Skip to content

Orchestrator can stall due to hashing output artifacts when the artifact has a lot of files #338

Description

@Ark-kun

Long-term solution: Move artifact hash calculation from Orchestrator-side to container-side.

Short-term solution: Introduce proper timeouts on artifact hash calculation.

A GCS directory with ~100000 files takes almost 40 seconds to list (100 HTTP requests), but sometimes the real number can jump to 2000+ seconds.

GCS client has timeout and retry/deadline parameters, but they do not work properly for list_blobs method due to paging. In fact, reducing timeout may increases the total time due to retries.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions