Three things the first pass left.
`serve` now names the flows asking for a GPU when the engine has none
declared. The placer already warned, but into the log, where a fresh install
that forgot `--gpus` does not read it — and the cost of missing it is GPU
nodes running concurrently, which is what the declaration exists to prevent.
The seed was the one field an export still had to coalesce: `--seed 1`
filled the run-level column and left `param.seed` blank, while a declared
seed filled the parameter and left the column blank. It is resolved like
every other input now, and the column carries the seed the run actually used
however it arrived — including when a parameter outranks the run's own,
where the column used to report the one that lost.
And the docs say plainly that declaring the card is what buys the worker
retirement: a node that imports jax without `resources={"gpus": 1}` never
gets CUDA_VISIBLE_DEVICES, so nothing marks its worker as one holding a card.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`fluksio serve` printed a log stream and nothing else, so watching an engine
meant a second terminal running `status --watch`, and stopping or pairing it
meant a third. At a terminal it now opens a dashboard: the health and flow
overview `status` draws, the recent runs as a table, and the engine's own
output in a pane below — which is what the earlier decision against this
was protecting, and it is still all there.
The engine is a child process running `serve --plain`, not a thread, so it
outlives the dashboard: q leaves it running and says so, s and r stop and
restart it, c cancels the selected run and e pairs with a portal. An engine
already serving this directory is adopted rather than duplicated, and it can
be stopped from here only because the pidfile and the token together prove
it is this installation's.
`--plain` and no terminal both keep the old behaviour, which is what the
container and CI run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`serve` moved off a port something else held and said so, which made
starting a second engine for one installation look like it had worked —
two engines on one SQLite file, which is not a supported shape. A serving
engine now records its pid beside its data, and a taken default port is
asked who it is: this directory's own engine is named and nothing is
started, another installation's Fluksio is named and the move happens
anyway, and anything else reads as it did before.
Whose engine it is is settled by the token, which is signed with this
directory's secret key — so a foreign one is only ever named, never
stopped from here. A directory with no credential yet asks
unauthenticated, and its answer is foreign, which is the side that stops
nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
A data-science environment installing fluksio waited for lxml, aiohttp and
the rest of a connector stack it has nothing to talk to. Outbound mail,
error reporting and the MQTT and InfluxDB clients moved to
`fluksio[server]`, which the image installs; each import is guarded and
names the extra. `tenacity` had no import site at all and is gone.
23 fewer packages and the compiled ones among them — a bare `pip install
fluksio` still serves, runs every python node, and registers the mqtt and
influxdb node types, which only need the library when one is actually
built. sentry-sdk arrives anyway underneath `fastapi[standard]`; what
changed there is that nothing of ours requires it.
The dev environment keeps every extra: the suite exercises the connectors
and strict mypy checks their call sites.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`fluksio status` listed recent runs and recent failures with no time on
them, so a red line said nothing about whether it was from a minute ago or
last week. Both carry an age now, spelled the way a duration is, and the
runs listing gained one too. The fields were already on the wire.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
GPU count is not detected, so FLOW_GPUS was 0 on a fresh install and a node
asking for one was silently clamped to zero and ran concurrently with every
other. Setting the variable serialised them, but it was an environment
variable only — `serve` had --max-runs and --max-workers and no --gpus.
The clamp warning now names the flag when nothing here declares a card.
The same flags are written into the environment before the settings are
built, so a value they refused died in a pydantic import naming no flag.
They are checked where they are typed instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
A worker that has run a jax node keeps holding the GPU after the run: XLA
preallocates most of the VRAM at import and never releases it, so the next
process OOMs on preallocation while a warm idle worker sits on the card.
Pools are kept warm on purpose — a library reads its environment at import,
so a warm worker cannot be re-told — but the end of a run is a point where
the memory should go back, and the environments carrying a GPU assignment
are exactly the pools that ran on one. Idle ones go now, busy ones when
they return.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`fluksio sync dev` refused a one-file-per-directory layout: s1..s4 each
holding `study.py` stopped the walk with "two files would both import as
'study'". The suggested `__init__.py` per study directory fixed discovery
and broke the bare `from study import ...` a test beside it does. A file is
now named for its path under the directory being synced — `dev/s1/study.py`
imports as `s1.study`, the directories between being namespace packages —
so nothing collides and nothing has to be added. A file at the root keeps
its bare name, and the generated shim follows: `import_root` walks one
directory per dotted segment and lands on the same root.
