[ci] Stop billing tags from overriding advanced_instance_config fields

[sai-miduthuri]

Description

Refines _annotate_cluster_compute (new-schema branch) so the billing-tag TagSpecifications land at exactly the advanced_instance_config levels that the cluster will actually use at launch. Previously every new-schema compute config got tags added at three locations unconditionally: top-level, head_node, and every worker_nodes[*].

Final rules

Input state	Base tagged?	Per-group tagged?
No advanced_instance_config anywhere	yes (auto-created)	—
advanced_instance_config at base only	yes	—
advanced_instance_config only on head and/or workers	no (not auto-created)	yes (only the groups that already have one)
advanced_instance_config at base AND on head/workers	yes	yes (only the groups that already have one)

In code: should_tag_base = has_base_aic or not (has_head_aic or has_worker_aic); per-group specs are tagged iff the user already supplied them.

Why

Anyscale resolves the effective advanced instance spec per node group using an either-or pick: when a per-group spec is present, it replaces (does not merge with) the cluster-level base spec. The previous behavior unconditionally created TagSpecifications-only entries on head_node and worker_nodes[*], which made those per-group specs non-empty and caused them to win — silently dropping any IamInstanceProfile/NetworkInterfaces/etc. set only at the cluster level.

This was observed on the iceberg_benchmark_* tests after their compute config was migrated to the new schema: workers no longer picked up the ray-autoscaler-v1 IAM instance profile that was declared at the top level of iceberg_benchmark_compute.yaml, so the test couldn't access the AWS resources it needed.

Test

release/ray_release/tests/test_cluster_manager.py:

• testClusterComputeExtraTagsNewSchema (updated) covers the no-AIC and base+per-group rows, including verifying that pre-existing IamInstanceProfile values on each level are preserved alongside the added TagSpecifications.
• testClusterComputeNewSchemaNoAic, testClusterComputeNewSchemaPerGroupHeadOnly, testClusterComputeNewSchemaPerGroupWorkerOnly (new) — one method per row of the table for clear regression coverage.
• testClusterComputeNewSchemaNonAws (new) — verifies the non-AWS early return.

Related issues

Related to #62863 (where this bug was first observed during the dataset compute-config migration).

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request modifies the cluster manager to only annotate the top-level advanced_instance_config with billing tags, preventing Anyscale from overriding cluster-level configurations when per-group specifications are present. Corresponding tests were updated to verify that node groups are no longer automatically annotated. A review comment suggests that the code should still annotate advanced_instance_config if it is explicitly provided by the user for specific node groups to ensure billing tags are not lost in those instances.

[elliot-barn]

should we throw an error here if the worker isn't a dict? not sure if we have cluster validation before this

[sai-miduthuri]

We do have cluster validation as a separate step when the compute-configs are loaded into memory for each each test through the test-collection:

ray/release/ray_release/scripts/run_release_test.py

Lines 92 to 95 in 2d9c453

	test_collection = read_and_validate_release_test_collection(
	test_collection_file or RELEASE_TEST_CONFIG_FILES,
	test_definition_root,
	)

ray/release/ray_release/config.py

Lines 108 to 112 in 2d9c453

	validate_release_test_collection(
	tests,
	schema_file=schema_file,
	test_definition_root=test_definition_root,
	)

ray/release/ray_release/config.py

Line 269 in 2d9c453

error = validate_test_cluster_compute(test, test_definition_root)

ray/release/ray_release/config.py

Lines 328 to 330 in 2d9c453

	is_new_schema = test.uses_anyscale_sdk_2026()
	cluster_compute = load_test_cluster_compute(test, test_definition_root)
	return validate_cluster_compute(cluster_compute, is_new_schema=is_new_schema)

ray/release/ray_release/config.py

Line 333 in 2d9c453

def validate_cluster_compute(

We implicitly ensure that the cluster_compute["head_node"] and cluster_compute["worker_nodes"] are of dict type if they are present in the compute-config.

Though I can add additional checks and throw an error here if it isn't a dict, I think we can remove the isinstance(..., dict) check instead and just assume that it is a dict.

What is your recommendation?

[elliot-barn]

Will leave as is. Approved!