Skip to content

Operate the DSPy workspace

Protect endpoint credentials and retain data, programs and measured results together. This operations reference describes the Limited image; the classifier is a CLI task, not a long-running server.

Setup and task lifecycle

Inspect workspace setup if it did not finish:

sudo systemctl status sushi-dspy-firstboot.service --no-pager
sudo journalctl -u sushi-dspy-firstboot.service --no-pager

The classifier exits after each task. There is no application daemon or API port to restart. Stop experiments before a planned reboot. Configuration and completed results persist on the root EBS volume, subject to your storage policy.

Reread saved results

Switch to ubuntu:

sudo -iu ubuntu

Provide an existing saved run directory:

read -r -p 'Full saved run directory: ' dspy_saved_run
/usr/bin/python3 -I -B /opt/dspy/readback.py "$dspy_saved_run" --require-provider

This rereads local files without another provider request. Keep all files from the task together: hashes, saved training demonstrations and predictions are checked as a group. Never overwrite a failed run to make it appear successful. After fixing a failure, choose a new output directory and explicitly rerun.

Backups and security

Back up /home/ubuntu/dspy-workspace/ and /home/ubuntu/dspy-results/ to encrypted storage with restricted readers. Workspace directories are mode 0700; provider configuration is 0600. Task outputs can contain your original ticket data. Termination can delete the root EBS volume, so arrange backups before termination.

Keep /opt/dspy/ and its virtual environment root-owned. Do not solve permission problems by making the runtime writable or installing into it as ubuntu. Use a reviewed replacement image or a separately managed environment for changes.

Support and evaluation

Review provider errors privately; terminal task errors omit request bodies and keys. Share only redacted symptoms with Software Sushi. Technical-support coverage requires a separate agreement. Upstream enterprise services, hosted inference and commercial upstream support are not included.

Assess the measured classifier scores on authorized data before relying on the program. Upstream Alpha metadata remains unchanged. Neither the package nor a completed example establishes production readiness or accuracy on your workload.

Last updated November 16, 2026 · Suggest an improvement