Why HyperLiquid?
Its dual-chain architecture is the perfect testbed for a fully capable financial data platform.
HyperCore
A real stress test in volume and state: billions of order-book events, with stateful processing challenges like the L4 address book — tracking state per address, per asset.
HyperEVM
The classic EVM pattern, approached in ELT style the way Dune and Glassnode do it: raw chain data in, smart-contract calls parsed, transformations and metric models built on top.
The stack
A pragmatic monorepo: ingestion → Kafka → Iceberg warehouse (Nessie + MinIO) → serving (ClickHouse, dbt, Spark, Trino), orchestrated with Airflow, with Flink for stateful streams.
Ingestion
Hyperliquid node & indexer services, replay-as-tap into Kafka topics.
Streaming
Apache Kafka 3.9, Flink 1.18 for bounded & unbounded stateful processing.
Warehouse
Apache Iceberg on Nessie + MinIO; transformations via dbt 1.12 & Spark 3.5.
Serving
ClickHouse 26.9 projections, Trino queries, Airflow-scheduled freshness.
