Apps and workloads

A workspace holds apps, and an app holds workloads. A workload’s decorator is its whole definition: image, resources, volumes, secrets, schedule, and scaling. Deploying makes the workspace match the code, with no dashboard setting or CLI flag on top, so infrastructure changes arrive as one reviewable diff.

Workload kinds

Kind	For	Triggered by	Scales
Function	Batch jobs, background work, GPU tasks	`.remote()`, `.map()`, `.spawn()`, `lazycloud run`	Autoscaler on queued calls
Schedule	Recurring work	A function’s `cron`	Like a function
Endpoint	An HTTP API from one function	HTTPS requests	Autoscaler on requests in flight
ASGI app	A FastAPI or Starlette app	HTTPS requests	Autoscaler on requests in flight
Realtime handler	A WebSocket service	WebSocket messages	Autoscaler on requests in flight
Pod	A container command, such as a model server	Connections, or `create()`	By hand
Sandbox	Generated code, agent workspaces	`create()` from your code	One container per `create()`

Containers start when work arrives and stop when idle, and every remote call is a task. A deploy starts every kind except sandboxes.

Define an app

import urllib.request

from lazycloud import App
from pydantic import BaseModel

app = App("page_stats")


class PageStats(BaseModel):
    bytes: int
    words: int
    links: int


class AnalysisStarted(BaseModel):
    task_id: str


@app.function(cpu=1, memory="512Mi", timeout_seconds=120)
def analyze_page(url: str) -> PageStats:
    with urllib.request.urlopen(url, timeout=30) as response:
        html = response.read().decode("utf-8", errors="replace")
    return PageStats(bytes=len(html), words=len(html.split()), links=html.count("<a "))


@app.endpoint(name="analyze", methods=["POST"])
def analyze(url: str) -> AnalysisStarted:
    call = analyze_page.spawn(url)
    return AnalysisStarted(task_id=call.task_id)

The endpoint answers at once with a task ID, and analyze_page fetches the page in its own cloud container, so a slow site never holds the request open. App names use lowercase letters, digits, and underscores, start with a letter, and run up to 63 characters. Workload names default to the Python name and are unique per kind within an app. Values that change between workspaces live outside the decorator, credentials in secrets and files in volumes.

Deploy

uv run lazycloud deploy page_stats            # every workload in the app
uv run lazycloud deploy page_stats:analyze    # just the endpoint

From Python, app.deploy() and analyze.deploy() do the same. Handler references covers the syntax.

Split an app across files

Each module imports the app and registers its own workloads:

# page_stats/app.py
from lazycloud import App

app = App("page_stats")

# page_stats/api.py
from page_stats.app import app
from page_stats.workers import AnalysisStarted, analyze_page


@app.endpoint(name="analyze", methods=["POST"])
def analyze(url: str) -> AnalysisStarted:
    return AnalysisStarted(task_id=analyze_page.spawn(url).task_id)

The CLI finds workloads by importing the module you name. page_stats.api imports page_stats.workers, so it reaches both:

uv run lazycloud deploy page_stats.api            # analyze and analyze_page
uv run lazycloud deploy page_stats.api:analyze    # just the endpoint

page_stats.app alone holds no workloads. When no module imports the rest, a small entry module that imports each one does.

Remove workloads

A workload deleted from the code stays deployed until a deploy with --prune:

uv run lazycloud deploy page_stats --diff --prune   # preview
uv run lazycloud deploy page_stats --prune

Pruning stops every deployed workload the app no longer declares and removes its schedules and routes. It needs the whole app, so a single-workload reference can’t prune. Volumes, secrets, and stored data stay, even when you delete the whole app. From Python, app.deploy(prune=True).

Call deployed workloads from other projects

uv run lazycloud app export page_stats

This writes a typed package to lazycloud_clients/page_stats, so another project can call the deployed functions with remote() and the endpoints with request(), without your source. ASGI routes get typed methods from the app’s OpenAPI schema, from /openapi.json unless --openapi-path or --openapi points elsewhere. The package holds no credentials, since callers sign in with lazycloud login. Pods and sandboxes aren’t exported, and a contract change needs a new export.

Pause and delete

uv run lazycloud app pause page_stats     # stop every workload, keep definitions
uv run lazycloud app resume page_stats
uv run lazycloud app delete page_stats

Workspaces

A workspace holds apps, deployments, tasks, and stored data. Within it, a name always means the same resource, so two apps mounting the models volume see the same files. Separate projects, or staging and production, fit separate workspaces. Tokens, a connected AWS account, joined machines, domains, and your plan belong to your account and apply to every workspace you own.

uv run lazycloud workspace list
uv run lazycloud workspace use staging

LAZYCLOUD_WORKSPACE overrides the stored choice for every command, which suits CI, and --workspace overrides it for one. Compute covers where a workspace’s workloads run, and the dashboard covers members.

Getting started

Core concepts

CLI reference

Examples

Account

Apps and workloads

Workload kinds

Define an app

Deploy

Split an app across files

Remove workloads

Call deployed workloads from other projects

Pause and delete

Workspaces

​Workload kinds

​Define an app

​Deploy

​Split an app across files

​Remove workloads

​Call deployed workloads from other projects

​Pause and delete

​Workspaces

Workload kinds

Define an app

Deploy

Split an app across files

Remove workloads

Call deployed workloads from other projects

Pause and delete

Workspaces