Module: Legion::Extensions::Llm::Discovery::Pipeline
- Includes:
- Helpers::Lex
- Defined in:
- lib/legion/extensions/llm/discovery/pipeline.rb
Overview
The shared discovery pipeline for EVERY lex-llm-* provider — the write
half of the inventory interface. A provider mixes it into its own
<Provider>::Runners::Discovery module and overrides ONLY the
genuinely provider-specific methods (fetch_raw_models, check_health,
build_offering_draft, build_callable, and — if its config keys differ —
catalog_base_url / auth_token / derive_physical_id). Everything else
(reconcile, claim, activate, probe, replace, weight publication, health
display, dormant weight tracking) is inherited.
module Vllm
module Runners
module Discovery
extend self
include Legion::Extensions::Llm::Discovery::Pipeline
def fetch_raw_models(instance_cfg:) = ...
def build_offering_draft(...) = ...
def build_callable(instance_cfg:) = ...
end
end
end
It is a STATELESS module in the LegionIO sense: it is not itself a live
runner and it lives OUTSIDE the framework-scanned runners/ dir (a
base in that dir is claimed by the builder as a live per-extension
runner). Each provider's <Provider>::Runners::Discovery is the live,
scanned runner; it is that module object which carries the per-instance
working state (states — the claim handles + last built drafts +
publish progress), one per provider. That state is the writer's working
state, NOT a second inventory: the Inventory::Registry remains the
single source of truth.
The base actor (Discovery::Actor) fires on the discovery interval and
dispatches runner_class.refresh each tick (manual, in-process — the
state and the Registry are both process-local, so the pipeline must run
on the owning node, not as a remote Runner task).
WEIGHT IS NOT COMPUTED HERE. Drafts are built identity-weighted; Inventory::WeightReconciler recomputes the write-time weight from live settings at publish (commit/activate). One weight owner, no duplication.
Defined Under Namespace
Classes: CatalogFetchFailure
Instance Method Summary collapse
- #activate_tracked_state(instance_id:, state:, probe_token:) ⇒ Object
- #activate_weight_snapshot(instance_id:, state:, offerings:, sequence:, probe_token:) ⇒ Object
- #auth_token(instance_cfg:) ⇒ Object
-
#build_callable(instance_cfg:) ⇒ Object
Build the provider's inference callable captured into the registry.
-
#build_offering_draft(instance_cfg:, instance_key:, model_id:, model_data:) ⇒ Object
Build the Inventory::OfferingDraft (evidence + metadata) for one model.
-
#build_offerings(instance_cfg:, instance_key:) ⇒ Object
── Offering build (weight-free draft; reconciler weights at publish) ─.
- #build_probe_enqueue(instance_id:) ⇒ Object
- #catalog_base_url(instance_cfg:) ⇒ Object
-
#check_health(instance_cfg:) ⇒ Object
Return an Inventory::ReadinessResult.
- #claim_and_activate_instance(name:, instance_cfg:) ⇒ Object
- #clear_settings_health(name:) ⇒ Object
-
#derive_physical_id(instance_cfg:) ⇒ Object
── Physical identity (secondary; dedup/diagnostics only) ───────────.
-
#discover_instances ⇒ Object
The provider module's public instance catalog — the single source of configured instances, shared with the dispatch path (legion-llm Call::Providers reads the same method).
-
#discovery_programming_error?(error) ⇒ Boolean
A programming bug in discovery must fail loud — swallowing it publishes ZERO offerings and makes an activated instance invisible.
- #dormant_weight_tracker ⇒ Object
-
#fetch_raw_models(instance_cfg:) ⇒ Object
Return an Array of raw model Hashes for this instance.
- #handle_reactive_probe(instance_id:, request:) ⇒ Object
- #health_path ⇒ Object
-
#model_id_from(model_data) ⇒ Object
The model's stable id within the catalog list.
-
#models_path ⇒ Object
OpenAI-compatible defaults — overridable per provider.
- #observe_dormant_weights ⇒ Object
- #offerings_equivalent?(previous, current) ⇒ Boolean
- #perform_readiness(instance_id:, state:, offerings:) ⇒ Object
- #physical_id_changed?(state, instance_cfg) ⇒ Boolean
- #provider_family ⇒ Object
- #provider_module ⇒ Object
-
#provider_namespace ⇒ Object
Resolved from the INCLUDING runner module's own name, because the pipeline is a mixed-in module: here
selfIS the provider's::Runners::Discovery module (a Module), so its name is self.name, notself.class.name. -
#publisher ⇒ Object
── Claim / activate ────────────────────────────────────────────────.
