Velocity Trading

Team Members: Tani Diament, Shimmie Hahn, Yisroel Newmark, Matias Salem

Overview

A distributed market maker that consumes exchange data and publishes quotes in real time, executing trades against them across a fault-tolerant cluster.

What makes the system interesting:

• 14-node K3s cluster (3 control plane nodes + 11 worker nodes) running 7 market-maker pods and three replicated upstream services (Exchange, Trading State, Exposure Reservation) each as 1 leader + 2 hot-backup replicas.
• Durable in-memory state — every IMap (positions, fills, quotes, reservations, external-orders) is write-through via a JPA MapStore to a single PostgreSQL StatefulSet, with InitialLoadMode = EAGER so a full cluster restart rebuilds state with no operator intervention.
• ZooKeeper-coordinated failover — Curator LeaderLatch elects the writable replica per service; clients resolve the current leader via an ephemeral /mm/endpoints/{svc} znode, so failover is invisible to callers.

Quickstart (local Docker Compose)

The full stack — 3× Zookeeper, Postgres, 3× Trading State, 3× Exchange, 3× Exposure Reservation, 7× Market Maker, Position UI, external publisher, nginx LB — runs locally via compose.yml.

mvn -DskipTests clean package
docker build -t market-maker .
docker compose up

Once everything reports healthy:

• Position UI — http://localhost:18080
• External publisher — http://localhost:18083
• Market Maker nodes — http://localhost:8081 … http://localhost:8087

For the production K3s deployment (air-gapped image bundling, ZK quorum, Headlamp, integration tests), see docs/installation.md.

Repo layout

src/main/java/edu/yu/velocitytrading/
├── exchange/              # Order matching, fill emission (REST + RSocket)
├── state/                 # TradingStateService — positions/fills owner; state.stream
├── exposurereservation/   # Global exposure budget; reservation lifecycle
├── marketmaker/           # MM node: quote generation, freshness, position tracking
├── cluster/               # Coordinator, symbol-shard assignment, Zk-watched listeners
├── ha/                    # LeaderLatch, LeaderGuardFilter, LeaderAwareRSocketClient
├── persistence/           # JPA entities + Hazelcast MapStores
├── memory/                # HazelcastRepository wrapper around IMap
└── model/                 # Records: Position, Fill, Quote, Reservation, StateSnapshot

docs/                      # design docs (see index below)
k8s/                       # Kustomize manifests for the 14-node K3s deploy
scripts/                   # PowerShell + bash helpers (image bundling, SSH setup, …)
compose.yml                # Local Docker Compose equivalent of the cluster

Project Plan

• Scope (scenarios covered)
• Distributed System Challenges
• Workflow Diagrams
• Error-Case Walkthroughs
• Software Architecture Diagrams
• Tools & Technologies

Installation and Usage Guide

• Installation Instructions
• API and Usage Documentation