ClickHouse
AnalyticsColumnar OLAP for per-workspace metrics and dashboards — without impacting transactional PostgreSQL.
Modular monolith. 29 domain modules. One compiled binary that completes 6× the taxi rides of the Node.js stack it replaced, on half the memory.
Benchmarks
Metric
Node.js (previous)
Rust (BetterSuite)
Taxi Rides per Minute
300
1,800
Throughput (RPS)
~8,850
~27,550
Avg. Server Latency
~46 ms
~11 ms
p99 Latency
131 ms
51 ms
Idle Memory
~400 MB
~196 MB
Peak Memory Under Load
~415 MB
~220 MB
Node.js (previous)
Rust (BetterSuite)
Taxi Rides per Minute
300
1,800
Throughput (RPS)
~8,850
~27,550
Avg. Server Latency
~46 ms
~11 ms
p99 Latency
131 ms
51 ms
Idle Memory
~400 MB
~196 MB
Peak Memory Under Load
~415 MB
~220 MB
Taxi rides, step by step
Same rides, hardware and limits as the table. The Node.js points at 300 and 360 come from a second run on a fresh stack, which is why its lines dip at 300; at 360 its process ran out of its 2 GB and was killed.
* Measured 7–8 November 2026 on a Hetzner CCX33, each backend alone on the same 2 dedicated vCPUs and 2 GB of RAM: Ridy 5.3.2 (3270d49) on Node.js 26.13.0, run as a single process, and BetterSuite on Rust 1.99.1. Taxi rides per minute: complete rides from quote to drop-off, sent the way the apps send them, with 2,000 drivers and 5,000 riders in London. The arrival rate rose every three minutes, and the figure is the highest rate at which quote, booking, acceptance and drop-off stayed under 300 ms at p95 and 99% of rides finished. Both drew routes as straight lines: BetterSuite (cfd25ce87) with its built-in estimate, Ridy through a local stand-in for Google's Routes API. With road routing on its own Valhalla server, BetterSuite passed 1,650. Other rows: 85% minimal GraphQL requests and 15% health checks, up to 1,000 concurrent clients, measuring per-request server cost rather than database time (BetterSuite 5f265364b).
Architecture
Ports-and-adapters design keeping business logic decoupled from infrastructure.
Queries, mutations, and subscriptions. Clients fetch exactly what they need.
Workspace scope required by the type system and enforced on every query. Defense in depth.
One compiled binary, one process. No service mesh, no per-service deploys.
Zero serialization, zero network latency between domain modules.
New workspaces live in seconds. Schema, scoping, config — all automated.
Ports-and-adapters design keeping business logic decoupled from infrastructure.
Queries, mutations, and subscriptions. Clients fetch exactly what they need.
One compiled binary, one process. No service mesh, no per-service deploys.
Workspace scope required by the type system and enforced on every query. Defense in depth.
Zero serialization, zero network latency between domain modules.
New workspaces live in seconds. Schema, scoping, config — all automated.
Dispatch
The backend infrastructure powering real-time dispatch, location matching, and concurrent workloads at scale — built on H3 hexagonal indexing over Redis sorted sets.
One booking, traced from a warm index to the last round before the deadline. Every value is the platform default each workspace starts from.
Every driver location ping writes ten H3 cell IDs at once, resolutions 0 through 9. Geography is never computed at request time — the reverse index is already sitting in Redis at whatever scale the search turns out to need.
39 drivers online · 390 cell memberships · resolutions 0–9, rewritten on every ping
Configurable
Sequential is one of two strategies, shown here on the shipped defaults. Broadcast offers a whole wave of drivers at once, and is the default fallback when the primary search comes up empty; reach, expansion, offer window, retry budget and the scoring weights are all set per workspace.
Toolchain
Every layer of the stack is purpose-selected — type-safe database access, background job processing, analytics at scale, and full distributed tracing.
Columnar OLAP for per-workspace metrics and dashboards — without impacting transactional PostgreSQL.
Compile-time verified SQL. Type-checked against the live schema at build time — zero ORM overhead.
Rust-native background processing with persistent storage, retries, and scheduling.
End-to-end distributed tracing across all 29 modules — GraphQL to DB to response.
Frontend
0
% Code Reuse
0
UI Components
0
Locales
0
FPS Render
Built on atomic design principles — from primitive atoms to complete production screens.
Atoms
Inputs, buttons, labels, icons
Molecules
Search bars, cards, list tiles
Organisms
Navigation, forms, data tables
Templates
Layouts, scaffolds, complete screens
Real-time
Sub-second location updates via GraphQL subscriptions. No polling.
Every state change streams instantly to all connected clients via WebSocket.
APNs + FCM + Web. Workspace-configurable rules for events and promotions.
Pricing, regions, currencies, delivery rules — configurable live from dashboard.
Sub-second location updates via GraphQL subscriptions. No polling.
Every state change streams instantly to all connected clients via WebSocket.
APNs + FCM + Web. Workspace-configurable rules for events and promotions.
Pricing, regions, currencies, delivery rules — configurable live from dashboard.
Database, workspace scoping, configuration, branding — orchestrated end-to-end without manual intervention.