run_coroutine_threadsafe refused it and it was never awaited, a RuntimeWarning
under pytest 9. The cancel still logs that it could not reach the worker.
Its entries are the cards present, and a local child's allocation is
named through them, so under =2,3 card 0 is device 2. Remote workers
keep mapping against their own mask.
DELETE /flows/library/{name} still answers 409 while nodes use the source.
With detach=true each of them keeps a copy of the code, in the published
flow and the draft, the source is deleted and the running users rebuild.
mqtt and http take websocket's vocabulary: their Mode is LISTENER/SENDER,
default names read mqtt_listener_… and http_listener_…, and mqtt's
publish_queue_size is send_queue_size. No alias stays behind, so the flow
store rewrites the old setting in every published and draft mqtt node when it
opens, in the same migration as trigger → debounce. Every mqtt and http
setting carries a one-line description for the editor, and the node-types
reference uses one wording for direction and says that an http sender
publishes a JSON object response on its outputs.
A push session that failed before it listened left only a task_crashed
event. The wrapper in _run_supervised now reports down with the exception,
unless the loop already said why it is down.
The store rewrites `"type": "trigger"` to `debounce` in every published and
draft flow when it opens, bumps the version and commits once. No alias stays
behind, so `trigger` is now an unknown type.
ws:// and wss:// URLs dial out and hold a supervised connection; a path is
served under /hooks/<flow>/ with an optional secret. Outputs only listen,
inputs send. Frames are JSON keyed by port, with json_key and on_connect.
The generated batch script exported nothing and passed `--token`, which
any user on a shared compute node reads from `ps`. It now exports
FLUKSIO_WORKER_TOKEN after `prerun`, which the worker already reads.
fluksio-worker now refuses ws:// and http:// engine and artifact URLs unless
they are loopback or --insecure is given; the engine still accepts both.
An artifact reference must be sha256: plus 64 hex digits before it is joined
into a path, and node code gets the host environment through the engine's
credential denylist instead of all of it.
The session db fixture now disposes the engine before deleting the file, since a
module under a no-op db (tests/sdk, tests/flow) may have connected first; the
sdk import tests forget the study packages they import; the supervision tests
keep the real quarantine length so it cannot expire mid-assertion.
A worker over FLOW_WORKER_MAX_RSS_MB (--worker-max-rss) when its node
returns is retired instead of kept warm, in the shared pool and in every
declared one; 0, the default, keeps them all. The engine now accounts its
own memory (FLOW_RAM / --ram, else the box or the cgroup limit less 1 GB),
so a node's `ram` declaration holds locally. A declared environment the
shared pool already has is that pool rather than a duplicate set of warm
workers.
A connector that misses one poll turned the health badge amber and the node
on the canvas red until it recovered. `LoadedNode.health_since` marks when the
status last actually changed — a rotating error detail under an unchanged
status does not move it, or a connector retrying with a different errno each
poll would never debounce — and `unhealthy_nodes` applies a 60 s floor to it.
The floor is on the surfacing, not on the transition: the stored health is the
truth the moment a node reports it, so a flow reading node health is never
told a stale story. Only going down is held back; recovery clears at once.
Home's "recently modified" order was a proxy — drafts first, then the
version counter, then the name. Both summaries now carry `updated_at`,
read as the mtime of the working copy: every write in the store commits
immediately, so a file's mtime is its commit time, and `git log -1 --
<path>` costs ~990ms across this instance's 15 documents (it walks the
history back to the last commit touching each one, so it is slowest for
the stalest) against ~1.8ms for the stats. No cache needed.
`DashboardSummary` also carries `footprint`: each widget as its type plus
the placement a panel resolves, which is the whole of what the mosaic
draws. That drops the document fetch per tile and with it the cap of
eight, past which tiles showed a name and nothing else. Verified against
this instance's 8 dashboards: the blocks are identical to what the old
client-side derivation produced.
A python node's ports and settings arrive as keyword arguments, so a declared
name its `process` does not take was a TypeError on every call — and a node
that loads fine and fails every time it runs is the quiet kind of broken: the
hosted demo did it 720 times an hour for two days and the health badge read
ok throughout. `_build_node` now reads a written body with `ast` and refuses
the mismatch at load, so the node is an error on the canvas and an issue on
publish. Skipped for `**kwargs`, a decorated or absent `process`, and the
template a new node opens with. The SDK's generated shim always takes
`**settings`, so synced flows are untouched.
