[
https://issues.apache.org/jira/browse/CAMEL-26872?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andrea Cosentino updated CAMEL-26872:
-------------------------------------
Fix Version/s: 4.25.0
(was: 4.24.0)
> camel-opa - manage security policy evaluators across context shutdown and
> restart
> ---------------------------------------------------------------------------------
>
> Key: CAMEL-26872
> URL: https://issues.apache.org/jira/browse/CAMEL-26872
> Project: Camel
> Issue Type: Bug
> Reporter: Andrea Cosentino
> Assignee: Andrea Cosentino
> Priority: Major
> Labels: camel-opa, lifecycle
> Fix For: 4.25.0
>
>
> OpaSecurityPolicy does not hand its evaluator's cleanup to CamelContext
> shutdown, despite the intention documented in beforeWrap().
> *Current behavior and cause*
> * OpaSecurityPolicy.beforeWrap() creates and caches an OpaPolicyEvaluator,
> then calls route.getCamelContext().addService(evaluator).
> * OpaPolicyEvaluator does not implement org.apache.camel.Service.
> OpaRestEvaluator and OpaWasmEvaluator implement AutoCloseable only.
> * AbstractCamelContext.addService() delegates to
> InternalServiceManager.doAddService(). Its lifecycle registration and
> deferred-start logic run only for objects implementing Service. Passing this
> evaluator therefore does not register its close() method.
> * The policy-owned REST HTTP transport is not explicitly closed on context
> shutdown. The WASM instance pool is not closed either. This is a component
> lifecycle/resource-ownership bug; the pool-race fix in CAMEL-26830 / PR
> #29633 should remain separate.
> *Reproduction / regression-test outline*
> # Configure an OpaSecurityPolicy with a component-owned REST client (no
> injected opaClient), healthCheckEnabled=false, and a local OPA-compatible
> HTTP fixture; guard a route with .policy(policy).
> # Start the CamelContext and send an allowed exchange.
> # Stop the context and verify that the owned evaluator/transport was closed
> exactly once. Currently it is not registered as a managed Service and close()
> is not called.
> # Restart the same CamelContext and policy instance; the route must authorize
> again using fresh owned runtime resources.
> # Cover the equivalent WASM policy with a valid generated module, checking
> pool closure and successful evaluation after restart.
> *Restart pitfall*
> Simply registering an adapter that calls close() on stop is insufficient.
> beforeWrap() only builds the evaluator when evaluator == null, so restart
> otherwise retains a closed pool/transport. Clearing only evaluator is also
> insufficient in REST mode: buildEvaluator() stores the internally-created
> client in the configurable opaClient field. A subsequent build sees opaClient
> != null, reuses that already-owned/closed client, and creates a new
> OpaRestEvaluator with no transport ownership. Distinguish injected
> configuration from internally-owned runtime state, or reset the owned client
> and related state correctly when rebuilding.
> *Acceptance criteria*
> * Context shutdown releases owned REST transports and WASM pools exactly
> once; repeated stop/close remains safe.
> * Stop/start of the same context and policy bean rebuilds functional owned
> runtime resources, including the cached-owned-REST-client case above.
> * A policy shared by multiple routes keeps working when one route stops while
> another still uses it; lifecycle ownership does not close shared resources
> prematurely.
> * Distinct policies/routes have independent ownership: stopping or replacing
> one cannot close another policy's runtime.
> * An explicitly injected opaClient remains caller-owned and is not closed or
> silently replaced by policy shutdown/restart.
> * Startup failure cleans up any already-created owned resources; lifecycle
> hooks do not accumulate duplicate registrations.
> * Tests exercise real route/processor lifecycle with deterministic close
> assertions and an in-place context restart, using Awaitility/latches where
> needed instead of sleeps.
> *Related work*
> CAMEL-24751 introduced guarded-processor reference counting for
> readiness-check registration. Preserve that behavior. Coordinate this fix
> with CAMEL-26836: stopping a reload watcher is not equivalent to closing its
> owned evaluator and pool. Reloadable evaluators need the same
> lifecycle/restart guarantees.
> Observed during review of https://github.com/apache/camel/pull/29633.
> Validated against the current lifecycle dispatch and policy source; no new
> runtime reproduction or Maven run was performed for this follow-up analysis.
> _Codex on behalf of oscerd_
> Generated by Codex via OSS Helper
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)