The collision message stays for the spelling that can still reach it: two
files named on the command line, each rooted at its own directory.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
A module defining nodes could not be run directly: the decorator refused
`__main__` while the module body was still executing, so a `if __name__ ==
"__main__"` self-check beside the nodes was impossible and the checks had
to live in a separate pytest file. The refusal now fires where the
generated body is written — document() and shims(), both, since sync writes
the document first — and the decorator hands the function back as it always
did. Syncing a __main__-defined node is still refused, with the same words.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
Inside a worker `import fluksio` is the reporter, which had emit and the
artifact calls but no logger — so `fluksio.logger.info(...)`, written
against the SDK's top-level export, died with AttributeError after the
training it was reporting on had already succeeded. Its records go to the
same capture a print does; the handler resolves sys.stderr per record
because a call runs under redirect_stderr.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`save_artifact` had no download counterpart: the run detail listed a run's
files and nothing in `fluksio --help` fetched one. `fluksio artifacts RUN`
lists them, `fluksio artifacts RUN NAME` writes one — under the name the
node saved it as, since the message name is chosen for the graph. The run
detail now carries that filename, which it held in the table and did not
report.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`export runs` dropped a `param.*` column whose value was constant across the
exported runs, so a downstream filter broke depending on which runs the
selection happened to hold. Every input the selection recorded is a column
now; `--params` still narrows it to a sweep's axis. The metrics default is
unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
A run submitted without explicit inputs recorded `params = {}`: Port
initials filled the values at node level and were never written back, so an
exported row had a blank `param.*` cell and the runs listing could not tell
a run that took every default from one submitted with those same numbers.
Declared initials are now folded in at submit, explicit values winning, and
the run-level seed still wins over a declared one.
`params_digest` is computed over the resolved values, so it changes shape
once: a run recorded before this does not dedupe against a newer identical
submit, and its stage-cache entries miss once.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
`fluksio run --param lr=0.002` died with a bare JSONDecodeError: `--param`
is not a `run` flag, so it became an input named `param` whose value
`lr=0.002` was json-decoded. The name check ran after the coercion, so the
decode error always won. Names are now checked first, and a coercion error
is a SyncError naming the input and its type, the way `_ask_params` has
always done it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc
Runs: a run can now be deleted (DELETE /runs/{id}, cancelling a live one
first), exported as csv from the screen's own filters, and its flow label
opens the flow. Its "Parameters" panel became "Inputs" and lists every
input the flow declares, marking the ones that took the flow's own value
rather than the run's — the comparison table resolves the same defaults
instead of printing "unset".
Home reads brain, dashboards, health, flows: the mosaic is one full-width
scrolling strip, and the flows list and the flow-activity rollups merged
into a single left-joined table so a flow's state and its numbers sit on
one row.
Charts take a drag to narrow the x window and a double click or tap to
come back out. UplotChart holds the scale and passes resetScales:false
while a window is held, which is what the old comment said made this
impossible.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019LrWVRguqbk33YzfEeUx5W
`GET /api/v1/search/` hands the client one flat index of everything worth
jumping to — flows and the nodes inside them, dashboards and the widgets on
them, panels, secrets, modules, workers and alert channels — and cmdk matches
it in the browser, so results narrow while typing without a round trip per
keystroke. A node hit is the one thing no list endpoint could answer: it opens
its flow with that node in focus.
The panel is reached from **Search** above Documentation in the sidebar, or
⌘K anywhere. The flow canvas palette moves to ⌘P, being the narrower of the two.
The panels dialog gains an address (`/dashboards?panels`) so a panel hit has
somewhere to land, and the sidebar logo gets `shrink-0`: the rail's width
animates while the logo is already back, and a flex item short of room is
squeezed rather than clipped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016vGH7jqcXxWKP9wZFPyVdU
`PUT /dashboards/{name}` had no backend test: three cover what it promises —
a first draft for a name nobody has used (200, not the 404 that came off in
4a2337f), an update that keeps everything the edit did not name, and a save
based on a version someone moved past.
