Skip to content
This repository was archived by the owner on Jan 12, 2026. It is now read-only.

implement basic watermarking#24

Open
iasoon wants to merge 1 commit into
ray-project:masterfrom
iasoon:watermarks
Open

implement basic watermarking#24
iasoon wants to merge 1 commit into
ray-project:masterfrom
iasoon:watermarks

Conversation

@iasoon

@iasoon iasoon commented Jun 15, 2022

Copy link
Copy Markdown
Member

Re-enabled the watermarking code so that the watermark assertions pass again

@iasoon

iasoon commented Jun 15, 2022

Copy link
Copy Markdown
Member Author

R: @pabloem

'wrong input watermark for %s. Expected %s, but got %s.' % (
ray.get(runner_execution_context.watermark_manager.get_stage_node.remote(
bundle_context_manager.stage.name)),
stage_node.output_watermark() == timestamp.MAX_TIMESTAMP), (

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This was indeed supposed to be output_watermark instead of input_watermark, right?

@pabloem

pabloem commented Jun 15, 2022

Copy link
Copy Markdown
Collaborator

TODO @pabloem - what is the produced watermark and why do stages differentiate

@iasoon iasoon force-pushed the watermarks branch 3 times, most recently from 4f0f971 to 015df34 Compare June 16, 2022 22:17
@iasoon

iasoon commented Jun 16, 2022

Copy link
Copy Markdown
Member Author

Ran into an issue when implementing this:

transform.spec.payload = data_spec.SerializeToString()

Overwriting the payload here also changes the transform in the original pipeline definition. This new format is different from what the watermark manager __init__ expects. I fixed it for now by making sure that the watermark manager initializes before the RayBundleContextManager runs, How is this expected to work? Should we copy the transform before we overwrite the payload?

@pabloem

pabloem commented Jun 21, 2022

Copy link
Copy Markdown
Collaborator

once we start parallelizing the execution of bundles, our watermarking will need to change (we need to wait for all parallel tasks of a given stage before advancing its watermark and executing the downstream stages).

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants