Skip to main content
The execution store keeps:
  • Each session’s background jobs, the ones cmd & starts, and each job’s output.
  • The processes the workspace runs, which ps lists.
  • Which instance of the workspace runs them, and when it was last seen.
The group key is store.execution. A job’s output is a stream other processes follow, so disk is refused here, and a disk store keeps its jobs in memory. In code, SqlWorkspaceStateStore keeps them in a SQL database, SQLite or Postgres, in two tables it makes on first use, so every process on the database lists the same jobs, lines and processes, and a restart on it keeps them. A reader following a job on another process there sees new output within a quarter of a second.

On Redis

Here the rest of the state stays on the machine’s disk, and every process on the machine that builds the workspace sees the same jobs.

Several instances of one workspace

Every process that builds the workspace under the same id on one Redis is an instance of it, on one machine or several, behind a load balancer or as replicas of one agent.
workspace.yaml
On the servers, workspace.json holds {"id": "agent", "config": {...}} with the same config on each. Then:
  • One instance runs the workspace’s lines at a time: the first to run one. Any other is refused with an error naming that instance, which a server answers with 409, until it stops, deletes or closes the workspace.
  • If it goes away without doing so, the workspace is free 30 seconds after it was last seen. The next instance to run a line reports the jobs it left as ended with status 129 and stops the programs they left running.
  • A server keeps its id under its state root, so a restarted server runs its workspaces again at once.
  • Sessions and symlinks are shared, since the whole store is on Redis.
A session sees only its own processes; a profile with processes: workspace widens that to every session’s. Shell describes the details.