clusterflux-public/docs/architecture.md
Clusterflux release a6ca9e26dd Public release release-be1f654ae8de
Source commit: be1f654ae8de7d5eeb2005b4e1a05bf1ebf3124f

Public tree identity: sha256:ab53679496be4be7d1e02bc00723072774d1a3f87e10cc56558a752f28f6abb1
2026-07-17 05:17:01 +02:00

57 lines
2.4 KiB
Markdown

# Architecture
Clusterflux separates a control plane from execution nodes.
## Coordinator
The coordinator owns identity scope, one-active-process admission, task
placement, attempt history, Debug Epoch coordination, VFS metadata, artifact
metadata, quotas, and secure relay reservations. Its async main may park and wake
without consuming node compute.
The coordinator does not execute arbitrary native commands or hosted
containers. It also does not become a durable artifact store merely because it
coordinates a download.
## Nodes
A node is an enrolled public-key identity. It reports capabilities and periodic
heartbeats. The coordinator derives whether it is live from the last accepted
heartbeat; a client-supplied "online" field is not placement authority.
Nodes resolve bundle environments, execute Wasm tasks, run native commands, and
retain artifact bytes. A node must prove the tenant, project, process, task, and
key scope on every signed request.
## Virtual process and tasks
A Coordinator Project admits one active virtual process. The process contains:
- one coordinator-hosted async main;
- logical task instances created by that main;
- one or more attempts for each logical task;
- joins tied to the logical task, not to an individual attempt.
"FailFast" makes a terminal failure visible to the join immediately.
"AwaitOperator" keeps the join pending while an operator accepts, cancels, or
restarts the failed task. Restart creates a distinct attempt but preserves the
logical task identity.
A terminal process releases the active slot automatically. A task parked for an
operator is not terminal and continues to occupy the slot.
## Durable and live state
Projects, identities, credentials, permissions, hosted policy records, relay
period usage, and relay reservation accounting may be stored in Postgres.
Active processes, task placement, Debug Epochs, transient VFS state, relay byte
streams, and live node state are bounded in memory and are not reconstructed as
running after a coordinator restart. Pre-restart relay reservations are
reconciled as abandoned before new links can be issued.
## Protocol lanes
Human clients use scoped sessions. Agents and nodes use separate signed
public-key identities. Operator actions use a separate administrative
credential. Authority comes from the authenticated lane and server-side scope,
never from identity fields in an untrusted request body.