Additional features

In the following, we introduce some features that are beyond the scope of above example workflow. For details and even more features, see Writing Workflows, Frequently Asked Questions and the command line help (snakemake --help).

Benchmarking

With the benchmark directive, Snakemake can be instructed to measure the wall clock time of a job. We activate benchmarking for the rule bwa_map:

rule bwa_map:
    input:
        "data/genome.fa",
        lambda wildcards: config["samples"][wildcards.sample]
    output:
        temp("mapped_reads/{sample}.bam")
    params:
        rg="@RG\tID:{sample}\tSM:{sample}"
    log:
        "logs/bwa_mem/{sample}.log"
    benchmark:
        "benchmarks/{sample}.bwa.benchmark.txt"
    threads: 8
    shell:
        "(bwa mem -R '{params.rg}' -t {threads} {input} | "
        "samtools view -Sb - > {output}) 2> {log}"

The benchmark directive takes a string that points to the file where benchmarking results shall be stored. Similar to output files, the path can contain wildcards (it must be the same wildcards as in the output files). When a job derived from the rule is executed, Snakemake will measure the wall clock time and memory usage (in MiB) and store it in the file in tab-delimited format. It is possible to repeat a benchmark multiple times in order to get a sense for the variability of the measurements. This can be done by annotating the benchmark file, e.g., with repeat("benchmarks/{sample}.bwa.benchmark.txt", 3) Snakemake can be told to run the job three times. The repeated measurements occur as subsequent lines in the tab-delimited benchmark file.

Modularization

In order to re-use building blocks or simply to structure large workflows, it is sometimes reasonable to split a workflow into modules. For this, Snakemake provides the include directive to include another Snakefile into the current one, e.g.:

include: "path/to/other.snakefile"

Alternatively, Snakemake allows to define sub-workflows. A sub-workflow refers to a working directory with a complete Snakemake workflow. Output files of that sub-workflow can be used in the current Snakefile. When executing, Snakemake ensures that the output files of the sub-workflow are up-to-date before executing the current workflow. This mechanism is particularly useful when you want to extend a previous analysis without modifying it. For details about sub-workflows, see the documentation.

Exercise

  • Put the read mapping related rules into a separate Snakefile and use the include directive to make them available in our example workflow again.

Automatic deployment of software dependencies

In order to get a fully reproducible data analysis, it is not sufficient to be able to execute each step and document all used parameters. The used software tools and libraries have to be documented as well. In this tutorial, you have already seen how Conda can be used to specify an isolated software environment for a whole workflow. With Snakemake, you can go one step further and specify Conda environments per rule. This way, you can even make use of conflicting software versions (e.g. combine Python 2 with Python 3).

In our example, instead of using an external environment we can specify environments per rule, e.g.:

rule samtools_index:
  input:
      "sorted_reads/{sample}.bam"
  output:
      "sorted_reads/{sample}.bam.bai"
  conda:
      "envs/samtools.yaml"
  shell:
      "samtools index {input}"

with envs/samtools.yaml defined as

channels:
  - bioconda
  - conda-forge
dependencies:
  - samtools =1.9

When Snakemake is executed with

snakemake --use-conda --cores 1

it will automatically create required environments and activate them before a job is executed. It is best practice to specify at least the major and minor version of any packages in the environment definition. Specifying environments per rule in this way has two advantages. First, the workflow definition also documents all used software versions. Second, a workflow can be re-executed (without admin rights) on a vanilla system, without installing any prerequisites apart from Snakemake and Miniconda.

Tool wrappers

In order to simplify the utilization of popular tools, Snakemake provides a repository of so-called wrappers (the Snakemake wrapper repository). A wrapper is a short script that wraps (typically) a command line application and makes it directly addressable from within Snakemake. For this, Snakemake provides the wrapper directive that can be used instead of shell, script, or run. For example, the rule bwa_map could alternatively look like this:

rule bwa_mem:
  input:
      ref="data/genome.fa",
      sample=lambda wildcards: config["samples"][wildcards.sample]
  output:
      temp("mapped_reads/{sample}.bam")
  log:
      "logs/bwa_mem/{sample}.log"
  params:
      "-R '@RG\tID:{sample}\tSM:{sample}'"
  threads: 8
  wrapper:
      "0.15.3/bio/bwa/mem"

The wrapper directive expects a (partial) URL that points to a wrapper in the repository. These can be looked up in the corresponding database. The first part of the URL is a Git version tag. Upon invocation, Snakemake will automatically download the requested version of the wrapper. Furthermore, in combination with --use-conda (see Automatic deployment of software dependencies), the required software will be automatically deployed before execution.

Cluster execution

By default, Snakemake executes jobs on the local machine it is invoked on. Alternatively, it can execute jobs in distributed environments, e.g., compute clusters or batch systems. If the nodes share a common file system, Snakemake supports three alternative execution modes.