The Playwright half is the same property through the editor. Pages and
sections are gone since 7ff29ca, so what the editor never draws — a widget's
md/sm placements and the dashboard-wide settings — is what a save has to
carry, and dragging one widget is what the spec makes it carry it through.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K1moruzue2kTJd3uVisgNk
The poll loop remembered what it read rather than what it published, so a
value the node could not publish counted as said: the next poll skipped it,
succeeded, and health went back to ok with the port still dark. Remember it
only after inject returns, and report ok last.
A node reporting itself down is now derived into its flow's issues on read
and counted on the health summary, so the canvas marks it and Home says so.
Being down does not stop the flow, and the issue clears by itself when the
node reports well again. The repeating poll warning is logged once per
outage rather than once per tick.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K1moruzue2kTJd3uVisgNk
ConnectorNode.stop cancelled its poll task and then caught CancelledError
around the await — the fourth site of the trap 93e4527 closed elsewhere,
swallowing a cancellation aimed at whoever asked for the teardown. It now
calls the shared Node._cancel_task, which keeps retrieving whatever the
loop raised on its way out, as the old `except (CancelledError, Exception)`
did.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K1moruzue2kTJd3uVisgNk
The Influx client was built with no timeout, so every query and write fell
through to influxdb-client's own 10 s default — invisible to a flow and
unchangeable. The param is in seconds like its peers; the client counts in
milliseconds, so the call sites convert.
The publisher backlog was a module constant, read once at import. It is the
depth at which the oldest payload is dropped and the node goes degraded, and
a node that bursts wants more than one that trickles, so it moves to Params
and is read where the queue is built.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K1moruzue2kTJd3uVisgNk
Without one, aiomqtt's disconnect acknowledgement has no deadline, so a
subscriber cancelled while its socket is dead never finishes unwinding and
teardown abandons the task. The knob is per node because brokers differ.
Node stop paths cancelled their background task and then caught
CancelledError around the await, which swallows a cancellation aimed at
the caller — the trap Supervisor._cancel already documents. One shared
Node._cancel_task now waits the way the supervisor does; mqtt's publisher
and subscription and delay's cron call it.
The api container also collected zombie python workers: orphaned when
--reload replaces the process holding their handle, they reparent onto a
PID 1 that reaps nothing but its own. `init: true` on the backend service.
A house's inverter broker dropped at 04:27 and the power flow was quarantined
20 seconds later. Quarantine was terminal — the supervised task returned and
only a publish or an engine restart could bring it back — so five hours of
power and battery readings are missing, and what ended it was an unrelated
`git pull` restarting uvicorn.
Two changes, both in that path:
- the failure budget is per task, not per flow. `power` runs an MQTT subscriber
and a Victron keepalive publisher against the same broker; they died together
and spent one shared budget in 41s, giving up before the 60s backoff step was
ever reached.
- quarantine is now a rest. The task sits out 5min, then 15, then an hour, and
each time gets its budget back and tries again, so a broker that comes back
is picked up without anyone watching. `quarantined` reads from whichever
tasks are currently resting.
The alert for it says when it will try again.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01C5H4uLCCpsbipL1R7WKCee
Three things the one-folder-per-study layout ran into.
**Discovery walks down.** A plain directory is now walked all the way, so
`fluksio sync dev` finds `dev/s1_baseline/study.py` and naming each study is
no longer the price of the layout. Hidden directories, `__pycache__`,
`node_modules` and virtualenvs are left alone, and a package is taken whole.
Two files that would import under one module name are refused, naming both:
Python keeps one module per name, so the second would silently *be* the first
— and a node's generated body imports by that name, so a worker would run the
wrong study's code. The message says the fix, which is an `__init__.py` per
study directory. A module that raises while importing is now a sentence
naming the file rather than an importlib traceback.
**`run` and `sweep` sync downwards too**, so the flow is found from the
repository root without the sync-then-`--no-sync` two-step. A study that will
not import is a warning rather than a stopped run, since a walk meets every
study and a half-finished one two directories away is not this run's problem.
The upload was already a no-op for a flow nothing changed in, so what the walk
costs is import time — `--sync PATH` narrows it, and skipping unchanged
subtrees would need a cache keyed on file state that is deliberately not here.
