Integrations
desflow runs from the tools people already use. An orchestrator such as Airflow runs a study as one task of a pipeline. An AI assistant gets the skill and the desflow commands from the plugin, which AI Assistants describes.
Pages
| Page | What it does |
|---|---|
| Airflow | a DAG with a BashOperator that runs the shop and a task that reads its result |
| Dagster | an asset that runs the shop through PipesSubprocessClient and keeps the figures as metadata |
| Prefect | a flow with a task that runs the shop and a task that reads its result |
| Kestra | a flow whose Python task installs desflow in a container and runs the shop |
One Shop for Every Orchestrator
The four orchestrator pages run one blueprint, shop.yaml, with one command. The shop takes an order every 2 seconds on average, and a packer takes 3 seconds on average to pack one. factor: 0 runs the model as fast as it can, so an hour of simulated time ends in under a second:
The command runs the shop with three packers for one simulated hour, at seed 7, and writes the run's statistics to result.json:
| docs/snippets/orchestrators/commands.sh:direct | |
|---|---|
Pass the Result Through a File
desflow logs to stderr and prints no result to stdout in a single run, so a later task reads the figures from the file that --summary writes, not from the output. Airflow, for example, keeps only the last line of a command's output, which is a desflow log line. Write the Result to a File lists the file's keys.
The Airflow, Dagster and Prefect pages read three rows of the file's statistics list, and the Kestra page logs the whole file:
tasks.pack.finished: the orders packed.tasks.pack.wait_mean: the mean wait for a packer, in seconds.resources.packers.utilization_avg: the packers' mean utilisation over the run, in percent.
The fixed seed gives every orchestrator the same figures as the command run in a terminal.
The command exits with 0 when the run finishes or is stopped by Ctrl+C or SIGTERM, 1 when the blueprint cannot be loaded, and 2 when a file is missing or an option is refused. Each orchestrator fails its task on any code other than 0. A stopped run flushes its sinks and writes no summary, so the task that reads result.json fails, not the task that ran desflow.