-
#reconcile_instances ⇒ Object
── Reconcile (each tick) ───────────────────────────────────────────.
-
#reconcile_weight_snapshot(instance_id:, state:, discovered_offerings:) ⇒ Object
── Weight publication (delegates to the shared reconciler) ─────────.
-
#refresh ⇒ Object
── Entrypoint (called by the actor's manual dispatch) ───────────────.
-
#refresh_instance(instance_id:, instance_cfg:) ⇒ Object
── Probe (refresh path) ────────────────────────────────────────────.
- #remove_all_instances ⇒ Object
- #remove_instance_state(instance_id) ⇒ Object
- #replace_if_changed(instance_id:, state:, instance_cfg:) ⇒ Object
- #replace_weight_snapshot(instance_id:, state:, offerings:, sequence:) ⇒ Object
- #report_probe_result(instance_id:, state:, probe_token:, readiness:) ⇒ Object
- #report_tracked_readiness(instance_id:, state:, probe_token:, readiness:) ⇒ Object
-
#reset_state! ⇒ Object
Test hook: drop this provider's working state for a fresh run.
- #run_cadence_probe(instance_id:, state:) ⇒ Object
- #state_mutex ⇒ Object
-
#states ⇒ Object
Per-provider working state, carried by the including runner module (one per provider).
- #tracked_unpublished?(instance_id:, state:) ⇒ Boolean
- #union_capabilities(instance_key) ⇒ Object
- #update_instance(name:, instance_cfg:) ⇒ Object
-
#write_instance_health(state) ⇒ Object
── Health display (write-back to settings, post-commit) ────────────.
Instance Method Details
#activate_tracked_state(instance_id:, state:, probe_token:) ⇒ Object
433 434 435 436 437 438 439 440 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 433 def activate_tracked_state(instance_id:, state:, probe_token:) Legion::Extensions::Llm::Inventory::WeightReconciler.activate_tracked!( settings: Legion::Settings, instance_id: instance_id, state_key: state.fetch(:name), state: state, states: states, mutex: state_mutex, probe_token: probe_token, activate: method(:activate_weight_snapshot), activation_sequence: ->(tracked) { tracked.fetch(:sequence) } ) end |
#activate_weight_snapshot(instance_id:, state:, offerings:, sequence:, probe_token:) ⇒ Object
426 427 428 429 430 431 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 426 def activate_weight_snapshot(instance_id:, state:, offerings:, sequence:, probe_token:) publisher.activate_instance_snapshot( instance_id: instance_id, publisher_token: state.fetch(:publisher_token), offerings: offerings, sequence: sequence, probe_token: probe_token ) end |
#auth_token(instance_cfg:) ⇒ Object
149 150 151 152 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 149 def auth_token(instance_cfg:) token = instance_cfg.dig(:credentials, :api_key) || instance_cfg[:api_key] token if token.is_a?(String) && !token.strip.empty? end |
#build_callable(instance_cfg:) ⇒ Object
Build the provider's inference callable captured into the registry. Abstract: a provider MUST implement how it executes inference.
137 138 139 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 137 def build_callable(instance_cfg:) raise NotImplementedError, "#{name} must implement #build_callable" end |
#build_offering_draft(instance_cfg:, instance_key:, model_id:, model_data:) ⇒ Object
Build the Inventory::OfferingDraft (evidence + metadata) for one model. NO weight — the reconciler computes it at publish. Abstract: a provider MUST implement its capability/operation evidence knowledge.
131 132 133 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 131 def build_offering_draft(instance_cfg:, instance_key:, model_id:, model_data:) raise NotImplementedError, "#{name} must implement #build_offering_draft" end |
#build_offerings(instance_cfg:, instance_key:) ⇒ Object
── Offering build (weight-free draft; reconciler weights at publish) ─
452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 452 def build_offerings(instance_cfg:, instance_key:) models = fetch_raw_models(instance_cfg: instance_cfg) models.filter_map do |model_data| model_id = model_id_from(model_data) next if model_id.empty? build_offering_draft(instance_cfg: instance_cfg, instance_key: instance_key, model_id: model_id, model_data: model_data) end rescue StandardError => e raise e if discovery_programming_error?(e) raise e if e.is_a?(CatalogFetchFailure) handle_exception(e, level: :warn, handled: false, operation: "#{provider_family}.runner.discovery.build_offerings") raise CatalogFetchFailure, "catalog build failed (#{e.class.name})", cause: e end |
#build_probe_enqueue(instance_id:) ⇒ Object
396 397 398 399 400 401 402 403 404 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 396 def build_probe_enqueue(instance_id:) proc do |request:| handle_reactive_probe(instance_id: instance_id, request: request) true rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.probe_enqueue", instance_id: instance_id) false end end |
#catalog_base_url(instance_cfg:) ⇒ Object
145 146 147 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 145 def catalog_base_url(instance_cfg:) normalize_api_base(instance_cfg[:base_url] || instance_cfg[:endpoint]) end |
#check_health(instance_cfg:) ⇒ Object
Return an Inventory::ReadinessResult. Default is the OpenAI-compatible GET /health. A provider with a different readiness probe overrides it.
