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
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>
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
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
poll() still calls a child that has exited but not been reaped alive, so a
node that ended itself with os._exit left its corpse on the idle queue and
the next call died on a broken pipe. Mark it the way the out-of-step paths
already do, and _release replaces it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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>
stroblmeandClaude Opus 5