This is hopefully the last beta release. We'll let this sit for a week and then cut a formal release. Any bug fixes done next week will likely just go into that release unless it's urgent.
With this release...
Map Task
We've added a map task. Map tasks are similar to the `map` functionality you find in Python and apply a list over an existing task.
workflow
def my_wf(a: typing.List[int]) -> int:
x = map_task(t1, metadata=TaskMetadata(retries=1))(a=a)
return t2(x=x)
Please see the [PR](https://github.com/flyteorg/flytekit/pull/405/files) and code comments for more information. Further documentation is forthcoming.
Note that map tasks are different (more optimized) than dynamic tasks because if you map a task over a thousand elements it still remains just _one_ task. A `dynamic` task would create a thousand nodes.
Task Resolver
Shoring up the execution side of flytekit (when the Flyte backend actually runs the task container on K8s) has been a long time coming. Again please see notes in the [PR](https://github.com/flyteorg/flytekit/pull/404) and code comments for more information, but this updates the way tasks are loaded at execution time and adds flexibility.
Previously, tasks were always loaded by one call to Python's `importlib.import_module()`. Now, at serialization time, the 'task resolver', a [default](https://github.com/flyteorg/flytekit/blob/master/flytekit/core/python_auto_container.py#L228) for which is provided, constructs arguments that, at execution time, the same resolver uses to resurrect/rehydrate the task. The default resolver merely does the same thing as previous, calling `import_module` and looking up the task name.
Misc
* Fixed an issue where subworkflows called in conditionals weren't getting included to the workflow closure.
* Fixed an issue where subworkflows were running even in false conditional branches during local execution
* Added custom Protobuf Type support
* Fixed the ability to construct launch plan from a `ReferenceWorkflow`