`/observability/summary` names a node that has failed in the last fifteen
minutes and reads degraded while it does, which is what would have made the
badge amber. `nodes.failing` carries the count.
Each was a loose end recorded under `### SDK` in the notepad.
`serve` takes its own pidfile down on SIGTERM. uvicorn restores the handler it
found and re-raises the signal it stopped on, so the default handler ended the
process without unwinding and the `finally` never ran — which is what a stop
sends, and what left `serve.pid` behind.
`serve.log` is cut back past 5 MB by the engine rather than by the screen that
started it, so an adopted engine is bounded too. Gated on its own stdout being
an appended regular file, which is what makes the cut safe: the kernel then
puts the next write at the new end.
Cards are counted from `/dev/nvidia[0-9]*`, so `FLOW_GPUS`/`--gpus` of 0 means
"work it out" the way `FLOW_CPUS` always has. The engine counts, not the
accountant — a remote worker builds one of those from its own inventory, and
detecting there would hand it the engine host's cards. The worker counts last:
what a batch job says it was granted still wins.
`GET /runs/metrics/names` is the distinct over a selection that `--list` and
the terminal's metric picker were approximating by reading the newest run that
had measured anything, which missed a name only an older run ever wrote.
`MetricSink` announces each batch it has written (`run_metric`, carrying the
names). Not a per-point event: one covers up to 500 points or two seconds of
them, and the rows stay the record. The terminal comparison fills in as the
first readings land instead of staying blank until reopened, and the browser
refetches the run and any comparison rather than the list behind them.
`retry --group` pages the list route by `before` instead of stopping at 500.
The terminal dashboard takes the terminal's colours (`ansi-dark`), and the web
UI can re-pair from Settings without disconnecting first.
The rate the media dtypes could carry was one frame every second or two: each
was a file on the data volume, an event on the socket, and a request back for
the bytes. This closes both halves of that, and they are one feature.
`save_artifact(..., volatile=True)` writes to a `VolatileStore` — the same
content-addressed store, in `/dev/shm`, bounded by size with the oldest falling
out (`ARTIFACT_VOLATILE_BYTES`, 48 MB under the container's raised `shm_size`).
Nothing sweeps it: a frame nobody kept is not worth walking the store to find.
`ArtifactStore.path` falls through to it, which is what lets a volatile frame be
an ordinary reference everywhere else — the dtype check, a panel's digest scope,
`load_artifact` in a node, and the widget's own fetch all work on one unchanged.
`adopt` copies one into the store when a run records it, so "returned media is
kept, emitted media is not" stays true.
The bytes then go down the flows websocket as a length-prefixed binary frame,
sent just ahead of the `message_value` naming them, so a tile has the frame when
it hears the value moved. Nothing is pushed unasked: a client names the messages
it is drawing (`{"type":"media","names":[…]}`), a panel's list is intersected
with the scope it already had, and only the newest frame per name in a batch is
sent — a client that fell behind is not handed frames it would draw over. The
tunnel relays text only, so a screen reached through a portal falls back to
fetching, which is why the rate table now has two rows.
Around the edges: the remote worker's fetch cache is bounded at last
(`FLUKSIO_ARTIFACT_CACHE_BYTES`), since content addressing means nothing in it
ever expires and a media stream fills it with chunks nothing asks for twice; a
port carrying an image draws the frame in the node panel rather than only
saying `image/png · frame.png · 1.79kB`; and an edge chip says that much instead
of a line of hash. The media screenshot stops waiting for `networkidle` — a
camera is a socket that never goes quiet, which is the point of it.
Follows the portal: the noun is "instance" everywhere the app says it —
UI strings, CLI output, error details, docs and comments. The wire keys
(`instance_id`, `instance_token`) and the hub route this calls move with it.
An existing cloud.json is adopted rather than refused: without the key
alias the dataclass fails to parse, which the caller swallows and reads as
"never enrolled" instead of "reconnect".
`instance_key` on a node type becomes `target_key`. It means the outside
thing a node points at, which is a different sense of the word, and keeping
both would put two meanings of "instance" in one codebase.
The log tee wrote through to the real stream unguarded, and the worker
pool tees a returned call's logs there after reading its result and
before handing it back — so a dead stdout, which `fluksio serve` makes
possible by running the engine as a child of the dashboard holding that
pipe, failed the node with its outputs already in hand. The capture half
runs first, so swallowing the write loses nothing.