**`serve` moves off a busy default port** — 8001, 8002, up to twenty — says
which it took, and writes that one into `client.json`. A port given with
`--port` still fails when it is taken, because naming one is asking for it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9Hdrmf2cwNABCnE5x9UJa
Comparing the per-node digest against an engine that does not record one is
comparing against nothing, and reporting every node as changed on every sync
for ever — which is what a client newer than its engine did, since `NodeDef`
drops fields it has never heard of. A node is named now only when both sides
carry a digest, so a no-op sync is `unchanged` again and the signal one syncs
for is back.
That silence had also been the only sign of the mismatch, so sync now names
it: one line saying the engine stored no record of what a node's code reaches,
with both versions in it and what to run. Bumped to 0.1.6 — the digest changed
the stored document's shape, and a version that does not move makes two
different engines indistinguishable, which is the thing it was made
load-bearing for a day ago.
`— draft` was printed whenever there was simply nothing to publish, which
reads as work left unfinished. It is said only when a draft is genuinely
there, and `— published` when one was.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9Hdrmf2cwNABCnE5x9UJa
`sync` follows each node function's imports through the project's own modules
— stopping at the standard library, at anything installed, and at Fluksio
itself, whose checkout would otherwise be most of every digest — and records
the file list with what it hashed to. The engine hashes those files again when
the run is claimed, so the fingerprint is live rather than a snapshot, and
falls back to what sync recorded when it cannot see them: a remote worker's
runs used to share one empty digest, and therefore one key.
Three things follow. Editing a helper a node calls into re-runs that node, as
before. Editing something the node never reaches no longer re-runs anything —
a notebook two directories away was invalidating every arm. And
`Run.code_digest` is now the hash of its nodes' digests, so it is neither
looser nor tighter than "the code behind these numbers", which is what makes
it worth joining an exported table on.
`sync` says so too: it compares the per-node digest against the stored one, so
a helper edit prints `train: updated (flow, fit)` instead of `unchanged`. The
digest is read when the document is built rather than when the flow is
declared, so a second `sync()` in one process sees an edit between them.
Also: `fluksio runs` shows only the inputs that differ from what the flow
declares, fitted to the terminal, so a flow taking a few kB of json no longer
wraps every line.
Every existing cache entry misses once — the fingerprint changed shape.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9Hdrmf2cwNABCnE5x9UJa
Three things the first export pass got wrong for a real study.
**Dotted paths.** A node returns a record, not a scalar — the numbers arrive
inside `final_metrics` — so `--metrics final_metrics.train_loss` yielded an
empty column and `--metrics final_metrics` yielded the whole record in one
cell. Both sides of the wide table now take dotted paths, and the defaults
reach the same depth: every number a result carries is a column named by its
path, and inputs are compared leaf by leaf, so two configurations differing in
one field give that field as the axis rather than two blobs that are merely
not equal. Lists stay whole — a curve belongs in the long table.
**`--list`.** Metric names are flow-qualified, so `--name train_loss` matched
nothing and said only that. `fluksio export metrics --list` prints the names
the selection carries, and an empty export made with `--name` points at it.
**A version to compare.** The CLI ships ahead of the engine and a stale one
answered a flat 404 with nothing anywhere in the API to tell how old it was.
The engine reports `version` on `/observability/summary`, `fluksio status`
prints it, and a 404 from export now names both versions — or says "older"
when the field itself predates the engine. Bumped to 0.1.5, which is what
makes the number worth reading.
Also formats `flow/metrics.py`, which had been committed unformatted and was
the last `ruff format --check` failure.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9Hdrmf2cwNABCnE5x9UJa
`fluksio export metrics` is the long table — a row per run, metric and step —
and `fluksio export runs` the wide one, a row per run with the inputs that
*vary* across the selection as columns beside its final numbers, status,
duration and the commit and digest of the code it ran. Both carry the run id
on every row, which is the join back to the run page and what makes an
exported file auditable. `Client.export_metrics`/`export_runs` answer the same
rows to a notebook.
The engine streams csv or jsonl from two routes declared above `/{run_id}`;
parquet is a client-side conversion behind the new `fluksio[parquet]` extra,
so nobody pays for pyarrow who does not want dtypes kept. The long export
reads each run through `_series`, so a cached node's curve comes with it, and
`--stride` thins each series rather than the concatenation of all of them.
