feat(BA-5419): support percentage values for rolling update max_surge and max_unavailable

[jopemachine]

Resolves BA-5419.

Summary

• Add support for float fraction values (e.g., 0.25 for 25%) for max_surge and max_unavailable rolling update parameters, in addition to existing absolute integer values
• Float values are resolved to absolute counts at execution time based on desired replica count, with max_surge rounding up and max_unavailable rounding down (matching Kubernetes semantics)
• Validation ensures float fractions are between 0.0 and 1.0, and existing integer-only configurations continue to work without changes

Rounding semantics (follows Kubernetes convention)

Field	Rounding	Reason
max_surge	ceil (rounds up)	Determines how many extra replicas can be created — rounding up ensures at least 1 can be created for progress
max_unavailable	floor (rounds down)	Determines how many replicas can be taken down — rounding down preserves availability conservatively

Example (desired_replicas = 10, both set to 25%):

• max_surge: ceil(10 * 0.25) = ceil(2.5) = 3 → up to 13 replicas simultaneously
• max_unavailable: floor(10 * 0.25) = floor(2.5) = 2 → at least 8 replicas always available

Both directions round toward safety: surge allows more creation headroom, unavailable restricts how many can go down.

Test plan

• DTO validation tests: float fraction input accepted, boundary values, negative rejected
• RollingUpdateSpec tests: fraction resolve with rounding, passthrough for integers, deadlock rejection for 0+0
• Strategy evaluation tests: correct route creation/termination with fractional specs
• GQL mutation tests pass unchanged
• CI passes lint/check/test

Resolves BA-5419

---

📚 Documentation preview 📚: https://sorna--10563.org.readthedocs.build/en/10563/

---

📚 Documentation preview 📚: https://sorna-ko--10563.org.readthedocs.build/ko/10563/

[Copilot]

Pull request overview

Adds support for percentage-based max_surge / max_unavailable for rolling updates (Kubernetes-style rounding), expanding validation/DTO/GQL surfaces and updating strategy evaluation to resolve percentages at runtime.

Changes:

• Introduce shared validate_int_or_percent() / resolve_int_or_percent() helpers and apply them to rolling update config/spec.
• Update rolling update strategy to resolve surge/unavailable against desired replicas at execution time.
• Expand DTO/GQL types and add unit tests covering percentage parsing and rounding behavior.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
tests/unit/manager/sokovan/deployment/strategy/test_rolling_update.py	Adds strategy/spec unit tests for percentage inputs, rounding, and deadlock prevention cases.
tests/unit/common/dto/manager/v2/deployment/test_request.py	Adds DTO validation tests ensuring percent strings are accepted/rejected as expected and numeric strings normalize to int.
src/ai/backend/manager/sokovan/deployment/strategy/rolling_update.py	Switches from raw spec fields to desired-based resolution helpers before computing budgets.
src/ai/backend/manager/models/deployment_policy/row.py	Updates RollingUpdateSpec schema to accept int-or-percent and adds resolver helpers.
src/ai/backend/manager/api/gql/deployment/types/policy.py	Broadens GQL fields to JSON to carry either int or percent string.
src/ai/backend/common/types.py	Adds shared int-or-percent validation and resolution helpers.
src/ai/backend/common/dto/manager/v2/deployment/types.py	Broadens response DTO fields to `int
src/ai/backend/common/dto/manager/v2/deployment/request.py	Broadens request DTO to accept int-or-percent with a shared validator.
changes/10563.feature.md	Adds changelog entry for the new percentage-based behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.