CustomContainerTrainingJob.run drops max_wait_duration=0 instead of requesting indefinite DWS wait

从 google/cloud/aiplatform/training_jobs.py 中的 _prepare_training_task_inputs_and_output_dir 开始，然后对比 issue 中引用的较低层 CustomJob 实现。验证省略、零值和正值 max_wait_duration 的调度输入；完成标准是零被序列化为 0s，同时省略状态仍保持区别，正时长得到保留。

Thanks for maintaining this library. CustomContainerTrainingJob.run(max_wait_duration=0) drops the value because the high-level training-job implementation uses a truthiness check. This prevents callers from requesting the documented indefinite wait for DWS Flex Start.

Environment details

• OS type and version: macOS
• Python version: 3.12
• google-cloud-aiplatform versions: reproduced in 1.148.1; source inspection confirms the same behavior in 1.150.0, 1.163.0, and current main

Steps to reproduce

1. Create a CustomContainerTrainingJob using any valid project, staging bucket, and container image.
2. Run it with Flex Start and max_wait_duration=0.
3. Inspect the scheduling inputs produced by _prepare_training_task_inputs_and_output_dir or the resulting TrainingPipeline request.

Code example

job.run(
    scheduling_strategy=custom_job.Scheduling.Strategy.FLEX_START,
    max_wait_duration=0,
)

Observed scheduling value:

max_wait_duration: None

Expected scheduling value:

max_wait_duration: 0s

For comparison, a positive input such as 14400 is preserved as 14400s.

Cause

The high-level training-job path converts the duration using a truthiness check, so zero follows the same branch as None:

https://github.com/googleapis/python-aiplatform/blob/v1.163.0/google/cloud/aiplatform/training_jobs.py#L1663-L1667

The generated API contract says that explicit zero means indefinite waiting, while omission defaults to 24 hours:

https://github.com/googleapis/python-aiplatform/blob/v1.163.0/google/cloud/aiplatform_v1/types/custom_job.py#L577-L582

max_wait_duration is a presence-aware google.protobuf.Duration, so explicit zero and an omitted field are distinct requests.

A previous fix correctly changed the lower-level CustomJob and hyperparameter-tuning paths to use is not None, explicitly noting that 0 is valid:

https://github.com/googleapis/python-aiplatform/commit/d9675fdf051233539f478187143f2833fd6e6af0

That commit did not update the CustomContainerTrainingJob path in training_jobs.py.

Suggested fix

Use an explicit is not None check in _prepare_training_task_inputs_and_output_dir, consistent with the lower-level implementation, and add tests covering omitted, zero, and positive durations.

Stack trace

No exception is raised. The problem is a silent request-semantics change: 0 is serialized as omission.