Carry pod_template override through map_task serialization

[1fanwang]

Tracking issue

Fixes flyteorg/flyte#7076

Why are the changes needed?

map_task(...).with_overrides(pod_template=PodTemplate(...)) silently drops the override at serialization. The override is accepted at workflow build time and stored on the Node (node._pod_template), but get_serializable_array_node_map_task in flytekit/tools/translator.py never reads that field — it builds TaskNodeOverrides with only resources, extended_resources, and container_image. The user thinks the override is applied, but the registered task spec has an empty overrides {} block.

This is the same class of bug as #6463, which #3270 fixed for the regular-task path of get_serializable_node. The array-node path has the same gap in a different function.

What changes were proposed in this pull request?

flytekit/tools/translator.py — get_serializable_array_node_map_task:

• Read node._pod_template, build the override pod spec via _serialize_pod_spec, and include pod_template in the emitted TaskNodeOverrides.
• Use entity.get_container(settings) (not _get_container) so prepare_target() substitutes the map-task command (pyflyte-map-execute) into the container args before the pod spec is built. ArrayNodeMapTask extends PythonTask directly (not PythonAutoContainerTask), so _get_container isn't on it anyway.
• For fast-serialization, prefix container args post-build (matching the existing pattern in get_serializable_task at translator.py:170) instead of swapping command_fn via _fast_serialize_command_fn. The latter wraps task.get_default_command(settings), which on ArrayNodeMapTask returns the inherited pyflyte-execute — wrong shape for a map task.

tests/flytekit/unit/test_translator.py:

• New parametrized regression test test_map_task_with_pod_template_override covering both fast_registration_enabled=True and False. Mirrors the structure of the existing test_task_with_pod_template_override from Fix issue where pod template override pod spec was missing #3270.

How was this patch tested?

• New parametrized test passes (both cases).
• Full tests/flytekit/unit/test_translator.py (11 tests) passes.
• Adjacent suites that exercise the same surfaces all pass with no regressions:
  • tests/flytekit/unit/core/test_array_node_map_task.py (46 tests)
  • tests/flytekit/unit/core/test_map_task.py (23 tests)
  • tests/flytekit/unit/core/test_node_creation.py (26 tests)

Check all the applicable boxes

• All new and existing tests passed.
• All commits are signed-off.
• I updated the documentation accordingly. (No user-facing behavior change beyond the bug fix; the documented behavior of with_overrides(pod_template=...) for map_task is what this PR makes work.)

[pingsutw]

Thank you for fixing it. LGTM

[pingsutw]

Could we also add another test that t has a podTemplate config and we use with_overrides() to override it?

 @task(pod_template=base_pod_template)
    def t(a: int) -> str:
        return str(a)

[1fanwang]

Done in 9dde146 — added test_map_task_with_pod_template_override_replaces_task_pod_template covering the case where @task(pod_template=base_pod_template) is then overridden via map_task(t)(...).with_overrides(pod_template=override_pod_template).

While writing the test I caught a real bug: when the underlying task already has pod_template set, entity.get_container(settings) returns None (it intentionally defers to get_k8s_pod), and _serialize_pod_spec would dereference primary_container.image on None. Fixed in the same commit by falling back to the inner task's _get_container under prepare_target() so the override pod spec still resolves image/command/args. Test asserts the override's OVERRIDE_KEY env wins over the base.

[pingsutw]

Looks good, thanks

[welcome]

Congrats on merging your first pull request! 🎉