Architecture Overview

Overview

Roomzin is a distributed in-memory inventory engine designed for booking workloads.

The application uses one logical inventory endpoint. Routing, shard topology, replication, and node failures are handled internally.

Why Booking Inventory Is Different

Booking queries combine availability, dates, geography, room types, pricing, amenities, and policies. Roomzin uses a fixed schema and specialized in-memory execution to make these queries predictable.

Platform Architecture

Request Flow

• SDK / RzProxy — client access
• Edge Router — routes requests to the appropriate zone
• Zone Router — routes requests to the appropriate shard
• RzBridge — routes within the shard
• Roomzin — executes the query

Applications do not need to know the active zone, shard, node, or leader.

Access Layer

SDKs

Supported SDKs:

• Java
• Go
• Python
• C#
• Node.js

Characteristics:

• Native binary protocol
• Connection pooling
• Asynchronous request/response

RzProxy

HTTP/JSON interface for legacy or non-native clients.

• Independently deployable
• Horizontally scalable
• Prometheus metrics
• HTTP/2 API

Routing Layer

RzRouter

One binary with two operating modes.

Edge Mode

• Client-facing entry point
• Routes requests to the appropriate zone
• Uses synchronized routing tables

Zone Mode

• Runs inside a zone
• Routes requests to the appropriate shard
• Uses RzBridge for shard-level routing

Routing occurs directly in the request path. No external proxy is required.

Bridge Layer

RzBridge

Shard-aware connector between routers and Roomzin nodes.

RzBridge knows:

• Nodes belonging to the shard
• Current shard leader
• Available followers

Routing:

RzBridge isolates shard topology from the routing layer.

Leader changes and node failures do not require application-level topology changes.

Storage Layer

Roomzin

Roomzin is an in-memory inventory engine.

The hot booking window is stored in memory and organized into replicated shards.

The query engine is optimized for:

• Geography
• Dates
• Availability
• Room type
• Price
• Amenities
• Policies

Each shard contains the complete dataset for that shard.

Query Execution

Fixed Schema

Booking attributes use fixed-size representations.

Attributes are represented as a bitmasks and numbers.

Benefits:

• Bitwise filtering
• SIMD-friendly execution
• No string comparisons
• No runtime hash lookups
• Compact memory layout
• Cache-friendly access

Maximum: 24 boolean attributes per each daily package.

Numeric Encoding

Human-readable values are converted to numeric identifiers during ingestion.

The query path operates on integers.

There is no runtime parsing or string interning.

Geospatial Indexing

Roomzin uses H3 for geographic indexing.

• Hierarchical hexagonal grid
• Numeric cell identifiers
• Radius/area filtering
• Variable geographic resolution
• No tree traversal in the query path

Segment-Based Execution

Inventory is partitioned into independent segments.

Properties:

• Independent processing queues
• No cross-segment locking
• No cross-segment synchronization
• Large segments can be split across cores
• CPU/core pinning for cache locality

Segments execute in parallel and results are combined after processing.

Request / Response Pipeline

Roomzin uses independent request and response streams.

Request stream

Clients can send multiple requests without waiting for previous responses.

Response stream

Responses are returned asynchronously and may arrive out of order.

Each request contains a correlation ID.

This avoids connection-level head-of-line blocking.

Replication

Each shard is a Raft cluster.

Leader

• Handles writes
• Replicates state through Raft

Followers

• Hold the complete shard dataset
• Serve reads

Full Replication

Every node in a shard contains the complete shard dataset.

Therefore:

• Reads do not require cross-node joins
• Any follower can serve any read
• Read capacity increases by adding followers
• Shard queries require no distributed coordination

Persistence

Roomzin uses lock-free snapshots for persistence.

Implementation uses Rust atomic references and copy-on-write.

Current characteristics:

Snapshot creation does not block query execution.

Global Distribution

Roomzin distributes inventory through zones and shards.

Routing tables determine:

Segment → Zone → Shard → Node

The application only sees the logical Roomzin service.

Control Plane

RzID

Lightweight service registry.

Responsibilities:

• Component registration
• Heartbeats
• Topology management
• Segment → shard ownership

RzID is not on the request hot path.

RzPoint

Customer-provided HTTP service for resolving logical service identities to hostnames.

Works with:

• VMs
• Kubernetes
• Cloud infrastructure
• Custom infrastructure

RzPoint separates logical service identity from physical deployment.

Network Protocols

Core Technologies

• Rust — Roomzin implementation
• TiKV Raft — consensus and replication
• Tokio — asynchronous runtime
• QUIC — internal node-to-node transport
• H3 — geospatial indexing

Architectural Constraints

Roomzin intentionally limits the query model.

• Maximum 24 attributes for room policy and rates
• Fixed schema
• Numeric-only query execution
• Full replication within a shard
• No distributed joins within a shard
• Hot booking window held in memory

These constraints are part of the execution model rather than configuration options.

Architecture Summary

Key properties:

• Fixed-size data representation
• In-memory execution
• Parallel segment processing
• Full shard replication
• Raft-based consistency
• Follower read scaling
• Topology-aware routing
• Lock-free snapshots
• Asynchronous request/response
• Custom binary protocol