Common Workflows

You usually arrive with a goal, not a tool in mind. Start with the story closest to what you're trying to do — each is a complete, end-to-end walkthrough — then branch into the deeper guides. New to spore.host? Do Your first instance first; it's the 15-minute backbone the stories below build on.

Running more than one thing?

spore.host has several ways to run work at scale — instance plugins, sweeps, arrays, instance queues, pipelines, MPI, and workflow-engine adapters. If you're not sure which fits, start with Which execution tool?.

Story A — Interactive research workstation

"I need a big machine for a few hours to explore some data, then it should go away."

truffle find "amd genoa 64gb" --region us-east-1     # 1. find a fit
spawn launch analysis \                              # 2. launch with guards
  --instance-type m8a.4xlarge \
  --ttl 8h \
  --idle-timeout 30m
spawn connect analysis                               # 3. work on it
spawn terminate analysis                             # 4. done — tear it down

• --ttl 8h is the hard deadline — the instance terminates then no matter what, so a forgotten session can't run all weekend.
• --idle-timeout 30m stops it early if you wander off; any activity resets the timer. Idle stops, TTL terminates — see Costs & safety guarantees.
• Cost: capped at 8 × the hourly rate; usually far less because idle stops it. Check the rate with truffle spot m8a.4xlarge before you launch.

→ Deeper: Finding the right instance · Jupyter/RStudio · Managing instances & data

Story B — Unattended batch computation

"Run this script on a big box; when it finishes, terminate — I'm not watching."

spawn launch simulation \
  --instance-type c8a.12xlarge \
  --ttl 12h \
  --command "./run-model.sh && spored complete --status success" \
  --on-complete terminate

What happens in each case:

• Succeeds → the && reaches spored complete, the completion sentinel appears, and --on-complete terminate tears the instance down.
• Fails → run-model.sh exits non-zero, spored complete never runs, so the instance stays up (for you to debug) until the TTL terminates it at 12h.
• Hangs → idle detection won't fire (a running process is activity), but the TTL still terminates at 12h. TTL is the backstop that always wins.

The distinction that trips people up: it's the completion sentinel appearing that triggers on-complete, not your command merely exiting — which is why the script calls spored complete explicitly. See Troubleshooting.

→ Deeper: GPU training jobs · Parameter sweeps · Batch queues

Story C — Scarce GPU capacity

"I need a p5.48xlarge, but there's never one available when I try."

1. Truffle confirms the type exists and you have quota.
2. Spawn tries to launch — and gets InsufficientInstanceCapacity.
3. Lagotto watches three regions every five minutes.
4. A notification arrives the moment capacity appears.
5. Lagotto launches it automatically with a 6-hour TTL — no one awake required.

This is the one workflow the "find → launch" path can't do alone, because quota and capacity are different things. Full walkthrough: Waiting for scarce capacity.

---

Choose what to learn next

I want to…	Go to
Find the right instance type & compare prices	Finding the right instance
Launch from a browser without learning the CLI	Portal detail levels
Run Jupyter or RStudio in the browser	Interactive workstation
Train a model on a GPU	GPU training jobs
Save up to 90% with Spot	Spot instances
Move data on and off instances	Managing instances & data
Run one job across many parameters	Parameter sweeps
Manage a group of related instances	Job arrays
Queue more work than fits at once	Batch queues
Run tightly-coupled multi-node jobs	MPI clusters
Chain jobs so each stage launches the next	Pipelines
Wait for a hard-to-find GPU	Waiting for scarce capacity
Control instances from Slack/Teams	Slack Setup
Drive compute from an AI assistant	AI Assistant (MCP)
Automate from Python	Python SDK
Run Nextflow / WDL / CWL / Snakemake / Airflow	Workflow engines