To pass a struct (such as a Python dataclass or Pydantic BaseModel) as an input to a Flyte workflow, define the struct using Python's dataclass or Pydantic BaseModel, then use it as a type annotation for the workflow input. Flytekit will handle serialization and deserialization automatically. You can provide the input as a JSON object when launching the workflow via the Flyte Console or CLI. For dataclasses, Flytekit uses MessagePack or JSON for serialization depending on the version, and the input form in the Console is generated from the struct's JSON schema. See the example below:

from dataclasses import dataclass
from flytekit import task, workflow

@dataclass
class MyStruct:
    a: int
    b: str

@task
def my_task(x: MyStruct) -> int:
    return x.a

@workflow
def my_wf(x: MyStruct) -> int:
    return my_task(x=x)

You can launch this workflow and provide the input as a JSON object, e.g.,

{"a": 1, "b": "hello"}

. For more, see the Flyte dataclass documentation and related issue. Would you like more detail or examples for Pydantic models or advanced struct usage? Sources: - Flyte dataclass documentation - Flyte issue: Unable to input dataclass and optional dataclass types in Flyte Console 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.

When using FlyteRemote to execute a workflow with a dataclass (struct) input, you must provide the input as a plain Python dictionary matching the dataclass structure, not as a dataclass instance. FlyteRemote expects the input to be serializable to JSON/struct, not a pickled binary object. For example, if your dataclass is:

from dataclasses import dataclass

@dataclass
class SinkTable:
    name: str
    id: int

You should call FlyteRemote like this:

inputs = {"sink_table": {"name": "my_table", "id": 123}}
execution = remote.execute(workflow, inputs=inputs)

Passing a dataclass instance directly will result in a type mismatch (

simple:BINARY

instead of

simple:STRUCT

). For more, see the Flyte dataclass docs and related issue. Would you like more detail or troubleshooting steps? Sources: - Flyte dataclass documentation - Flyte issue: Unable to input dataclass and optional dataclass types in Flyte Console 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.

Can you provide the versions of Flytekit (on your client) and the Flyte backend you are using, and confirm whether you have set the environment variable FLYTE_USE_OLD_DC_FORMAT to true when running FlyteRemote? 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.

Set the FLYTE_USE_OLD_DC_FORMAT environment variable in the Python process before you create or use FlyteRemote. You can do this at the top of your script:

import os
os.environ["FLYTE_USE_OLD_DC_FORMAT"] = "true"

from flytekit.remote import FlyteRemote
# ... rest of your code

This ensures Flytekit serializes dataclasses using the legacy Protobuf struct format, which may be required for compatibility with older Flyte backends. See Flyte dataclass documentation and Flyte PR 5991. Would you like more detail on version compatibility or troubleshooting? Sources: - Flyte dataclass documentation - Flyte PR 5991 👍 1 button 👎 0 button 💡 Hint: Mention @ancient-wolf-19325 in the thread for followups.

To pass a dataclass to subworkflows in Flyte, simply use the dataclass as an input/output type in both the parent and subworkflow signatures. Flytekit will handle serialization and deserialization automatically, allowing you to pass the dataclass instance directly between workflows and tasks. See the example below:

from dataclasses import dataclass
from flytekit import task, workflow

@dataclass
class MyData:
    foo: int
    bar: str

@task
def t1(x: MyData) -> MyData:
    return x

@workflow
def subwf(x: MyData) -> MyData:
    return t1(x=x)

@workflow
def main_wf(x: MyData) -> MyData:
    return subwf(x=x)

This pattern is supported and tested in Flyte, including for nested dataclasses. For more, see this integration test example. Would you like more detail or examples with nested dataclasses? Sources: - Flyte dataclass_with_optional_wf.py example 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.