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 111 def check_health(instance_cfg:) conn = build_connection(base_url: catalog_base_url(instance_cfg: instance_cfg), instance_cfg: instance_cfg, timeout: 5, open_timeout: 3) response = conn.get(health_path) Legion::Extensions::Llm::Inventory::ReadinessResult.new( ready: response.status == 200, reason: "#{health_path} returned #{response.status}", metadata: { status: response.status } ) rescue Faraday::ConnectionFailed => e handle_exception(e, level: :warn, handled: true, operation: "#{provider_family}.runner.discovery.health") readiness_failure(error: e) rescue StandardError => e raise e if discovery_programming_error?(e) handle_exception(e, level: :warn, handled: true, operation: "#{provider_family}.runner.discovery.health") readiness_failure(error: e) end |
#claim_and_activate_instance(name:, instance_cfg:) ⇒ Object
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 260 def claim_and_activate_instance(name:, instance_cfg:) instance_id = name.to_s instance_key = Legion::Extensions::Llm::Inventory::Identity::InstanceKey.new( provider_family: provider_family, instance_id: instance_id, physical_id: derive_physical_id(instance_cfg: instance_cfg) ) begin offerings = build_offerings(instance_cfg: instance_cfg, instance_key: instance_key) rescue CatalogFetchFailure log.debug { "#{provider_family} discovery: catalog fetch failed at claim for #{instance_id} — deferring readiness" } offerings = nil end callable = build_callable(instance_cfg: instance_cfg) probe_coordinator = Legion::Extensions::Llm::Inventory::ProbeCoordinator.new( instance_key: instance_key, enqueue: build_probe_enqueue(instance_id: instance_id) ) publisher_token = publisher.claim_instance( instance_id: instance_id, physical_id: instance_key.physical_id, callable: callable, probe_request_handle: probe_coordinator ) state = { name: instance_id, instance_key: instance_key, instance_cfg: instance_cfg, callable: callable, probe_coordinator: probe_coordinator, publisher_token: publisher_token, sequence: 0, offerings: offerings || [] } Legion::Extensions::Llm::Inventory::WeightReconciler.track_initializing!( states: states, state_key: instance_id, state: state, mutex: state_mutex ) perform_readiness(instance_id: instance_id, state: state, offerings: offerings) if offerings end |
#clear_settings_health(name:) ⇒ Object
494 495 496 497 498 499 500 501 502 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 494 def clear_settings_health(name:) instance_settings = settings.dig(:instances, name.to_sym) return unless instance_settings.is_a?(Hash) instance_settings.delete(:health) instance_settings.delete(:capabilities) rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.clear_health", instance_name: name.to_s) end |
#derive_physical_id(instance_cfg:) ⇒ Object
── Physical identity (secondary; dedup/diagnostics only) ───────────
516 517 518 519 520 521 522 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 516 def derive_physical_id(instance_cfg:) host_port = extract_host_port(url: catalog_base_url(instance_cfg: instance_cfg)) token = auth_token(instance_cfg: instance_cfg) return "#{host_port}/ak:#{::Digest::SHA256.hexdigest(token)[0, 6]}" if token host_port end |
#discover_instances ⇒ Object
The provider module's public instance catalog — the single source of configured instances, shared with the dispatch path (legion-llm Call::Providers reads the same method). A provider that defines no discover_instances does not populate the inventory.
197 198 199 200 201 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 197 def discover_instances return {} unless provider_module.respond_to?(:discover_instances) provider_module.discover_instances end |
#discovery_programming_error?(error) ⇒ Boolean
A programming bug in discovery must fail loud — swallowing it publishes ZERO offerings and makes an activated instance invisible.
