app/docs/code/nodes.md

# Writing node code

A Function node is a Python file: no base class, no decorator, no framework
import unless you want one.

```python
def process(temperature, setpoint=21.0):
    """Ask for heat when the room is below the comfort point."""
    return {"heat": temperature < setpoint}
```

## The rules

**One function called `process`.** If the file defines exactly one public
function under another name, that one is used instead. Two, and the node
refuses to load rather than guessing.

**Arguments come from ports and settings, by name.** `temperature` above is an
input port; `setpoint` is a setting typed into the node's panel. Both arrive as
keyword arguments, which is why a setting may not share a name with a port.
See [Where a node's values come from](../concepts/values.md).

**The return value is a dict keyed by output ports.** Every value is checked
against the port's declared type before it is published. A key that is not a
declared port is an error, not a silent drop. Nothing leaves a node except
through a port it declared.

**Nothing else is importable from the engine.** Node code runs in a separate
process, on a separate interpreter, with none of Fluksio's own modules on its
path. What it can import is what the [Modules](../interface/operations.md)
screen installed, which can include [your own project](#your-own-code-as-a-package),
so a node need not be a self-contained file.

**A node is a pure function of its inputs.** No context object, no global
store, no handle to reach for. A running total or a debounce timer has a
specific shape; see [Keeping state in a flow](../concepts/state.md).

## Producing values over time

A node that produces values *during* its execution is a generator. Every
`yield` is a dict keyed by output port, published the instant it happens:

```python
def process(lr, steps):
    loss = 1.0
    for _ in range(steps):
        loss = train_one_step(lr)
        yield {"loss": loss}          # published now
    return {"final_loss": loss}
```

Whatever the generator `return`s at the end is the node's result, which is what
downstream nodes read. If you never `return`, the last thing you yield is the
result instead.

Mark the port so the flow says what it does:

```json
{"name": "loss", "dtype": "float", "stream": true}
```

Two consequences. In a [run](../concepts/runs.md), the whole series is kept as
that run's metrics, which is why there is no `log_metric()` anywhere in the
API. And **the node's timeout starts measuring silence rather than duration**:
each emission resets the deadline, so a node yielding every few seconds can run
for hours under a timeout of 300.

### `fluksio.emit`

Where a `yield` cannot reach, because the value comes from inside somebody
else's callback and they call you rather than the other way round:

```python
import fluksio


def process():
    model.fit(callbacks=[LambdaCallback(
        on_epoch_end=lambda epoch, logs: fluksio.emit(loss=logs["loss"])
    )])
    return {"weights": ...}
```

Same ports, same type checking, same publication. Prefer `yield` where you can
reach it; `emit` where you cannot.

## Bytes: artifacts

Messages are JSON, which is what lets the same value pass through Redis, the
work queue and the worker protocol unchanged. A checkpoint is not that.

```python
import fluksio


def process(dataset):
    path = fluksio.load_artifact(dataset)     # → a local path to read
    ...
    return {
        "weights": fluksio.save_artifact("model.pt", media_type="application/octet-stream"),
        "score": 0.94,
    }
```

`save_artifact` takes bytes or a path, stores them by their SHA-256 digest, and
returns a small reference (digest, size, media type, name) which is what an
`artifact`-typed port carries.

Because the address is the content's hash, a sweep whose fifty configs share
one preprocessed input stores it once, and a reference stays valid wherever the
store is reachable from, including on another machine.

`fluksio artifacts RUN NAME` is how one comes back out at a terminal.

### `fluksio.logger`

A node's `print` is kept as that node's logs, and so is anything on
`fluksio.logger` is the same logger the SDK exports at top level, so code that
runs both inside a node and outside one says it the same way:

```python
fluksio.logger.info("resuming from epoch %d", start)
```

### Media

Say what the bytes are and the port can be typed for them:

```python
def process(speech):                     # an `audio` port
    clip = fluksio.load_artifact(speech)
    words = transcribe(clip)
    return {"transcript": words}         # a `str` port
```

```python
def process(camera_url):
    for index, jpeg in enumerate(grab(camera_url)):   # a generator
        frame = fluksio.save_artifact(
            jpeg, f"frame-{index:05d}.jpg",
            media_type="image/jpeg", volatile=True,
        )
        frame["meta"] = {"seq": index}
        yield {"frame": frame}                        # an `image` stream port
```

An `image`, `audio` or `video` port is an artifact reference whose media type
has to match, so a node declaring `audio` never receives a video by accident.
See [Payload types](../reference/payload-types.md#image-audio-video) for what
each carries and what rates are realistic.

`volatile=True` is for a frame rather than a result. The bytes go to a ring in
memory instead of the data volume, and the engine pushes them down the
websocket to whichever screens are drawing that message — so a camera runs at
ten frames a second without writing anything to disk. They last as long as it
takes newer frames to need the room. Leave it off for a clip somebody asked to
keep.

!!! warning "Emitted media is not kept; returned media is"

    Only what a node *returns* is recorded against its run. Frames yielded
    along the way are replaced in state by the next one, and the artifact sweep
    removes bytes nothing refers to any more, which is what stops a camera
    filling the disk. If a particular frame matters, return it: a volatile one
    is copied out of the ring when the run records it, so returning it is also
    what makes it outlive the next few seconds.

