ALEC-305: Remove the Ludwig deep-learning engine

[joseanesONMS]

Tracked by ALEC-305.

Ludwig is retired. ALEC now offers exactly two correlation engines: DBSCAN clustering and the LLM-based clustering engine. This removes the TensorFlow/Ludwig-trained deep-learning engine and the standalone cluster engine (the UI's mislabeled "Deep Learning" option).

Removed

• Modules: engine/deeplearning + features/deeplearning (incl. ludwig_model.yaml and the bundled TensorFlow model saved_model.pb), and the now-orphaned wrap/tensorflow.
• Concrete cluster engine: ClusterEngine / ClusterEngineFactory. engine/cluster keeps AbstractClusterEngine and the graph/spatial infra that DBSCAN and the LLM engine extend.
• Wiring: Maven modules/deps (7 poms), Karaf features (alec-engine-deeplearning, alec-features-deeplearning, alec-dependency-tensorflow) and their inclusion in alec-opennms-standalone; the deeplearning/cluster branches in EngineRestImpl, EngineParameterImpl, ProcessAlarms; smoke-test parameters; the UI's ENGINE_DEEP_LEARNING constant and DeepLearning icon/label; and the orphaned tensorflow/json-simple/httpclient/string-similarity pom entries.

The UI's engine-status widget now shows CLUSTERING for DBSCAN and LLM for the LLM engine. AccountSettings already offered only DBSCAN/LLM, so no selector was orphaned.

Docs

• Removed the Ludwig deep-learning page and all references.
• Added a dedicated LLM Clustering Engine (experimental) reference page + nav entry, and clarified throughout that the LLM clustering engine (correlation — forms situations) is distinct from LLM Root Cause Analysis (explanation — describes a formed situation). The two share the LLM Setup but are enabled independently. Verified the Root Cause Analysis page is accurate against the current feature.

Verification

• mvn clean install — compile across all modules, features.xml resolution, and KAR assembly all succeed (BUILD SUCCESS), which validates the Karaf feature graph has no dangling references to the removed bundles.
• vue-tsc clean; 86 UI tests green; ui-ext bundle rebuilt (no "DEEP LEARNING" strings remain).

Given this touches the Karaf feature graph, I recommend a smoke deploy (ALEC boots and registers with only DBSCAN + LLM) before merge — happy to run it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UB6PGc2rpPbojnqbTHU5ND

[joseanesONMS]

@cgorantla — friendly re-ping on this one too. ALEC-305: remove the Ludwig deep-learning engine, leaving DBSCAN + LLM as the only correlation engines (per the decision that Ludwig is retired). AbstractClusterEngine and the graph/spatial infra stay; only the concrete ClusterEngine/ClusterEngineFactory and the whole deeplearning/tensorflow stack are removed.

• mvn clean install green (compile + features.xml resolution + KAR assembly); the meaningful CircleCI stages pass (only the pre-existing distributed-container smoke flake is red).
• Built and deployed to a local OpenNMS: ALEC boots and registers with only DBSCAN + LLM — no deeplearning/tensorflow bundles in the KAR or the running system.
• Docs updated: removed the Ludwig page and added a dedicated LLM Clustering Engine page, clarifying that LLM clustering (correlation) is distinct from LLM Root Cause Analysis (explanation).

This and #168 (the ALEC-301 LLM engine fixes, just verified live) are the last two open in the series — a review on both when you have a moment would let them land. Thanks!

[joseanesONMS]

@cgorantla friendly re-ping — this one's ready for review. ALEC-305: remove the Ludwig deep-learning engine, leaving DBSCAN + LLM as the only correlation engines.

The earlier smoke-test-commit failure was the known flaky DistributedRedundantCorrrelationTest.canCorrelateAlarmsAfterFailure (testcontainers ContainerLaunchException — container startup timeout), not a regression: every build stage passed. I've pushed an empty commit to re-run CI. Would appreciate your review when you have a moment.

[joseanesONMS]

Fixed the smoke failure in 61af07a0 — and it turned out not to be the chronic infra flake, but a real regression from this PR.

Root cause: removing the standalone ClusterEngineFactory left the smoke tests still deploying alec-engine-cluster, which now carries only AbstractClusterEngine and registers no EngineFactory OSGi service. So alec-driver-main couldn't resolve:

Unable to resolve org.opennms.alec.driver.main:
  missing requirement osgi.service; filter:="(objectClass=…EngineFactory)"

ALEC never started in the Sentinel container, so the test timed out as a testcontainers ContainerLaunchException — which looks identical to the infra flake but was deterministic (failed every run on this branch; #168's smoke, which keeps the cluster engine, passed).

Fix: deploy DBSCAN instead (same cluster base infra, real EngineFactory), matching what DistributedStandaloneCorrelationTest already does:

• DistributedRedundantCorrrelationTest / DistributedUDLTest: engine supplier "cluster" → "dbscan"
• sentinel-overlay-redundant boot: hardcoded alec-engine-cluster → the %ALEC_ENGINE_FEATURE% placeholder (so the supplier drives it, like the standalone overlay)
• opennms-integrated-overlay boot: alec-engine-cluster → alec-engine-dbscan, so IntegratedCorrelationTest/IntegratedUDLTest don't break smoke-test-full on develop post-merge

smoke-test compiles; the on-commit suite (DistributedRedundantCorrrelationTest) now resolves an EngineFactory and boots ALEC.

[joseanesONMS]

@cgorantla re-ping — this one pairs with #169 to finally get smoke-test-full green.

#169 fixes the distributed smoke OOM + OIA-version-skew (red since March). The last remaining full-suite failure is DistributedStandalone[deeplearning] (the TensorFlow engine timing out on model load) — which this PR removes, since ALEC-305 drops the deeplearning engine and reduces DistributedStandalone to {dbscan}. So #166 + #169 together = green full smoke.

On its own gate (smoke-test-commit = DistributedRedundant only), #166's DBSCAN fix already resolves it. Ready for your review.

[cgorantla]

LGTM!