Five concurrent training nodes, each sizing its thread pool to every core,
left the engine's own event loop unscheduled: the API stopped answering
within 10 s and every client died. The same shape on a GPU deadlocked a run
for 21 minutes at 0% utilisation with nothing failing and nothing to read --
it just sat in `running`.
@node(resources={"cpus": 2}) is the declaration. The engine holds that much
for the length of the execution, so more of them than the machine has room
for wait their turn rather than oversubscribing it, and a `gpus` node holds
its card exclusively. FLOW_CPUS defaults to every core but two, and those two
are what keeps the engine answering.
Because a thread cap is read when the process imports the library, a warm
worker cannot be told a different one -- so an environment gets a pool of its
own and nodes deriving the same one share it, rather than paying a cold start
per call on exactly the nodes whose imports are slowest. XLA_FLAGS is never
derived: it is a composed, version-dependent string, so it travels in
resources.env where it is visible.
A node that declares nothing is not accounted for and behaves as it always
did -- it just gets FLOW_CPUS/FLOW_MAX_WORKERS as a thread cap, which is the
half of this that fixes the reported incident without anybody declaring
anything. An operator who set OMP_NUM_THREADS themselves still wins.
Resources are claimed strictly before a worker slot, so the two blocking
waits cannot deadlock. A node queued for them publishes node_queued and shows
on GET /workers/resources, because waiting and hanging looked identical.
Accounted, not enforced: no cgroups, no rlimits. Scheduling across machines,
flavours and enforcement are the next steps.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four things the python SDK turned up, each fixed where every client sees it.
A key no port declares is now an error rather than a silent drop, on the
return, the yield and the emit alike — the contract the docs already stated.
The SDK reads literal yields at sync time, so a typo fails before anything
runs, and an emission of one fails the call rather than being logged where
nobody looks.
NaN and infinity are refused at the port. JSON cannot spell either, so one
that travelled came back as a 500, a socket frame that stopped the canvas, or
a metric batch the database dropped whole.
An artifact input takes `@run:<id>.<output>` or a bare digest, resolved on the
engine — so the CLI, the run dialog and a python caller mean the same thing,
and a sweep can pass one at all.
Node timeouts are off by default. The clock measured silence, which a training
node is full of, and remote workers had already stopped enforcing it — their
heartbeat reset it. Now a heartbeat proves the agent rather than the node,
ninety seconds of nothing fails the call either way, and the engine touches
work it is still running so a long node is not redelivered at sixty seconds.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019V5bsYGNxcgPs4xXmTPx69
A node's timeout now covers its body only: the pool loads the source into the
worker it picked, off the node's budget, so imports that outlast the timeout no
longer make a node impossible to run. Draft checks compile without caching, so
saving does not evict what a busy node is serving calls from. Requests carry an
id the worker echoes and the pool checks, a reply is encoded once, and the
remote-exception cache is bounded.
DELETE /flows/{name} answers 409 while the flow has a running or queued run,
which is what was letting run_node rows outlive their run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01StpRc2C6au1WJ1EUU7fsfu
One process owns this database — the image has run a single uvicorn
worker for that reason since the four-engines bug — so a file beside the
flows is the honest shape for it, and it is what lets `fluksio serve`
need no infrastructure at all. Live values, node execution and the work
queue never came here anyway; what does is a rollup a minute at a time,
a row per cascade and the run history, and WAL keeps the readers going
while that one writer works.
DATA_DIR is now the one setting that moves everything an installation
keeps; the rest derive from it and the images still spell theirs out.
The schema is prepared in-process at startup, so the prestart service is
gone, and the ten Postgres-only revisions collapse into one portable
baseline.
Three things only worked because psycopg was casting for us: a token's
subject arriving as a string where the column is a UUID, `greatest`, and
`date_bin`. The timestamps needed a column type of their own — SQLite
stores no offset, and a naive datetime read back either raises against an
aware `now` or serialises as local time.
Postgres stays in the stack only for Umami, behind the analytics profile.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A cluster or GPU host installs `pip install fluksio-worker` and gets the
agent and the runner, not psycopg, numpy and the MCP SDK. The engine
depends on it as a workspace member, so the file it launches node code
with is the same file a remote worker runs — which is what keeps a node
unable to tell the difference. Copying the two files by hand still works.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A wheel whose top-level module is `app` collides with anything else in a
user's venv, so the package that is about to be published takes the name
it is published under. Only the Python package moves; the repo, the
Docker WORKDIR and the compose project keep theirs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A python node's settings are constants of its own function, so they are passed
the way its ports are: by name. The controller binds them to the compiled
function, the `params` field is gone from the worker and remote protocols, and
a setting sharing a port's name is reported as a node error rather than
shadowing it. The panel's scaffold follows suit and keeps the header in step
with both ports and settings.