470 471 472 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 470 def discovery_programming_error?(error) error.is_a?(NameError) || error.is_a?(ArgumentError) end |
#dormant_weight_tracker ⇒ Object
79 80 81 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 79 def dormant_weight_tracker @dormant_weight_tracker ||= Legion::Extensions::Llm::Inventory::DormantWeightTracker.new end |
#fetch_raw_models(instance_cfg:) ⇒ Object
Return an Array of raw model Hashes for this instance. Default is the OpenAI-compatible GET /v1/models -> body. A non-OpenAI provider overrides this entirely.
94 95 96 97 98 99 100 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 94 def fetch_raw_models(instance_cfg:) conn = build_connection(base_url: catalog_base_url(instance_cfg: instance_cfg), instance_cfg: instance_cfg, timeout: 15, open_timeout: 5) response = conn.get(models_path) raise CatalogFetchFailure, "catalog fetch returned HTTP #{response.status}" unless response.status.between?(200, 299) Array(Legion::JSON.load(response.body).fetch(:data, [])) end |
#handle_reactive_probe(instance_id:, request:) ⇒ Object
376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 376 def handle_reactive_probe(instance_id:, request:) state = state_mutex.synchronize { states[instance_id] } return unless state coordinator = state[:probe_coordinator] return unless coordinator.begin_probe(request: request) probe_token = publisher.readiness_probe_started(instance_id: instance_id, publisher_token: state[:publisher_token]) readiness = check_health(instance_cfg: state[:instance_cfg]) coordinator.finish_probe(request: request) report_probe_result(instance_id: instance_id, state: state, probe_token: probe_token, readiness: readiness) rescue StandardError => e begin coordinator.finish_probe(request: request) rescue StandardError => finish_e handle_exception(finish_e, level: :warn, handled: true, operation: "#{provider_family}.runner.discovery.reactive_probe.finish", instance_id: instance_id) end handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.reactive_probe", instance_id: instance_id) end |
#health_path ⇒ Object
143 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 143 def health_path = '/health' |
#model_id_from(model_data) ⇒ Object
The model's stable id within the catalog list. OpenAI-compatible /v1/models lists carry it under :id; Ollama's /api/tags carries it under :name.
105 106 107 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 105 def model_id_from(model_data) model_data[:id].to_s end |
#models_path ⇒ Object
OpenAI-compatible defaults — overridable per provider.
142 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 142 def models_path = '/v1/models' |
#observe_dormant_weights ⇒ Object
442 443 444 445 446 447 448 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 442 def observe_dormant_weights Legion::Extensions::Llm::Inventory::WeightReconciler.observe_dormant!( settings: Legion::Settings, provider_family: provider_family, states: states, mutex: state_mutex, tracker: dormant_weight_tracker, dormant_logger: ->(key) { log.info("[llm][#{provider_family}] action=dormant_weight weight_key=#{key.inspect} no_lane_published=true") } ) end |
#offerings_equivalent?(previous, current) ⇒ Boolean
474 475 476 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 474 def offerings_equivalent?(previous, current) offering_comparison_multiset(previous) == offering_comparison_multiset(current) end |
#perform_readiness(instance_id:, state:, offerings:) ⇒ Object
291 292 293 294 295 296 297 298 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 291 def perform_readiness(instance_id:, state:, offerings:) reconcile_weight_snapshot(instance_id: instance_id, state: state, discovered_offerings: offerings) probe_token = publisher.readiness_probe_started(instance_id: instance_id, publisher_token: state[:publisher_token]) readiness = check_health(instance_cfg: state[:instance_cfg]) report_probe_result(instance_id: instance_id, state: state, probe_token: probe_token, readiness: readiness) rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.readiness", instance_id: instance_id) end |
#physical_id_changed?(state, instance_cfg) ⇒ Boolean
234 235 236 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 234 def physical_id_changed?(state, instance_cfg) state[:instance_key].physical_id != derive_physical_id(instance_cfg: instance_cfg) end |
#provider_family ⇒ Object
185 186 187 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 185 def provider_family @provider_family ||= provider_namespace.split('::').last.to_s.downcase.to_sym end |
#provider_module ⇒ Object
189 190 191 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 189 def provider_module @provider_module ||= Kernel.const_get(provider_namespace) end |
#provider_namespace ⇒ Object
Resolved from the INCLUDING runner module's own name, because the
pipeline is a mixed-in module: here self IS the provider's
self.name, not self.class.name.