## Printing

`print` works and is captured. The first 16 KB per call is kept and shown in
the flow editor's log panel and on the run's per-node record; the rest is
dropped, so a node printing in a loop cannot fill anything up.

Use it to debug, not to record results. A number worth keeping is an output
port, not a line of text.

## Errors

An exception fails that node's execution, not the flow. The message you see is
one line from the frame in *your* code, not a stack through the engine.

The node keeps its last error visible after it recovers, so a failure that
fired an alert at 03:00 still says what it was at 09:00. It can also be
acknowledged from the canvas.

## Timeouts

`timeout` on a node is how many seconds its code may be *silent* before it is
stopped. A yield or an `emit` resets the clock, and the first call's imports are
not charged to it: a node importing torch is not being slow, it is loading.

There is no timeout by default. Training runs for hours and a node that reports
nothing is usually working, so the engine waits: what fails a call is the worker
dying, which arrives at once rather than after a deadline. Set a timeout on the
nodes where silence means stuck (an HTTP call that should answer in seconds, a
loop that can spin) either per node or, for all of them,
with `FLOW_NODE_TIMEOUT`. `timeout = 0` says explicitly that this node has none.

## Running a node somewhere else

A node declares the label of the machine it needs:

```json
{"id": "train", "device": "gpu", "device_policy": "require", "timeout": 7200}
```

`require` (the default) waits for a worker carrying that label; `prefer` runs
locally when none is attached. A node bound to a device is compiled *on that
machine*: a node importing `torch` is correct on the GPU box and a missing
module on the engine, so checking it here would fail something that is fine.

See [Remote workers](workers.md).

## Sharing code between flows

A node's source can be promoted to the shared library from its panel, and other
flows can then use it by reference. There is one copy and one place to edit, and
every flow using it runs the edit.

Shared sources live in `_lib/` in the flow repository, so they are versioned
with everything else.

## Packages

Node code runs in a virtual environment of its own, on the instance's data
volume, separate from the one Fluksio itself runs on.

Declare what you import in [Modules](../interface/operations.md), or over the
API:

```sh
curl -X POST $FLUKSIO/modules/apply -H "Authorization: Bearer $TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"requirements": "numpy>=2\npandas\n"}'
```

It is a pip manifest installed with `uv pip sync`, versioned alongside your
flows. An install takes effect immediately; nothing restarts.

### Your own code as a package

!!! tip "If Fluksio is installed in the venv you work in, skip this"

    Node code then runs on that environment, so your project and everything it
    imports are already importable; see
    [Getting started: data science](../getting-started/data-science.md). What
    follows is for a Fluksio with a venv of its own, which is what a container
    always has.

A manifest line can name a directory, so the project you already have is
installable like any other dependency:

```text
-e /home/you/my-research
numpy>=2
```

A node body then imports it, and the logic stays where it already lives: in
your repository, under your own version control, importing its own siblings:

```python
from myresearch.train import fit


def process(lr, epochs):
    return fit(lr, epochs)
```

That is the whole of it. The node is three lines, `myresearch` can be as many
modules as it likes, and nothing was copied.

!!! tip "You can have those three lines written for you"

    Decorate `fit` with `@node(...)` where it is defined, say which nodes make
    a flow with `Flow(...)`, and `fluksio sync` generates the body above,
    along with the flow document, so there is nothing to PUT by hand. The
    declaration lives beside the function it describes and is checked against
    its signature. See
    [Getting started: data science](../getting-started/data-science.md).

!!! warning "Editable, but not live"

    `-e` means edits reach the venv without reinstalling, but a node's process
    already holds the imported module in memory. The engine's workers are
    long-lived, so a change to your code is picked up when they are retired,
    which is what **Apply** does. Pressing it after an edit is the loop.

    A [worker](workers.md) you attach yourself is the exception: it starts a
    process per call, so it reads your code fresh every run. If you are
    iterating on the code many times an hour, point one at your own
    interpreter (`fluksio-worker --python "$(which python)"`) and mark the
    node with its label.

!!! note "The path is a deployment detail"

    It is resolved on whichever machine runs the node, and the manifest is
    committed to the flow repository, so an absolute path from your laptop
    means nothing inside a container or on a GPU box. Those need their own
    install of the same project; a VCS requirement
    (`myresearch @ git+ssh://…@a1b2c3d`) travels where a path does not.

## A worked example

The repository ships a small supervised fit as a seedable demo: three nodes,
a batch flow, streaming metrics, artifacts between stages, and a GPU-labelled
node that falls back to the engine when no worker is attached. It is the
shortest complete thing to read:

```text
prepare ──dataset(artifact)──▶ train ──weights(artifact)──▶ evaluate
                                 │
                                 └── loss (streaming float) ──▶ chart
```

The example is built by a seed script the maintainers run against a live
stack; the flow above is what it produces.

## See also

- [Where a node's values come from](../concepts/values.md)
- [Keeping state in a flow](../concepts/state.md)
- [Runs: pipelines that finish](../concepts/runs.md)
- [Node types](../reference/node-types.md) — the ones you do not have to write