Simple comparison of Snakemake, Nextflow and Cromwell/WDL

A very simple workflow (download FASTQ files and align them to a reference genome) implemented in Snakemake, Nextflow and WDL. This README contains a subjective comparison of the pipelines.

• reddit thread about the pipelines.
• biorxiv preprint sharing experiences from rapid prototyping a pipeline with different workflow managers.

Criteria

• Support for conda envs
• Support for singularity
• User-friendlyness
• How easy to submit on cluster?

Snakemake

✓ conda ✓ singularity

• very good documentation

• published 2012, >500 citations.

How to run on cluster:

snakemake --use-conda --cluster qsub --jobscript /path/to/wrapper_script.sh --jobs 32

• --use-conda is required to create an environment for each task from a corresponding yaml file
• --cluster qsub tells snakemake to submit jobs using qsub
• --jobscript points to a wrapper script for qsub containing configuration options
• --jobs max number of jobs submitted to the cluster in parallel. Snakemake takes care of resubmitting.
• --profile allows to configure system-wide profiles to run on the cluster.

The jobscript looks something like this:

#$ -cwd
#$ -V
#$ -e ./logs/
#$ -o ./logs/

Nextflow

✓ conda ✓ singularity

• published 2017, 180 citations
• works with conda in principle, but currently not because of #1195.
• very good documentation
• Nf-core to publish pipelines.

Differences to Snakemake

• has no 'dry-run' mode
• change executor only via config files (bad for testing)
• file-name-abstraction. That means: not all files are exposed to the user but stay in a hidden working directory (every job gets an own directory). Debugging is easy, as one can change to the job-directory and check what's going on. Result files can be explicitly exposed via publishDir directive.

How to run on cluster

• via configuration file. Either global or local.
• local configuration file: put into project directory nextflow.config

Looks like this:

To run on cluster:

process {
  executor = 'sge'
  penv = 'smp'
    clusterOptions = { "-V -S /bin/bash " }
  }

To run locally:

executor {
  cpus = 10 /* max cpus on local machine */
}

Cromwell/WDL

✗ conda ✓ singularity

• mediocre documentation for getting the user started, but it's ok.
• finishes all running tasks, even if a later task fails.
• no support for conda environments

How to run on cluster

• via configuration file.

backend {
  default = "SGE"
  providers {
    SGE {
      ...
      submit = "qsub -V ..."
    }
  }
}