Also: `flow_events` catches the RuntimeError a peer leaving mid-send
raises, which is a disconnect by another route, and the remote agent no
longer raises out of the task when its subprocess died before it could
be written to — the read below reports that and ends the call.
The worker already sent it and the log panel already got it; the failure
outcome kept the one-line error and the node's stdout and dropped the
rest, so reading a failure back meant reproducing it under `run --local`.
It rides in the node's logs now — no schema change, and the API row, the
run detail page and `RunHandle.failures` carry it as they are.
`_record_node` keeps the tail of the log cap rather than the head, so a
chatty node cannot push the traceback past it, and `fluksio run` prints
what each node said when a run does not end ok.
`brain_graph` read `self.issues` alone, which is what a build found — so a
node that loaded and then lost its device was a well neuron on Home, and
that is what someone comes to this view to find.
The player is the one tile that both reads and publishes, so it has two
bindings: it shows a `record` describing what is playing — title, artist,
album, status, and position and duration in seconds — and publishes transport
words back to one `str` message (`toggle`, `next`, `prev`, `seek:<seconds>`).
Those are a streamer's own vocabulary rather than this app's, which is what
lets one tile drive whatever is on the other end.
The position counts forward in the browser between readings, so the bar moves
at one second while the device behind it is polled at whatever rate suits it;
every reading that arrives is taken as the truth and the count restarts there.
That is also why this is one record rather than five messages — a tile drawn
from five would redraw itself five times, and show a new title against the old
duration in between.
Being both is why `INPUT_WIDGETS` does not gain it: what that set means is "the
message this widget publishes is its only binding", which is exactly what a
player is not. Its reading is checked the usual way and its `target` separately.
The fader beside it needed nothing new. `ui/core` has had `orientation` on the
slider all along and all three looks draw it; only the widget never passed it,
so a volume control — the one thing reached for without looking, where up is
louder — could not be a column. Now it can, and the tile's height is the track.
A `webpush` alert channel, and the PWA it needs to arrive. The payload is
encrypted to the subscription (RFC 8291) and the request signed with this
installation's own keypair (RFC 8292), both over `http-ece` — `pywebpush`
does the same in one call but brings `requests` and `aiohttp` with it, two
HTTP stacks beside httpx on a machine that may be a Raspberry Pi.
The manifest and the worker are hand-written rather than `vite-plugin-pwa`:
there is nothing worth precaching when the page carrying the credential is
`no-store`, so the worker handles `push` and `notificationclick` and nothing
else. `registration.scope` is the app's root in both places it runs, which is
why the payload carries no URL.
A run finishing in error is the first event worth waking someone for; `ok`
and `cancelled` describe to nothing, so a nightly batch that works stays
quiet. The events were already on the bus — only the filter changed.
`WEBPUSH_FILE` is a derived path, so the keypair lands on the data volume
with the alerts beside it. Off it, a rebuild would silently stop every phone
being notified: the key they subscribed against would be gone.
The two sides both touched `submit_ready`'s readiness check, for unrelated
reasons, so the conflict is textual rather than semantic and both changes
stand:
- `e0962a4` completes a node that is not ready instead of passing over it,
so a producer that can never run stops stranding its consumers.
- the audit branch has `_is_node_ready` return the values it read, so the
node runs on them instead of asking state for the same keys again.
Merged as: read once, keep the values whether or not the answer is yes, and
take the not-ready branch from `e0962a4`. Its reasoning holds under the
merge — by the time readiness is consulted, `in_degree` is zero and every
in-wave producer has finished, so the answer cannot change later in the
wave.
Also fixes a fixture this branch added: the module-scoped row cleanup in
`tests/conftest.py` assumed a schema, and `tests/flow` overrides `db` with a
no-op because those tests need no database. It only showed when that
directory ran on its own.
746 tests green, and each directory green alone. Engine throughput is
unchanged by the merge (559 msg/s on the memory backend, against 639 before
it and 262 at the start of the audit).
A three-node cascade publishes 13-16 events and each one crossed to the
event loop on its own. They are one `call_soon_threadsafe` now — whatever
was published between two turns of the loop goes over together — and every
subscriber still receives every event, oldest still dropped first when one
falls behind.
The socket end of the same path:
- **any frame from the client ended its stream.** `receive_text` was
awaited once, outside the loop, so a keepalive — or anything else a
client decided to say — satisfied it and was read as the client going
away. It is recreated per iteration; only a disconnect ends the stream.