"Legion::Extensions::Llm::Vllm::Runners::Discovery" -> "Legion::Extensions::Llm::Vllm"
181 182 183 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 181 def provider_namespace @provider_namespace ||= name.split('::')[0..-3].join('::') end |
#publisher ⇒ Object
── Claim / activate ────────────────────────────────────────────────
256 257 258 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 256 def publisher @publisher ||= Legion::Extensions::Llm::Inventory::Publisher.new(provider_family: provider_family) end |
#reconcile_instances ⇒ Object
── Reconcile (each tick) ───────────────────────────────────────────
205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 205 def reconcile_instances configured = discover_instances configured_ids = configured.keys.map(&:to_s) tracked_ids = state_mutex.synchronize { states.keys } tracked_ids.each do |instance_id| remove_instance_state(instance_id) unless configured_ids.include?(instance_id) end configured.each do |name, instance_cfg| update_instance(name: name.to_s, instance_cfg: instance_cfg) rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.instance", instance_name: name.to_s) end observe_dormant_weights end |
#reconcile_weight_snapshot(instance_id:, state:, discovered_offerings:) ⇒ Object
── Weight publication (delegates to the shared reconciler) ─────────
408 409 410 411 412 413 414 415 416 417 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 408 def reconcile_weight_snapshot(instance_id:, state:, discovered_offerings:) Legion::Extensions::Llm::Inventory::WeightReconciler.commit_if_changed!( settings: Legion::Settings, instance_id: instance_id, state: state, discovered_offerings: discovered_offerings, mutex: state_mutex, equivalent: lambda do |previous, current| !states[instance_id].equal?(state) || offerings_equivalent?(previous, current) end, replace: method(:replace_weight_snapshot) ) end |
#refresh ⇒ Object
── Entrypoint (called by the actor's manual dispatch) ───────────────
156 157 158 159 160 161 162 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 156 def refresh(**) reconcile_instances { success: true } rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.refresh") { success: false } end |
#refresh_instance(instance_id:, instance_cfg:) ⇒ Object
── Probe (refresh path) ────────────────────────────────────────────
332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 332 def refresh_instance(instance_id:, instance_cfg:) state = state_mutex.synchronize { states[instance_id] } return unless state status = publisher.snapshot.publication_status(instance_key: state[:instance_key]) if status.state == :initializing begin offerings = build_offerings(instance_cfg: instance_cfg, instance_key: state[:instance_key]) rescue CatalogFetchFailure log.debug { "#{provider_family} discovery: catalog fetch failed for #{instance_id} — staying :initializing" } return end perform_readiness(instance_id: instance_id, state: state, offerings: offerings) return end replace_if_changed(instance_id: instance_id, state: state, instance_cfg: instance_cfg) run_cadence_probe(instance_id: instance_id, state: state) end |
#remove_all_instances ⇒ Object
164 165 166 167 168 169 170 171 172 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 164 def remove_all_instances(**) tracked = state_mutex.synchronize { states.keys } tracked.each { |instance_id| remove_instance_state(instance_id) } state_mutex.synchronize do states.clear dormant_weight_tracker.clear! end { success: true } end |
#remove_instance_state(instance_id) ⇒ Object
238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 238 def remove_instance_state(instance_id) state = state_mutex.synchronize do tracked = states[instance_id] next unless tracked publisher.remove_instance(instance_id: instance_id, publisher_token: tracked[:publisher_token]) states.delete(instance_id) if states[instance_id].equal?(tracked) tracked end return unless state clear_settings_health(name: state[:name]) rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.remove_instance", instance_id: instance_id) end |
#replace_if_changed(instance_id:, state:, instance_cfg:) ⇒ Object
352 353 354 355 356 357 358 359 360 361 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 352 def replace_if_changed(instance_id:, state:, instance_cfg:) begin new_offerings = build_offerings(instance_cfg: instance_cfg, instance_key: state[:instance_key]) rescue CatalogFetchFailure log.debug { "#{provider_family} discovery: catalog fetch failed for #{instance_id} — keeping last snapshot" } return end changed = reconcile_weight_snapshot(instance_id: instance_id, state: state, discovered_offerings: new_offerings) write_instance_health(state) if changed end |
#replace_weight_snapshot(instance_id:, state:, offerings:, sequence:) ⇒ Object
419 420 421 422 423 424 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 419 def replace_weight_snapshot(instance_id:, state:, offerings:, sequence:) publisher.replace_instance_snapshot( instance_id: instance_id, publisher_token: state.fetch(:publisher_token), offerings: offerings, sequence: sequence ) end |
#report_probe_result(instance_id:, state:, probe_token:, readiness:) ⇒ Object
300 301 302 303 304 305 306 307 308 309 310 311 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 300 def report_probe_result(instance_id:, state:, probe_token:, readiness:) committed = if readiness.ready? && tracked_unpublished?(instance_id: instance_id, state: state) activate_tracked_state(instance_id: instance_id, state: state, probe_token: probe_token) else report_tracked_readiness(instance_id: instance_id, state: state, probe_token: probe_token, readiness: readiness) end write_instance_health(state) if committed committed rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.report_probe", instance_id: instance_id) false end |
#report_tracked_readiness(instance_id:, state:, probe_token:, readiness:) ⇒ Object
317 318 319 320 321 322 323 324 325 326 327 328 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 317 def report_tracked_readiness(instance_id:, state:, probe_token:, readiness:) state_mutex.synchronize do return false unless states[instance_id].equal?(state) if readiness.ready? publisher.readiness_succeeded(instance_id: instance_id, probe_token: probe_token) else publisher.readiness_failed(instance_id: instance_id, probe_token: probe_token, reason: readiness.reason) end true end end |
#reset_state! ⇒ Object
Test hook: drop this provider's working state for a fresh run.
