Web

The web app ships inside the server binary and serves at its root, :8080 by default. It is a client of the ConnectRPC API and nothing else. Every button maps to an RPC you can call yourself, so anything you do here is scriptable later. Three objects make up the model, and the navbar follows them:

• A connection is a configured connector. Credentials plus connection-scoped settings, saved once and reused.
• A pipeline is a graph of connections. Edges route resources from a source to a sink, each carrying a read mode and a write mode.
• A run is one execution of a pipeline route, tracked from request to verified write.

​

Observability

The landing page. Records and bytes moved, run durations, and success rate over a timeframe you pick, plus a feed of recent runs across every pipeline. The charts come from the metrics service, which aggregates the same run rows the rest of the app reads.

​

Connections

Sources and Sinks each list what you have already connected. A card carries the connector it uses, how many pipelines depend on it, and its current version, which is the optimistic lock that keeps concurrent edits from overwriting each other. New source opens the connector catalog. It shows the connectors compiled into the running server and labels each one alpha, beta, or stable. Connecting one generates a form from that connector’s config schema. The connector declares its fields, types, defaults, and visibility rules, and the form renders exactly those. Postgres CDC settings, for example, appear only once you set replication to cdc. Two details worth knowing:

• Scope. Connections hold connection-scoped fields only, the ones that describe reaching the system. Pipeline-scoped fields like a destination schema are set later, per pipeline.
• Secrets. Secret-typed fields are extracted on save and stored through the configured secrets provider. Reads return them masked, and editing leaves a stored secret in place unless you submit a new value.

Before saving, the server checks required fields, value types, and other local configuration rules. It does not make a live connection at this point, so an unreachable hostname or rejected password can still fail during discovery or a run.

​

Pipelines

The list view shows each pipeline with its schedule, recent run outcomes, and last duration. Opening one gives you the canvas, the history, and the settings.

​

Canvas

The canvas edits the pipeline graph. Nodes are connections and edges route selected resources between them. Standard sources let you choose full or incremental reads and an appropriate destination write mode. CDC sources use the connection’s change stream and let the route append events or merge them into current state. Incremental reads may also need a cursor column. Replication modes explains the valid combinations. Edits are validated as you work. Unmet requirements appear on the edge itself, so a table without a primary key blocks an upsert route before you save rather than failing at run time. Saving creates a new immutable pipeline version. In-flight runs keep executing against the version they started with.

​

History

Every run, newest first, with per-resource record and byte tallies and a final status. An active run streams, so progress arrives as the engine publishes it rather than on a poll. Statuses map directly to engine states. A failed run that saved resumable progress enters PARTIAL and continues from its last checkpoint when requested again. See Runs and recovery.

​

Settings

Name and description, a cron schedule with a timezone and an overlap policy, and the worker configuration that sizes the machines runs execute on. Metadata edits apply in place and do not create a version.

​

Contributing to the web app

For contributors, the web app lives in ui/ as a React 18 and TypeScript app built with Vite:

• Routing and data. TanStack Router with file-based routes in src/routes/, and TanStack Query for server state. API calls go through ConnectRPC’s web transport with connect-query bindings.
• Generated types. src/gen/ holds the protobuf-generated TypeScript. Never edit it by hand. pnpm codegen (or just gen at the repo root) regenerates it with buf after a proto change.
• Design system. Components, tokens, and theming come from @galaxy-io/dls. The app mounts its stylesheet and GalaxyThemeProvider in main.tsx. Browse the catalog at storybook.getgalaxy.io before building anything new.
• Styling and the canvas. Component-local styles use Linaria. The canvas is built on @xyflow/react. Icons are Phosphor, with project SVGs compiled by pnpm assets.
• Formatting. Biome formats and lints the UI.

The server serves the compiled bundle only when built with the embedui tag, which just binaries and the published images do. For development, just ui runs Vite on :5173 against a local API. See Local development for the rest of the workflow.