Description

SandboxInClusterConnectionConfig.use_pod_ip (default False) is documented as: "By default, connects via stable K8s DNS … When use_pod_ip=True, connects directly to the pod IP." (models.py).

As of 636e032 (#800), both Sandbox and AsyncSandbox unconditionally pass get_pod_ip to SandboxConnector:

# sandbox.py:65 / async_sandbox.py:68
get_pod_ip=self.get_pod_ip,   # was: self.get_pod_ip if use_pod_ip else None

InClusterConnectionStrategy.connect() then always attempts pod-IP routing first and silently falls back to cluster DNS. Net effect: the use_pod_ip flag is dead — setting it to False (the documented default) does not force DNS routing.

Impact

• Doc/behavior mismatch: the default is now pod-IP-first, not "stable K8s DNS".
• Environments that require DNS/Service routing (strict NetworkPolicy, service mesh, secure overlay) can't opt out — the same scenario [Go SDK] Support configuration option to opt-out or toggle direct Pod IP routing #916 raises for the Go SDK. [Go SDK] Support configuration option to opt-out or toggle direct Pod IP routing #916's premise that the Python SDK can toggle this via use_pod_ip no longer holds at HEAD.
• Stale tests: test/unit/test_async_sandboxclient.py asserts use_pod_ip is "derived from connection_config", a derivation the wiring no longer performs.

Suggested fix

Restore the gating — get_pod_ip=self.get_pod_ip if use_pod_ip else None in both Sandbox and AsyncSandbox — or make connect() honor config.use_pod_ip directly.

Note

Introduced in #800; present at HEAD. Released 0.4.5 still has the correct gating, so consumers pinned to 0.4.5 are unaffected.