Two things they needed on the way: `GET /runs` takes `?since=` and `?before=`,
so a long history pages by the last row's own timestamp instead of an offset
that shifts under it; and a read that reaches no engine now says so in half a
second rather than seven, because `runs`, `flavors`, `export` and an unwatched
`status` pass `retries=0`. Everything that submits keeps them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9Hdrmf2cwNABCnE5x9UJa
A batch run opens no cascade, and FlowRun was written only from
cascade_started — so `fluksio run` showed on /runs and in `fluksio status` and
was simply absent from Home. The collector now folds the run_started and
run_finished events RunService already published. Such a record is exempt from
the staleness sweep in both places: a training step of an hour is a normal one,
and only run_finished ends it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A bool was excluded from the history as "not a measurement", so a true/false
port had no curve in the node panel and none on an edge — only the word. It is
recorded as 0/1 now and drawn as steps, since a bezier through two states
slopes through readings that never happened. The axis is pinned to 0..1, so a
flag that was never on sits at the floor rather than mid-box.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The newest metric bucket is upserted every flush while its minute runs, and
both rollup endpoints summed it in — so every curve on Home ended on a fall
that was only the clock. The timeseries now stops at the last closed bin, and
the flow rollups' window ends on the last closed minute, so all sixty slices
are whole ones.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Concurrent runs sat at 4 whatever FLOW_MAX_CASCADES said: that setting bounds
cascades, and the run drivers read a hardcoded MAX_PARALLEL nobody could reach.
FLOW_MAX_RUNS is the knob they read now, --max-runs/--max-cascades/--max-workers
are the same three as flags on serve, and the engine says which numbers it
started with — which is the only way to tell that a settings file was read.
Events keep the run they happened in. The payload always carried it and the
persist path dropped it, so reading one run's failures meant filtering the
engine-wide list; a batch run's id reaches those events now too, since a run
has no journaled item to name itself by.
Also: a provisioner's 0 means "no deadline" rather than "cancel on the next
reconcile", and a command that reaches no engine says how to start one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015sbYeYaVgYQqm1sbx7wPdL
The timer thread promoted due work on a fixed one-second tick, so every
delayed item was 0-1000ms late whatever the load — measured on the house
at 705ms mean on a rollershutter stop, which is 2-4% of a 26-second
travel and accumulates in the position the motor node believes it is at.
It now sleeps to the soonest deadline and is woken when a nearer one is
scheduled, which measures 0.9ms end to end through Redis.
A promoted timer also went to the back of the queue. It goes into a due
lane of its own that `claim` reads first, so work that has waited out a
deadline is not held up by work that is merely queued.
Beside it, in the same code: seeding a message now bumps its version, so
a re-put flow's synchronous nodes no longer wait forever on a value that
is sitting in state; the consumer group drops the consumers of engines
that are gone (138 had accumulated on this installation); and the cast
that closes the long-standing `xclaim` mypy error.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Slurm is not a machine that attaches and stays; it is a queue somebody else
owns. So nothing here submits a node to it. It submits a job whose payload is an
ordinary worker dialling back in, and everything downstream — the protocol, the
artifacts, cancellation, the books — already worked and did not have to learn
what Slurm is.
The alternative, which Covalent takes, is to stage a serialized call and a
runner onto the login node, poll squeue and copy the result back: a second way
of running a node beside the one that exists. The cost of not doing that is one
assumption, that a compute node can open a connection outward. Where that is
false, _payload is the single method a staged variant would replace.
Clusters are configured in provisioners.json beside the alerts, since this is
infrastructure an operator writes rather than anything a flow says. The script
is generated with the system ssh and no new dependency, and prerun owns the
environment — deliberately no pip install, because what is on a cluster is
somebody's decision.
One outstanding request per profile, cancelled if it never attaches and on the
way out. Nothing autoscales.
The run gate needed the same hook: a run held before it starts never reaches the
placer's own wait, so it would have queued forever on a machine nothing had
asked for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A6HeySA27EkGANZN95QySW
Raw cpus and gpus are a property of the machines an installation has, so a node
written against a cluster quietly stops meaning anything when the cluster is
replaced. A node says "gpu-small" instead, and what that is stored here —
editable, and read again every time the node is built, so changing the flavor
changes what the next run gets.
Memory joins the schema properly (`ram`, in MB, accepting "2G"), along with
`duration_s` for how long a node is expected to take. That one is recorded and
shown and nothing else yet: a statement for whoever is planning around the node,
not a limit — the limit is still `timeout`.
