builder)
        └─▶ Athena query over S3 ──▶ DynamoDB (segment membership)

Cognito ──▶ API Gateway ──▶ Lambda (management API)
Next.js static export ──▶ S3 ──▶ CloudFront

Component breakdown

Ingestion pipeline

1. Client SDKs send events (track, identify, page, group) to the ingest endpoint
2. API Gateway HTTP API receives the request with write-key authentication
3. Ingest Lambda validates the event against Zod schemas, enriches with server-side context (IP, timestamp), and writes to Kinesis

Stream processing

4. Kinesis Data Stream buffers events with 7-day retention
5. Firehose delivers raw events to S3 in GZIP-compressed NDJSON format
6. Processor Lambda consumes from Kinesis, performs identity resolution (linking anonymous → known users), upserts profiles in DynamoDB, and fans out to SQS

Destination delivery

7. SQS queues events per destination with dead-letter queues for reliability
8. Connector Lambda executes the destination connector (webhook, S3 export, or custom)

Segmentation

9. EventBridge Scheduler triggers hourly
10. AWS Glue runs the audience-builder PySpark job, which reads S3 via Spark SQL and writes segment membership to both S3 (Parquet) and DynamoDB

Admin

11. Management API Lambda provides CRUD for sources, destinations, segments, and profiles
12. Admin UI is a Next.js static export served via CloudFront

Compute decisions

Component	Compute	Why
Ingest API	Lambda	Zero idle cost, auto-scales
Stream processor	Lambda	Short-lived, stateless
Audience builder	AWS Glue (PySpark)	Serverless Spark for complex segment queries, no VPC needed
Destination connectors	Lambda	Event-driven, short-lived
Management API	Lambda	Low-traffic CRUD