Skip to main content

Dataflow

How bytes actually move: commits to running systems, requests to services, metrics to dashboards, prompts to tokens.

Git → runtime (reconciliation)​

Each path is pull-based except the edge: hosts and the cluster poll Git, so a partitioned node converges the moment it reconnects. The edge is push-triggered but immutable per deployment.

User request paths​

Public web (anon):

  • Static-heavy apps (www, links, docs) serve from the ASSETS binding; the Worker is a thin shell.
  • fade does its heavy lifting client-side: files never leave the machine — parsing (EXIF/RAF), demosaic, and rendering run in WebGPU shaders and Web Workers (darkroom package).
  • reiya is the only main-branch app with a database (D1 reiya), reached through drizzle-orm.

Tailnet service (operator):

Agent mesh (A2A):

Media pipeline (servarr)​

  • Indexing: prowlarr feeds the *arr apps tracker results.
  • Acquisition: qbittorrent downloads; xqbit (qbittorrent-cleanup) reaps missingFiles and resumes recoverable errored torrents.
  • Organization: the *arr apps import, rename, and manage the library; bazarr/metatube/maintainerr attach subtitles, metadata, and cleanup.
  • Consumption: jellyfin streams (transcoding on the x86 nodes); immich serves photos with ML tagging from its own model pod.

Inference path (prompt → tokens)​

Model selection is the x-ai-eg-model header; each route is a backend chain with remote-provider fallback tiers. The binding constraint is the Halo unified GTT pool shared between weights and KV cache — full story in AI infrastructure.

Observability path (signal → dashboard)​

Secret flow (sops)​

Secrets are never stored plaintext anywhere except the target's memory at activation time; rotation = re-encrypt + commit + reconcile.