A flavor and a number for the same thing is refused, compared by value so an
editor writing the whole object back with its defaults still round-trips. A name
nothing stores is refused at the save, which covers the canvas and `fluksio
sync` at once, and deleting one a node still asks for says which node.
Four sizes are seeded on an installation that has none, and never re-seeded:
re-adding one somebody deliberately removed is an argument nobody wins.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A6HeySA27EkGANZN95QySW
The engine answered "where does this node run" twice, in two ways that could
not see each other: a device sent it to a worker carrying that label, and
resources were counted against the engine's own cores. Declaring both meant the
second answer won and nothing was counted at all — which the data-science
getting-started page and the worked example both do.
One question now, in flow/placement.py: of every machine attached, which could
grant what this node asked for, and which of those has it free. The books move
onto each machine — one accountant per worker, built from the inventory it
reported — and the waiting moves above them, where one condition variable can
be woken by a release anywhere or by a worker attaching. Locks go one way:
placer, then a machine's books, never back.
So a node asking for a card now finds the box that has one, rather than being
clamped down to none and run here. When nothing can grant the ask at all it is
still cut down and run — a flow written on a cluster has to work on a laptop —
but the ceiling is one real machine now, since taking the largest of each
dimension separately can describe a machine nobody has.
Two things fixed on the way. A device on a connector node held every batch run
of its flow forever, waiting for a worker that could never run an entry point.
And `prefer` falling back to the engine skipped the books, so the fallback held
nothing.
The bench flow's node has taken a `params` argument that with_settings has not
forwarded for some time, so the benchmark could not run at all: 62 ms median
submit-to-result with this, against the 61 ms on record.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A6HeySA27EkGANZN95QySW
A worker reported its labels and nothing about the machine behind them, so the
engine could route a node to a GPU box but not tell whether that box had a GPU
free. Inventory — cores, GPUs, memory — now arrives with the hello frame, and
the run frame carries back what the engine allocated for that call.
Which is protocol 2 on both ends. GPUs are never probed: asking a vendor tool
would make the one dependency two, so a GPU is what the batch job says it was
given or what --gpus says. A worker that reports nothing still attaches and is
scheduled by its label alone.
Two things a job scheduler needs: --max-idle stops a worker started for one job
rather than letting it hold its allocation to the walltime, and a refusal is now
fatal instead of a reconnect loop that reads as a hang in a job's log.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A6HeySA27EkGANZN95QySW
A port may now declare `image`, `audio` or `video`. Each is the artifact
reference the engine already had, narrowed by the `media_type` on it, so a
speech recogniser declares what it eats rather than taking any bytes at all and
finding out. Bytes still never travel as a message and nothing on the wire
stops being JSON: a camera publishes one reference per frame, a microphone one
per chunk, and a reference may carry a `meta` dict nothing here interprets.
Streaming media is therefore an ordinary streaming port — with one change to
what that means. An emission used to journal an item with no payload, so
downstream read whatever was current when the item was claimed; a consumer
slower than its producer saw only the newest chunk and the ones between were
lost. That is right for a training curve and wrong for a second of speech, so
an emission now journals a `kind="emission"` item carrying its values, and the
executor hands them to the nodes reading that message instead of writing them
to state again. The value in state stays the latest, which is what everything
else reads, and the wave is filtered by what actually changed rather than
walking everything reachable. No queue serialization change — the existing
`outputs` field carries it.
Continuous media makes the store's missing GC a real problem, so this closes
it: `sweep_artifacts` runs hourly, keeps every digest a `run_artifact` row
records or a live message holds, spares anything written in the last hour, and
stands aside entirely while a run is in flight, since a node may store a
checkpoint long before it returns the reference to it. That also collects the
orphans a deleted flow has always left behind. `ARTIFACT_GC_INTERVAL_S=0` turns
it off.
Around the edges: `GET /artifacts/{digest}` serves the media type the caller
passes and answers ranged requests, so a browser plays a clip rather than
downloading it; `PUT` spools to disk instead of holding the whole body in
memory, as does `save_artifact` given a path; a Media widget draws whatever its
message points at, and a wall panel may fetch the bytes its own tiles are
showing and nothing else; and a connector gets `save_artifact`, for a device
whose readings are bytes.
