Multimodal Routing Architecture

This repository provides a high-level public overview of a high-performance multimodal routing platform built in Go.

At its core, the system includes an in-memory graph database optimized for multimodal route search, capable of computing realistic itineraries across flights, trains, suburban trains, buses, ferries, and in-city transfers.
The platform exposes this capability as a white-label B2B/B2C solution for travel companies.

This document describes the architecture and engineering approach only — it contains no production code, datasets, or client-specific logic.

Problem

Building a practical multimodal routing engine at scale requires:

• combining multiple transport modes (air/rail/bus/ferry/local transfers)
• handling schedules for millions of flights, trains, and buses updated in near real time
• respecting transfer rules and minimum connection windows
• supporting millions of combinations of business preferences to prioritize or suppress directions, nodes, and routes
• delivering low-latency responses suitable for B2B/B2C user-facing applications

While several research algorithms exist (e.g., RAPTOR, CSA and related approaches), none of them solve real-world multimodal routing "as is". Practical deployments require substantial adaptation, hybrid strategies, and domain-specific optimizations to handle time-dependent schedules, multiple transport modes, transfers, and business constraints at global scale.

Core Solution

1. Multimodal Search Engine (Go)

The routing engine:

• stores the full routing graph in memory
• represents segments for flights, trains, buses, ferries, and local transfers
• runs best-K path search with RAM- and CPU-optimized algorithms
• supports flexible business rules for ranking and filtering routes

Key implementation techniques:

• stateful services with pre-loaded dictionaries, configuration, and the entire graph
• sharded graph representation to split the workload
• bitmasks and direct references instead of hashmaps where possible
• reusable structures via memory pools on top of arrays
• multi-level caching: RAM, local (sidecar) Redis, remote Redis
• immutable dictionaries for repeating parts of replies
• compact request/response packets over bidirectional streams to reduce network overhead

These techniques were combined to maintain low latency and reasonable memory usage over a very large dataset.

2. White-Label Multimodal Platform

Around the routing core, the system exposes a full-featured white-label platform enabling:

• airlines, railway operators, bus companies, and travel agencies to offer multimodal connectivity
• configuration of business preferences to promote or suppress specific routes, nodes, or directions
• detailed and aggregated statistics for each request and result
• ML-based components used to estimate optimal cache TTLs for ticket availability and pricing data

Technology Stack

Core services and APIs:

• Go for all backend services
• Protobuf + gRPC for internal communications
• Streaming REST API for front-end and integration clients

Data, storage, and messaging:

• Postgres
• Mongo (after an initial phase with Cassandra)
• Redis
• ClickHouse (analytics and ML data)
• RabbitMQ initially, later migrated to Kafka

DevOps and operations:

• CI/CD automation
• Full suite of load tests, unit tests, and custom-DSL-based integration tests
• Docker and Kubernetes (cloud and bare metal), supporting multiple fully isolated environments
• Ruby on Rails for the admin panel (later migrated to ReactJS)
• Prometheus + Grafana for metrics, ELK stack for structured logs
• Python with TensorFlow and SciKit-Learn for ML pipelines

Additional Context

• Selected Engineering Challenges

Contact

For collaboration or technical inquiries:

• Email: konstantin@disciplinedware.com
• Maintainer: Konstantin Trunin — https://www.linkedin.com/in/ktrunin