[litellm-agent] Staging → litellm_internal_staging (5/15/2026)

[oss-pr-review-agent-shin]

Merged PRs (2)

#	Title
#27990	feat(chart): add support for autoscaling behavior in HPA
#27927	feat(proxy): add blocked flag to models for pause/resume from the UI

---

Auto-updated by litellm-agent on each merge.

[oss-pr-review-agent-shin]

@greptile please review

[greptile-apps]

Greptile Summary

This PR merges two features: HPA autoscaling behavior support in the Helm chart, and an admin-controlled blocked flag for proxy models to enable pause/resume from the UI.

• HPA behavior: Adds an optional behavior block to hpa.yaml via toYaml | nindent, guarded by {{- if .Values.autoscaling.behavior }}, with a new helm-unittest test suite covering both presence and absence of the field.
• Model blocking: Adds a blocked column to LiteLLM_ProxyModelTable (safe migration, DEFAULT false), propagates the flag through router filtering (_filter_blocked_deployments, _is_deployment_blocked, get_fully_blocked_model_names), enforces proxy-admin-only authorization in patch_model, hides fully-blocked models from /v1/models, and invalidates the in-memory router via clear_cache() → add_deployment() after a PATCH.
• Redis: Adds socket_timeout and socket_connect_timeout to the allowed cluster kwargs set.

Confidence Score: 5/5

Safe to merge. Both features are additive, the migration is backward-compatible, and the model-blocking logic is applied consistently across all routing and credential-lookup paths.

The blocked flag flows correctly from DB → get_model_info_with_id → router model list → every routing entry point (sync, async, pass-through, credential helpers). The auth gate is correctly placed before the DB write. clear_cache() triggers a synchronous DB reload so the block takes effect immediately after the PATCH. The migration uses IF NOT EXISTS with DEFAULT false so existing rows are unaffected. Tests are mock-only and cover all new paths.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/router.py	Adds _filter_blocked_deployments, _is_deployment_blocked, and get_fully_blocked_model_names; applies blocked filtering at every routing entry point (sync/async, pass-through, direct dict, credential helpers). Logic is consistent and well-tested.
litellm/proxy/management_endpoints/model_management_endpoints.py	Auth gate (proxy-admin only) added before the DB update; update_db_model threads blocked into the Prisma payload; clear_cache() triggers a full DB reload so the block takes effect synchronously.
litellm/proxy/proxy_server.py	Propagates DB blocked column into model_info["blocked"] during model load; filters fully-blocked model names from /v1/models in both code branches of model_list.
litellm-proxy-extras/litellm_proxy_extras/migrations/20260514120000_add_blocked_to_proxy_model_table/migration.sql	Safe additive migration: ADD COLUMN IF NOT EXISTS "blocked" BOOLEAN NOT NULL DEFAULT false — existing rows are unaffected.
litellm/types/router.py	Adds blocked: Optional[bool] = None to ModelInfo and updateDeployment. ModelInfo uses extra="allow" and a custom .get() so dict-style access in the hot path remains compatible.
litellm/proxy/route_llm_request.py	Adds _is_deployment_blocked check to the credential-resolution path for file/batch/passthrough endpoints, preventing bypass of an admin pause.
deploy/charts/litellm-helm/templates/hpa.yaml	Adds optional behavior block to HPA spec using standard `toYaml
tests/test_litellm/test_router.py	Comprehensive mock-only tests for all new blocking paths: sync/async healthy deployment filters, pass-through, credential helpers, and get_fully_blocked_model_names.

_{Reviews (3): Last reviewed commit: "fix: pass socket timeouts to Redis clust..." | Re-trigger Greptile}

[greptile-apps]

The commented-out hint behavior: {} documents the option but an empty map is falsy in Go templates, so uncomment-and-leave-empty silently produces no output. A brief inline example with at least one meaningful key makes the expectation clearer and reduces user confusion.

Suggested change

	# behavior: {}
	# behavior:
	# scaleDown:
	# stabilizationWindowSeconds: 300
	# scaleUp:
	# stabilizationWindowSeconds: 0

[codecov]

Codecov Report

❌ Patch coverage is 75.00000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/proxy/route_llm_request.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[oss-pr-review-agent-shin]

@greptile please review

[oss-pr-review-agent-shin]

@greptile please review