paraffin 0.2.0

Parallel execution of DVC stages

paraffin

Paraffin, derived from the Latin phrase parum affinis meaning little related, is a Python package designed to run DVC stages in parallel. While DVC does not currently support this directly, Paraffin provides an effective workaround. For more details, refer to the DVC documentation on parallel stage execution.

[!WARNING] paraffin is still very experimental. Do not use it for production workflows.

Installation

Install Paraffin via pip:

pip install paraffin

Usage

The paraffin submit command mirrors dvc repro, enabling you to queue and execute your entire pipeline or selected stages with parallelization. If no parameters are specified, the entire graph will be queued and executed via dvc repro --single-item.

paraffin submit <stage name> <stage name> ... <stage name>
# Example: run with a maximum of 4 parallel jobs
paraffin worker --concurrency=4

Parallel Execution

Due to limitations in Celery’s graph handling (see Celery discussion), complete parallelization is not always achievable. Paraffin will display parallel-ready stages in a flowchart format. All stages are visualized in a Mermaid flowchart.

flowchart TD
        subgraph Level0:1
                A_X_ParamsToOuts
                A_X_ParamsToOuts_1
                A_Y_ParamsToOuts
                A_Y_ParamsToOuts_1
        end
        subgraph Level0:2
                A_X_AddNodeNumbers
                A_Y_AddNodeNumbers
        end
        subgraph Level0:3
                A_SumNodeAttributes
        end
        Level0:1 --> Level0:2
        Level0:2 --> Level0:3
        subgraph Level1:1
                B_X_ParamsToOuts
                B_X_ParamsToOuts_1
                B_Y_ParamsToOuts
                B_Y_ParamsToOuts_1
        end
        subgraph Level1:2
                B_X_AddNodeNumbers
                B_Y_AddNodeNumbers
        end
        subgraph Level1:3
                B_SumNodeAttributes
        end
        Level1:1 --> Level1:2
        Level1:2 --> Level1:3

Queue Labels

To fine-tune execution, you can assign stages to specific Celery queues, allowing you to manage execution across different environments or hardware setups. Define queues in a paraffin.yaml file:

queue:
    "B_X*": BQueue
    "A_X_AddNodeNumbers": AQueue

Then, start a worker with specified queues, such as celery (default) and AQueue:

paraffin worker -q AQueue,celery

All stages not assigned to a queue in paraffin.yaml will default to the celery queue.

[!TIP] If you are building Python-based workflows with DVC, consider trying our other project ZnTrack for a more Pythonic way to define workflows.