84 85 86 87 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 84 def reset_state! @states = nil @dormant_weight_tracker = nil end |
#run_cadence_probe(instance_id:, state:) ⇒ Object
363 364 365 366 367 368 369 370 371 372 373 374 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 363 def run_cadence_probe(instance_id:, state:) coordinator = state[:probe_coordinator] return unless coordinator.begin_probe probe_token = publisher.readiness_probe_started(instance_id: instance_id, publisher_token: state[:publisher_token]) readiness = check_health(instance_cfg: state[:instance_cfg]) coordinator.finish_probe report_probe_result(instance_id: instance_id, state: state, probe_token: probe_token, readiness: readiness) rescue StandardError => e finish_probe_safely(coordinator) handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.cadence_probe", instance_id: instance_id) end |
#state_mutex ⇒ Object
75 76 77 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 75 def state_mutex @state_mutex ||= Mutex.new end |
#states ⇒ Object
Per-provider working state, carried by the including runner module (one per provider). Lazy module-ivar accessors — a module has no constructor. The Inventory::Registry is the source of truth; this is only the writer's in-flight working state for this provider.
71 72 73 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 71 def states @states ||= Concurrent::Map.new end |
#tracked_unpublished?(instance_id:, state:) ⇒ Boolean
313 314 315 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 313 def tracked_unpublished?(instance_id:, state:) state_mutex.synchronize { states[instance_id].equal?(state) && !state.fetch(:published) } end |
#union_capabilities(instance_key) ⇒ Object
504 505 506 507 508 509 510 511 512 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 504 def union_capabilities(instance_key) capabilities = Set.new publisher.snapshot.lanes_for(instance_key: instance_key).each do |lane| lane.capability_evidence.each do |capability, evidence| capabilities << capability if evidence.supported? end end capabilities.to_a.sort end |
#update_instance(name:, instance_cfg:) ⇒ Object
222 223 224 225 226 227 228 229 230 231 232 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 222 def update_instance(name:, instance_cfg:) state = state_mutex.synchronize { states[name] } if state && physical_id_changed?(state, instance_cfg) remove_instance_state(name) claim_and_activate_instance(name: name, instance_cfg: instance_cfg) elsif state refresh_instance(instance_id: name, instance_cfg: instance_cfg) else claim_and_activate_instance(name: name, instance_cfg: instance_cfg) end end |
#write_instance_health(state) ⇒ Object
── Health display (write-back to settings, post-commit) ────────────
480 481 482 483 484 485 486 487 488 489 490 491 492 |
# File 'lib/legion/extensions/llm/discovery/pipeline.rb', line 480 def write_instance_health(state) instance_key = state[:instance_key] snapshot = publisher.snapshot display = display_fact(instance_key: instance_key, snapshot: snapshot) instance_settings = settings.dig(:instances, state[:name].to_sym) return unless instance_settings.is_a?(Hash) instance_settings[:health] = health_hash(display) instance_settings[:capabilities] = union_capabilities(instance_key) rescue StandardError => e handle_exception(e, level: :warn, operation: "#{provider_family}.runner.discovery.write_health", instance_id: state[:instance_key].instance_id) end |