What this cannot do is live video: a frame every second or two is a glance, and
the honest answer above that is the camera's own stream, which the widget takes
as a URL and the browser plays from source.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The pieces were tested one at a time, so removing the one line in
`_build_node` that wraps a worker-backed function with its sink broke nothing
visible — while losing every generator node's last yield and returning None
in its place.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The store answers the new-node template when nothing was ever written for a
node, so such a node ran — returning {} on every call, reporting active and
ok, and saying nothing anywhere. Unreachable through `fluksio sync`, which
writes every body before it publishes; the editor end was open.
A run of a flow holding one is now refused, and the flow carries a
missing_source issue so it is visible before anybody runs it. A draft is
exempt: a node being written legitimately has no published body yet.
The generated client is regenerated for the new issue code, which also
catches up the drift left by earlier backend work (resources, code_digest,
idempotency_key).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A run_artifact row is keyed by the message the bytes left on, and that was
also the only name it could answer with — so an `@run:` reference resolved
through the row was the same bytes under a name its producer never chose.
The row now records the file name beside the message name; rows written
before the column answer as they always did.
The fallback also checks the bytes are still in the store, which the bare
digest spelling beside it has always done. A missing blob now fails at
submit rather than in the middle of the run that wanted it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A cache key held qualified input names, so the same node reading the same
values through two flows keyed differently and only a node with no inputs
could ever hit across one. The fingerprint beside the key already says what
the node is, and it has been flow-agnostic since it moved ahead of
assign_flow — the names were the last thing tying an entry to one flow.
Inputs now reduce by the node's own name for them; a name belonging to
another flow keeps its prefix, since reading it is part of what the
execution is. Every stored entry misses once and is re-run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The worker held each yield one behind, because the last one is the node's
result when the generator returns nothing of its own. Only the engine knows
what ports a node declared, so the check happened when the *next* yield
arrived — a pass late, which for a training loop is however long one epoch
takes.
The worker now sends every yield as it happens and returns whatever its
generator returned; EmitSink holds the last one back and decides at the end
of the call what it was. Old "emit" frames are still handled, so a remote
agent that has not been restarted keeps working.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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>
Three faults with one root: the stored body of a code-defined node is an
import shim, and nothing that mattered was ever read from the code itself.
- The run stamp could not identify what ran. The shim imports whatever is on
disk when the worker starts, and an uncommitted tree stamps <commit>-dirty
for every run it ever produces. Run.code_digest hashes the repository's .py
files, memoized on their stat state, and it is read again when the run is
actually claimed -- so a sweep queued for hours records the code each of its
runs executed, not the code that was there when it was submitted.
- The stage cache adopted code that was too new. The fingerprint hashed the
shim, which is invariant under any edit to the imported function or anything
it calls into, so a re-run was served from cache and answered without the
outputs the edit added. It now carries the repo digest and the node's
declared ports. Every fingerprint changes once, which invalidates the
existing cache; a canvas flow has no repository and keys as before.
- An interrupted sync looked like a hand-edited canvas. The engine answers a
new-node template for a node with no stored body, and the template carries
no marker, so the drift check read "somebody edited this" and demanded
--force -- for the one state that re-running the sync is the fix for.
NodeSource.missing states the fact, and sync skips those and reuses the
bodies it read instead of asking for each one twice.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A driver script died of one slow answer: httpx.ReadTimeout out of
RunHandle.refresh() with a 30 s read timeout and no retry anywhere, which
cost a sweep 78 of its 84 runs.
- Split the timeout (5 s connect, 120 s read): a wrong URL fails at once,
and a busy engine gets longer than the slowest thing it does on purpose
(a 60 s compile, a 15 s rebuild wait).
- Retry idempotent calls three times on a transport error or 502/503/504.
503 is the engine's own "ask again" — it is what RebuildBusy answers.
- Submit carries a key the engine stores with the run, so a retry after a
timeout returns that run instead of starting a second. A sweep keys every
entry, so a half-created one recreates only what is missing.
- wait() and --follow tolerate five failed polls in a row; a 404 still stops
at once, because that is an answer rather than a gap.
- CLI says "engine not answering" and names the run still on the engine,
instead of printing a traceback.
