app/docs/concepts/runs.md

# Runs: pipelines that finish

A flow that automates a house never ends — a value arrives, nodes fire, and it
waits for the next one. A research pipeline is the other shape: parameters go
in, stages execute in order, and at some point it is *done* and has produced
something worth keeping. Fluksio calls the second one a **run**, and it is the
same engine either way.

This is what makes Fluksio usable where Kedro, MLflow or ClearML would be: a
run has parameters that identify it, a result, per-step metrics, artifacts and
a place in a queryable history — without a second server, and without paying a
project bootstrap on every execution.

## A batch flow

Set `mode: "batch"` on the flow and name the messages its result should hold:

```json
{
  "name": "train_polymer_gnn",
  "mode": "batch",
  "outputs": ["final_loss", "report"],
  "inputs": [
    {"spec": {"name": "lr", "dtype": "float"}, "initial": 0.1},
    {"spec": {"name": "steps", "dtype": "int"}, "initial": 3000}
  ],
  "nodes": [{"id": "train", "timeout": 7200, "device": "gpu", "...": "..."}]
}
```

A batch flow is built and validated like any other — it appears on the canvas,
its ports are type-checked — but it is never *activated*: no subscriptions, no
schedules, no webhooks. It runs when a run asks it to, and not otherwise.

Its `inputs` are its parameters. A run supplies values for them; anything it
does not supply keeps the declared initial value.

One thing a batch flow may not do is rate-limit a port (`interval`). A rate
limit holds a value back for a timer to release, and a run has no timer — the
value would be dropped rather than delayed, so submitting is refused instead.

## Submitting

```bash
curl -X POST $FLUKSIO/runs/flows/train_polymer_gnn \
     -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
     -d '{"params": {"lr": 0.3, "steps": 4000}, "seed": 7}'
```

The answer is immediate and the run is `queued`; a training run is measured in
hours, so nothing waits for it. Poll `GET /api/v1/runs/{id}` for its status,
result, per-node record and artifacts.

Wrong parameters are refused before anything executes — an undeclared name, or
a value of the wrong type, comes back as a 422 naming the problem.

### Sweeps

An ensemble is the same parameters at different seeds; a grid search is the
parameters spread out. Both are one call, and the caller builds the list:

```bash
curl -X POST $FLUKSIO/runs/flows/train_polymer_gnn/sweep \
     -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' -d '{
  "runs": [{"params": {"lr": 0.1}, "seed": 1}, {"params": {"lr": 0.3}, "seed": 1}]
}'
```

From a terminal that is `fluksio sweep train_polymer_gnn --param lr=0.1,0.3`,
which builds the product of the lists you give it and posts the same call.

They share a `group_id`, so `GET /api/v1/runs?group=…` is the sweep, and they
execute in parallel. That is safe because **each run has a state backend of its
own**: message names are global keys, so two runs of one flow would otherwise
overwrite each other's values. They do not.

## Producing values before you are finished

A training loop has numbers worth keeping long before it has a result. Those
numbers are **outputs**, not logs: a node declares a port for them and produces
them over time, which in Python is a generator.

```python
def process(lr, steps):
    loss = 1.0
    for _ in range(steps):
        loss = train_one_step(lr)
        yield {"loss": loss}          # published now, on the `loss` port
    return {
        "weights": fluksio.save_artifact(dump(model), "weights.npz"),
        "final_loss": loss,
    }
```

Mark the port it streams on, so the flow says what it does:

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

Every yield is published the instant it happens — same port, same type check,
same place on the canvas as any other value. Whatever the generator `return`s
is the node's result, and is what downstream nodes read. If you never
`return`, the last thing you yield is the result instead.

This is the whole reason the framework does not have a logging API. A metric
that escapes through `log_metric()` is undeclared: invisible to validation,
absent from the canvas, and stored somewhere the graph knows nothing about.
A metric that leaves through a port is a message — so a chart binds to it
directly, a downstream node can consume it, and the run keeps its series
without anyone asking.

**Where a yield cannot reach** — the value comes from inside somebody else's
callback, and they call you rather than the other way round — `fluksio.emit`
writes the same ports the same way:

```python
import fluksio


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

### What a run does with them

Every number a node emits is kept as the run's series, stepped by the count of
emissions on that message. Read one back with
`GET /api/v1/runs/{id}/metrics?name=<flow>.loss` — or leave `name` off for
every series the run kept — or compare runs:

```
GET /api/v1/runs/series/compare?ids=<a>,<b>,<c>&metric=<flow>.loss
```

That answers in the `series` shape a chart widget already draws, so three
training curves side by side is a widget binding. During a run the values also
arrive live on the flow socket, so a chart bound to the port fills in as the
training goes.

A streaming port may set `interval` to thin out what reaches the canvas — the
run's history still keeps every value, because the interval is asking for the
display not to be flooded, not for the curve to have holes in it.

Emitting has a second effect: **a node's timeout measures silence, not
duration**. A node that yields every few seconds can run for hours under a
`timeout` of 300; one that says nothing for longer than its timeout is killed.
Set `timeout` on a long node to how long it may plausibly go quiet.

In a *live* flow, an emission also wakes whatever is downstream of it, exactly
as a subscriber publishing does. In a run it does not: a run's graph is
scheduled once, and three thousand mid-node cascades would leave "the run has
finished" with nothing to mean.

## Artifacts

Bytes never travel as a message. `save_artifact` writes them to a
content-addressed store and returns a small reference — digest, size, media
type, name — which is what an `artifact`-typed port carries:

```python
def process(weights):          # requires: weights, dtype "artifact"
    path = fluksio.load_artifact(weights)
    ...
```

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. Artifacts a run produced are listed on it and
downloadable at `GET /api/v1/artifacts/{digest}`.

## Stage caching

A run mostly does not redo what an earlier one already did. Before a node
executes it is fingerprinted — a sha256 over its source, its settings and the
values it is about to read — and if some earlier run of that same fingerprint
finished, what that one returned is restored into this run's state and the node
is skipped. It is recorded with the status `cached` and a duration of zero, and
its artifacts are listed on the new run as well, so they stay downloadable from
either.

The settings go into the key raw, so a secret contributes its `{"$secret":
name}` reference and never its value. An artifact input counts as its content
digest: the same bytes under a different filename are the same input. Each node
on a run carries the `cache_key` it was looked up by.

The run history *is* the cache; there is no second store. A node's returned
outputs are kept on its run record as canonical JSON, up to 256000 characters —
a node returning more than that is simply not cacheable that run. An entry
whose artifact bytes have since left the store is a miss, not an error.

Only `python` nodes are cached, and by default all of them are. A built-in node
type or a connector node has side effects and no source to fingerprint, so
neither is ever a candidate. Turn it off for one node with
`@node(..., cache=False)` — the flow document carries it as `cache`, so the
canvas and the API can change it too — or for one run with
`fluksio run --no-cache`, `fluksio sweep --no-cache`, or `"no_cache": true` in
the submission body.

What a cached node does not bring back is what it emitted on the way. Its
returned outputs are restored; the values it published mid-execution are not,
because those were the story of an execution that is not happening this time.
So a skipped training node contributes no loss curve to the new run — if you
want the curve, that run has to actually train.

## Objects that cannot be serialized

A live model, a `DataLoader`, a JAX-compiled function — these do not cross a
node boundary, and no framework flag will make them. There are exactly two
patterns, and they are both deliberate:

* **Keep them in one node.** Stages that must share live memory are one node.
  Building the model and training it is one stage; the fact that Kedro would
  make them two nodes is Kedro's problem, not a structure worth reproducing.
* **Cross at a checkpoint.** Save what matters as an artifact and rebuild from
  it on the other side. That is the boundary that also survives the next node
  running on a different machine.

## Running a node somewhere else

A node that needs a GPU declares the label of the machine that has one:

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

A worker on that machine dials **out** to the engine, because the engine
generally cannot reach it — different network, no inbound route — and because
nothing should expose Redis across hosts. Install it on the box, mint it a
token, and start it:

```bash
curl -X POST $FLUKSIO/workers/tokens -d '{"name": "gpu-dev"}'   # once, as an admin

pip install fluksio-worker

fluksio-worker \
  --url wss://api.example.com/api/v1/workers/attach \
  --token "$FLUKSIO_WORKER_TOKEN" \
  --labels gpu,cuda12 \
  --python /opt/torch-venv/bin/python
```

`fluksio-worker` is its own distribution — the agent, the node runner, and
`websockets`. Nothing of the engine, so a GPU box does not install a database
driver to run a training step. Where pip is not an option, the two files still
work copied into one directory and run with `python agent.py …`; the engine
serves the runner at `GET /api/v1/workers/runtime`.

`--python` is the interpreter node code runs on, which is how the GPU box keeps
its CUDA wheels without the engine ever installing them. The node's source
travels with every call, so nothing has to be deployed there.

A few consequences worth knowing:

* `import fluksio` inside a node is the worker's own reporter — `emit`,
  `save_artifact`, `load_artifact` — installed before the node's code runs, so
  the installed `fluksio` package (if the box has one) never shadows it.
* 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 on the engine would fail a node that is fine.
* If nothing carrying the label is attached, the run stays `queued` and says
  what it is waiting for. Submit first, switch the GPU box on later.
* Cancelling a run kills what it is executing, there or here, and leaves other
  runs of the same node alone.
* If the worker disappears mid-call, the run fails in seconds with
  `worker went away mid-call` rather than waiting out its timeout.
* `device_policy: "prefer"` runs locally when no such worker is attached;
  `"require"` (the default) waits for one.

## Durability

Submitting journals the run to a Redis stream of its own, separate from the
one the automations use — a burst of five hundred sweep runs must not stand
between a house and its heating. An engine that is down when a run is
submitted picks it up when it starts.

From the moment a run is claimed, its database row is the record and the queue
is finished with it. Redelivering two hours of training because an
acknowledgement was late is not recovery; instead a running run refreshes a
lease, and one whose lease goes stale is marked `abandoned` — which is what a
run whose engine was killed mid-training becomes.

## What this costs, compared

The repository ships a benchmark that measures submitting a run against a
Kedro project doing the same nothing:

```
fluksio — submit accepted     median   15.3 ms
          submit -> result    median   60.8 ms
kedro   — kedro run           median 1109.5 ms
```

The difference is not the orchestration; it is that Fluksio does not boot a
project per run. The engine is already up, and the workers already have the
node's code compiled. On a 510-run sweep, that gap is about nine minutes of
pure startup that never happens.

## See also

- [Getting started: data science](../getting-started/data-science.md) — the same
  material as a first setup
- [Writing node code](../code/nodes.md) — generators, `fluksio.emit`, artifacts
- [Remote workers](../code/workers.md) — attaching the machine with the GPU
- [The HTTP API](../code/api.md) — every endpoint used above