feat: add spinloop gateway, a fleet's OpenAI-compatible front door - #170
Draft
outofcoffee wants to merge 8 commits into
Draft
feat: add spinloop gateway, a fleet's OpenAI-compatible front door#170outofcoffee wants to merge 8 commits into
outofcoffee wants to merge 8 commits into
Conversation
outofcoffee
added
documentation
A new `spinloop gateway` command serves a fleet.yaml behind one OpenAI-compatible endpoint: it answers /v1/models with what the fleet is running, routes completion requests with the fleet's own selector, wakes a node from its own Spinloop source when nothing is serving, swaps the caller's authorisation for the engine key its fleet entry names, and honours a new top-level `wake: on|off` setting in the fleet file. A Spinloop's FLEET may now name an endpoint (a URL) as well as a fleet file: the launch is pointed at it with the OpenAI-compatible prefix added where needed, authenticating with the endpoint's token through the launch's existing key chain, and `spinloop fleet route` answers it without querying or starting a node. A node wake now carries the Spinloop's own BASEURL into the pushed deploy config, translated to the engine's bind exactly as a local serve translates it. Without it a node with no preset woke onto the engine's default bind - loopback for llama.cpp - which the gateway then correctly refused to route to. examples/gateway-docker/ brings up two daemon nodes and two gateways (one waking, one with wake: off) in containers and is the gateway integration test, run per PR by a new CI job.
The status branch sources the serving facts (runner, model id, served name) from the environment's deploy config — the same source the stats reply reads — not the daemon, whose model is the on-disk weights path, not a name a router can match against. The fleet and remote views now name a model the same.
A running remote environment reported its state and model but no engine address: the control plane's published base url was dropped in the status mapping, so the gateway could list a remote node's model and then refuse to route a request to it. The status now carries the engine's host, port and path from that address, and routing resolves a reported engine host in place of the fleet file's (which names a remote node by environment alone).
A daemon reports running once the engine process exists, which is before llama.cpp has fetched and loaded weights — minutes on a large model, during which nothing is listening on the engine's port. Selection read only the state, so the gateway proxied there and every request failed on the dial. Nodes reporting not-ready are now skipped, and a node already loading the wanted model is waited for rather than passed over: it serves sooner than a cold start, and waking a second node would leave two engines up for one request. An absent readiness reading still routes, so older daemons and runners with no health-check convention are unaffected.
…ng per request Reading CPU costs a host command that takes longer the busier the host is — `top -l 1` on a loaded macOS machine takes several seconds — so collecting inside the metrics handler blocked it past the fleet client's 5s timeout, and a node rendered as unreachable on the dashboard while it was answering fine. The figures were also collected twice: once by the sampler for the retained history, and again on every request. One collection per tick now serves both readers, and /v1/metrics reports the last reading. Its errors move onto the sample, which is the only place a broken source can be reported from once the handler no longer collects. Two cadences change alongside. The catch-up interval now also requires a scrape target, so an engine whose runner exposes no metrics endpoint settles at the tick rather than running the host commands every second for counters that are never coming. The dashboard polls every 5s rather than 2s: it is one call per machine per tick, and the figures it draws are sampled every 15s.
/v1/models listed only what answered now, so a client could not see the models it could ask for. The gateway already resolves each node's own source to wake it, so it now lists that model beside what the running nodes report, each source read at most once in a short window. The list stays bounded by what the gateway can start: a running engine is never displaced, a remote environment is never woken by a request, and a fleet whose wake is off lists nothing beyond what runs.
…ures The readiness selection (skip a not-ready engine, wait for one already loading, name a pinned node still starting) belongs to add-fleet-gateway, so it goes into that change's choosing and routing requirements. The daemon metrics work is separate: the sampler's last system reading serves /v1/metrics and the catch-up interval bounds itself with a scrape target, so it gets its own change.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Serves a fleet under one OpenAI-compatible endpoint: an agent's Spinloop names the gateway's address in its
FLEET, and the agent needs only the gateway's token.Summary
spinloop gatewaycommand andinternal/gateway: one OpenAI-compatible endpoint in front of afleet.yaml./v1/modelslists what the fleet is running; completion requests are routed with the fleet's own selector; a node is woken from its own Spinloop source when nothing is serving, and the request is held until the engine answers; the caller's authorisation is swapped for the engine key the node's fleet entry nameswake: on|offsetting in the fleet file decides whether routing may start an engine; off refuses without starting anything, naming the node and thespinloop fleet startcommand that wouldFLEETmay now name an endpoint (a URL) as well as a fleet file: the launch is pointed at it, with the OpenAI-compatible prefix added where needed, and authenticates with the endpoint's token through the launch's existing key chain.spinloop fleet routeanswers an endpoint without querying or starting a nodeBASEURLinto the pushed deploy config, translated to the engine's bind exactly as a localservedoes. Without it a node with no preset woke onto the engine's own default bind — loopback for llama.cpp — which nothing but the node itself could reachexamples/gateway-docker/: two daemon nodes and two gateways over them (one waking, onewake: off) in containers. It is the gateway's integration test, run per PR by a new CI jobdocs/commands/gateway.md, the endpoint form ofFLEETindocs/spinloop-file.md, thewakesetting indocs/commands/fleet.mdImplementation details
internal/fleet; the handler reuses a fan-out reading taken within the last two seconds rather than fanning out per requestengine:override taking responsibility for reachability, is never selected — and when it is the only candidate, the failure says so