- runs: clamp the params column to 80 characters; events() takes the
flow/since/until the endpoint already had; RunHandle.failures answers
"what killed this run" from the run's own node rows.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The engine was I/O-bound on its own state backend. `RedisState.lock()` is one
key — `pipeline:_lock` — for the whole process, taken five times a message at
two round trips each, and every cascade and every node read queued behind it.
Inside it, reading a node's inputs was three round trips per input (an EXISTS
for `in`, then EXISTS and GET for the value), writing was two updates that a
single transaction already gives, and the version counters went one INCR at a
time.
Replaced with the atomic command that was always available: `get_present` is
one MGET and tells a missing key from one holding null, so the lock it used to
be read under bought nothing; value and timestamp land in one `update`, which
is a MULTI/EXEC; `increment_multi` pipelines the counters. `values()` — what
every websocket snapshot calls — is two reads whatever the message count
instead of two per message.
Beside that: every webhook did its blocking XADD on the asyncio event loop
(MQTT already used `to_thread`); the per-execution `NodeOutcome` was built and
validated even with no run watching; `_minute` built a tz-aware datetime per
event on the loop thread to key a dict, and now keys on an int; `move_due`
promoted delayed items one round trip each, every second; `FLOW_MAX_CASCADES`
makes the in-flight ceiling a setting rather than a constant.
`orjson` replaces stdlib json where a message pays for it — state, the
journal, the engine side of the worker pipe. `fluksio-worker` stays
dependency-free, and the run-cache digest stays on stdlib so no stored key is
invalidated. A non-finite number now stores as `null` rather than the bare
`NaN` that was never JSON.
Measured with `scripts/bench_engine.py` against a real Redis, 200 messages:
a five-node chain went from 43.9 to 103.1 msg/s with p50 latency 2110ms →
782ms and p95 3913ms → 1439ms; one source into twenty consumers went from 5.4
to 33.7 msg/s. In memory, twenty consumers went from 187 to 448 msg/s.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BpfSinyCBfjuieikyfMPbf
The limit was applied in `apply_outputs`, which the executor reaches after the
item is off the queue — so a subscriber told to publish every 15s still cost a
queue entry, a `cascade_started`, a run record and a walk of everything
reachable from it per inbound message. Seven relay nodes behind one inverter
ran 192 times a minute to publish six.
Two halves, matching the two shapes it takes:
`trigger()` now keeps a value whose every port is inside its window and
journals nothing at all. The window split came out of `_throttled` as a
read-only `_window_split`, so the question is asked the same way in both
places and the exact split is still made once, at claim time.
A cascade carries the names it actually published, and the wave runs only the
nodes something in that set feeds. A node whose triggering inputs were all
held back is completed without running, which frees its own consumers to be
judged the same way — the case where a node re-published 619 messages a minute
off inputs that changed six times. Redeliveries and emissions carry no such
set and still walk everything, since one has a half-finished wave to finish
and the other is the value already being in state.
Skipping a node can make one ready that the scheduling pass has already walked
past, so `submit_ready` runs to a fixpoint. That also closes the same latent
hole on the replay path, where a done-marker skip could strand a join with no
future outstanding to come back for it.
Measured with the new `scripts/bench_engine.py`, 500 messages through the
house's shape: a limited source went from 500 cascades / 3500 node runs /
5009 events to 1 / 7 / 19, publishing the same 8 values; an unlimited source
into limited relays took the node reading them from 500 runs to 1.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BpfSinyCBfjuieikyfMPbf
`RedisWorkQueue.stats` read XPENDING, which counts entries delivered to a
consumer and not yet acknowledged — work in progress. Entries sitting in the
stream undelivered were counted nowhere, so an engine hours behind reported
itself idle: on the house, `pending: 4` while the group's lag was 1554.
The group's own `lag` is the missing number. `backlog` now carries it on both
queues (`len(_items)` in memory), leads the health tile, and a sustained one
publishes `engine_degraded` from the timer thread — named with the flow most
of the waiting work belongs to, sampled from the undelivered tail, since that
is the actionable half. It is a summary problem rather than a /utils/health
503: a backlog should not restart the container.
Also drops the keyspace `scan_iter` `stats()` did per poll to count parked
items — it walked every state and idempotency key twice per ten seconds — for
a set the park/unpark path maintains.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BpfSinyCBfjuieikyfMPbf
stroblmeandClaude Opus 5