F

Flyte enables you to build & deploy data & ML pipelines, hassle-free. The infinitely scalable and flexible workflow orchestration platform that seamlessly unifies data, ML and analytics stacks. Explore and Join the Flyte Community!

Checkpointing Issues in Flyte v1.11.0

Summary

The user is using Flyte v1.11.0 and has implemented Intratask checkpoints, which work with normal tasks but not with mapped tasks. They reference documentation suggesting the use of ArrayNode for support and have imported map_task from flytekit.experimental, but still encounter an error regarding checkpointing availability. The user is questioning if version 1.11.0 supports checkpointing in mapped tasks. Additionally, they mention that the flyte.binary (v1.11.0) pod is restarting due to a fatal error: concurrent map writes, indicating possible issues with the array node. The user is using Python 3.11 and flytekit version 1.11.0 and plans to create a minimal example to share.

Status
open
Tags
    Source
    #ask-the-community
      e

      eric901201

      10/6/2024

      And then I’ll ping maintianers

      e

      eric901201

      10/6/2024

      Please provide example

      e

      eric901201

      10/6/2024

      Ok

      s

      stephfen

      10/6/2024

      That is a fair suggestion, unfortunately I have another issue with later flyte versions and the main reason I have not tried it. Will wait for that to be fixed (i.e. being able to define and run flyte within a jupyter notebook) and after that will try it.

      What I mainly want to know if map_array is something that I should even try to use in v1.11.0 for intra task checkpoints or if I should just wait until I upgrade.

      e

      eric901201

      10/6/2024

      I’m not familiar with map task, but I can try my best to help you

      e

      eric901201

      10/6/2024

      Thank you

      e

      eric901201

      10/6/2024

      And also try the latest flytekit

      e

      eric901201

      10/6/2024

      I’ll help you ping someone who might know the answer after you try it

      e

      eric901201

      10/6/2024

      Flyte

      e

      eric901201

      10/6/2024

      Can you try the latest flute?

      s

      stephfen

      10/6/2024

      Actually, flyte.binary (v1.11.0) pod is restarting due to an error fatal error: concurrent map writes so maybe the array node has some issues. (/flyteorg/build/flytepropeller/pkg/controller/nodes/array/execution_context.go:29 ). The python version is 3.11 and flytekit version is 1.11.0. I will try to create a minimal example and share.

      e

      eric901201

      10/6/2024

      Can you provide your

      1. Python version
      2. Flytekit version
      3. Example code to reproduce this error?
      s

      stephfen

      10/6/2024

      I currently am with Flyte v1.11.0 and started to use Intratask checkpoints. This works well when using normal tasks but fails when used inside mapped tasks. According to the docs (https://docs.flyte.org/en/v1.11.0/user_guide/advanced_composition/map_tasks.html) I would need to use ArrayNode in order to get support. So I imported the map_task from flytekit.experimental, but unfortunately, the same error Checkpointing is not available, please check the version of the platform. . Does v1.11.0 support checkpointing in mapped tasks?