The demo's `pace` moves from a flow input to a setting of the training node,
which is what it always was.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NUb8YpL2s3gmN9WTACTt4q
The first cut had node code call fluksio.log_metric, which was a second,
undeclared way for data to leave a node: invisible to validation, absent from
the canvas, and stored where the graph could not see it. That is precisely the
MLflow discrepancy this framework exists to avoid, so it is gone.
A node that produces values over time is a generator. Every yield is a dict
keyed by output port, published the instant it happens — same port, same type
check, same place on the canvas as any other value — and what it returns is
its result. A port doing this declares stream: true, and a run keeps every
number one takes, so experiment tracking is a consequence of the graph rather
than an API beside it: a chart binds to a training curve the way it binds to a
temperature. fluksio.emit writes the same ports imperatively, for where a
yield cannot reach — inside a training framework's callback.
In a live flow an emission also wakes what is downstream, as a subscriber
publishing does; in a run it does not, because a run's graph is scheduled once
and mid-node cascades would leave 'finished' with nothing to mean. The
enqueued item carries no payload: the value is already in state, and one
carrying it would re-apply an old emission after the node returned.
Verified on the stack: 30 loss values arrived live on the flow socket during a
run, attributed to the node that produced them, and the same node run on the
remote worker streamed its curve back across the socket.
Also caches remote compile results per worker, so attaching a GPU box does not
put a network round trip in every rebuild.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AD8SfVhzXBG2nAfFcVh3iD
A checkpoint is not a message. DType.ARTIFACT carries a reference — digest,
size, media type, name — so everything on the wire stays JSON and thirty
megabytes never sit in Redis, which answers the vision's open binary-payload
question by narrowing it: inline codecs would only serve payloads too small to
be worth a round trip, and nothing asks for that.
The store is content-addressed rather than per-run, for three reasons that all
pay later: a sweep whose fifty configs share one preprocessed input stores it
once, a reference stays valid however it is passed around because it names
content instead of a location, and the digest is what a stage cache will
compare — so building it in now is what keeps that from being a change to the
message contract.
Node code calls fluksio.save_artifact/load_artifact and cannot tell whether it
is writing the engine's own directory or putting bytes over HTTP, which is
what will let the same flow run on a remote worker unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AD8SfVhzXBG2nAfFcVh3iD
A training loop has numbers worth keeping thousands of steps before it has a
result. Node code now imports fluksio and calls log_metric/progress, which
sends a line back without ending the call; the engine writes those to
run_metric in batches from the run's own driver rather than folding them off
the event bus, which drops what it cannot keep up with.
Two things fall out. Each event resets the worker deadline, so a node's
timeout measures silence rather than duration — which is what lets a two-hour
training keep a liveness contract instead of racing it. And the worker pool's
_running is now keyed by (run, node), so cancelling one config of a sweep
kills that training and leaves the rest alone.
Fixes a latent framing bug: read_line returned whatever a read had taken,
which was fine while a worker only ever sent one line per request and
unparseable as soon as it sent several. It now keeps the remainder.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AD8SfVhzXBG2nAfFcVh3iD
The worker script is handed to the interpreter by path, so app/flow was
sys.path[0] for every node: `import queue` got the engine's. It now drops
its own directory before anything else imports, and runs with the
deployment's credentials scrubbed out of its environment.
Also: reload builds off the event loop, the pool wakes what is blocked on
it when it stops, a refused metrics flush is kept for the next one rather
than dropped, and the cascade events are paired through failures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017MeiWk3Yq12n2pTvnQWYvt
User code no longer execs in the engine. A pool of persistent worker
subprocesses speaks one JSON object per line; the controller installs a
proxy as the node's function, so every execution path funnels through it
and the pipeline is untouched. A crash costs one subprocess, a per-node
timeout is a kill, and cancelling from the canvas is that same kill.
The workers run a venv of the user's own on the data volume, filled from
a pip manifest versioned beside the flows. Applying it retires the
workers and rebuilds, so a package lands without restarting the engine.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017MeiWk3Yq12n2pTvnQWYvt
stroblmeandClaude Opus 5