- events go out in one frame per wave (`{"type": "batch", "events": [...]}`,
capped at 64), serialised once with orjson rather than per client with
the stdlib's `json.dumps` through `send_json`. The client unpacks a batch
and still understands single frames, so an older engine behind a newer
bundle keeps working.
- authenticating and building the snapshot happen on a thread. Both were on
the event loop: one is a database round trip, the other reads the whole
of state, per connect and again per `dashboard_changed` per panel.
`Pipeline.values()` — what that snapshot is — no longer SCANs the whole
Redis namespace. It scanned five bookkeeping keys for every message to find
the messages; `RedisState` keeps a set of the names beside them and answers
from it. Maintained wherever a message is written, so a seeded value or a
deleted flow keeps it exact.
On the client, while in the same file:
- a `node_health` event invalidates the flow's detail. The canvas draws
health from the server-derived `issues`, so a node going down or
recovering only showed on mount, navigation or a rebuild. The store had
a health map of its own that nothing ever read; it and `useNodeHealth`
are gone rather than wired up, since the server's view is the one the
canvas already uses.
- a reconnect invalidates the five key families this socket feeds instead
of the entire cache, and the backoff is jittered. The usual reason a
socket dropped is the engine restarting, so every tab and every wall
panel refetched everything, together, at the moment it was least able to
answer.
- a frame that will not parse costs the frame, not the connection. It was
the one unguarded `JSON.parse` in the app; an exception there escaped to
`window.onerror` and left whatever it had already applied behind.
Measured with `make bench-engine` against a real Redis: 103.6 -> 164.4
messages a second on a five-node chain (p50 latency 2125 -> 1171 ms) and
34.8 -> 63.2 on a fan-out of twenty. Against the memory backend, which is
what a pip install runs on, 262 -> 626.
The two that bought most of it:
- `StateBackend.record` puts a published value, its timestamp, its series
and its version counter in one round trip. They were four calls building
four pipelines, and a value crossing an edge pays them twice. A released
rate-limit hold rides along instead of a DEL per port.
- the readiness check reads a node's inputs and hands them to the node,
rather than reading the triggering ones to count them and having the node
read the same keys again a moment later.
`apply_outputs` was a second copy of `_record_outputs` and is now the same
code plus the event that distinguishes it.
The rest, each small:
- `_derive` builds a node-by-id map and a `consumes` index, so dispatching
an item and publishing a value stop scanning every node in the
installation.
- `read_all` is memoised against the store revision — it sits on the
publish path, so a dashboard slider was reading and validating every
flow file per value. Same mechanism `_wiring` already uses.
- the `message_value` source block is built once per node instead of per
emission.
- both timer threads ask the queue to promote only when something is
actually due, which takes an idle engine from ~4 Redis round trips a
second to one.
- the shared httpx client is bounded (32 connections, one retry); its
default pool is 100 with no per-host cap, so one slow endpoint could
take it and every other sender node with it.
- the MQTT and delay nodes no longer log a line per message at INFO.
Robustness, in the same pass:
- `MemoryWorkQueue._done` was a set nothing ever removed from — one entry
per non-idempotent node per item, for the life of the process, in the
default configuration. Capped, the way the Redis side expires its
markers.
- a saturated engine can claim from the due lane past the cascade limit.
The capacity gate sits in front of the claim, so the due lane's priority
— decided inside it — did not apply while every slot was held: a motor's
stop was not behind the long nodes, it was unread. Only after a slot has
genuinely failed to free for half a second, and briefly, so the backlog
is not starved in turn.
- `reclaim_stale` dispatches through that same gate. It could return sixty
entries and push in-flight far past the limit the gate exists to hold.
- a flow's nodes are stopped together rather than one after another. Each
gets `NODE_STOP_TIMEOUT`, so a flow whose broker was unreachable took
five seconds per node — long enough to outlast `REBUILD_WAIT` and 503
the deploy.
- the worker pool and the HTTP client are closed on a thread, not on the
event loop, and a run closes the state backend it built (on Redis, a
client and a connection pool per run).
- the five background tasks say something when they die. Each catches
exceptions inside its loop, so one raised anywhere else left the engine
serving with no metrics, no alerts or no artifact sweep, silently.
`tests/flow/test_round_trips.py` counts the state operations one message
costs — four, where it was about eleven — because none of the above would
fail a behavioural test if it were undone.
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.
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.
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.
stroblme
Melvin Strobl