In cluster environments, compute jobs are usually submitted as shell scripts via commands like qsub. Snakemake provides a generic mode to execute on such clusters. By invoking Snakemake with

$ snakemake --cluster qsub --jobs 100

each job will be compiled into a shell script that is submitted with the given command (here qsub). The --jobs flag limits the number of concurrently submitted jobs to 100. This basic mode assumes that the submission command returns immediately after submitting the job. Some clusters allow to run the submission command in synchronous mode, such that it waits until the job has been executed. In such cases, we can invoke e.g.

$ snakemake --cluster-sync "qsub -sync yes" --jobs 100

The specified submission command can also be decorated with additional parameters taken from the submitted job. For example, the number of used threads can be accessed in braces similarly to the formatting of shell commands, e.g.

$ snakemake --cluster "qsub -pe threaded {threads}" --jobs 100

Alternatively, Snakemake can use the Distributed Resource Management Application API (DRMAA). This API provides a common interface to control various resource management systems. The DRMAA support can be activated by invoking Snakemake as follows:

$ snakemake --drmaa --jobs 100

If available, DRMAA is preferable over the generic cluster modes because it provides better control and error handling. To support additional cluster specific parametrization, a Snakefile can be complemented by a Cluster Configuration (deprecated) file.

Using –cluster-status

Sometimes you need specific detection to determine if a cluster job completed successfully, failed or is still running. Error detection with --cluster can be improved for edge cases such as timeouts and jobs exceeding memory that are silently terminated by the queueing system. This can be achieved with the --cluster-status option. The value of this option should be a executable script which takes a job id as the first argument and prints to stdout only one of [running|success|failed]. Importantly, the job id snakemake passes on is captured from the stdout of the cluster submit tool. This string will often include more than the job id, but snakemake does not modify this string and will pass this string to the status script unchanged. In the situation where snakemake has received more than the job id these are 3 potential solutions to consider: parse the string received by the script and extract the job id within the script, wrap the submission tool to intercept its stdout and return just the job code, or ideally, the cluster may offer an option to only return the job id upon submission and you can instruct snakemake to use that option. For sge this would look like snakemake --cluster "qsub -terse".

The following (simplified) script detects the job status on a given SLURM cluster (>= 14.03.0rc1 is required for --parsable).

#!/usr/bin/env python
import subprocess
import sys

jobid = sys.argv[1]

output = str(subprocess.check_output("sacct -j %s --format State --noheader | head -1 | awk '{print $1}'" % jobid, shell=True).strip())

running_status=["PENDING", "CONFIGURING", "COMPLETING", "RUNNING", "SUSPENDED"]
if "COMPLETED" in output:
  print("success")
elif any(r in output for r in running_status):
  print("running")
else:
  print("failed")

To use this script call snakemake similar to below, where status.py is the script above.

$ snakemake all --jobs 100 --cluster "sbatch --cpus-per-task=1 --parsable" --cluster-status ./status.py

Using –cluster-cancel

When snakemake is terminated by pressing Ctrl-C, it will cancel all currently running node when using --drmaa. You can get the same behaviour with --cluster by adding --cluster-cancel and passing a command to use for canceling jobs by their jobid (e.g., scancel for SLURM or qdel for SGE). Most job schedulers can be passed multiple jobids and you can use --cluster-cancel-nargs to limit the number of arguments (default is 1000 which is reasonable for most schedulers).

Using –cluster-sidecar

In certain situations, it is necessary to not perform calls to cluster commands directly and instead have a “sidecar” process, e.g., providing a REST API. One example is when using SLURM where regular calls to scontrol show job JOBID or sacct -j JOBID puts a high load on the controller. Rather, it is better to use the squeue command with the -i/--iterate option.

When using --cluster, you can use --cluster-sidecar to pass in a command that starts a sidecar server. The command should print one line to stdout and then block and accept connections. The line will subsequently be available in the calls to --cluster, --cluster-status, and --cluster-cancel in the environment variable SNAKEMAKE_CLUSTER_SIDECAR_VARS. In the case of a REST server, you can use this to return the port that the server is listening on and credentials. When the Snakemake process terminates, the sidecar process will be terminated as well.

Constraining wildcards

Snakemake uses regular expressions to match output files to input files and determine dependencies between the jobs. Sometimes it is useful to constrain the values a wildcard can have. This can be achieved by adding a regular expression that describes the set of allowed wildcard values. For example, the wildcard sample in the output file "sorted_reads/{sample}.bam" can be constrained to only allow alphanumeric sample names as "sorted_reads/{sample,[A-Za-z0-9]+}.bam". Constraints may be defined per rule or globally using the wildcard_constraints keyword, as demonstrated in Wildcards. This mechanism helps to solve two kinds of ambiguity.

  • It can help to avoid ambiguous rules, i.e. two or more rules that can be applied to generate the same output file. Other ways of handling ambiguous rules are described in the Section Handling Ambiguous Rules.

  • It can help to guide the regular expression based matching so that wildcards are assigned to the right parts of a file name. Consider the output file {sample}.{group}.txt and assume that the target file is A.1.normal.txt. It is not clear whether dataset="A.1" and group="normal" or dataset="A" and group="1.normal" is the right assignment. Here, constraining the dataset wildcard by {sample,[A-Z]+}.{group} solves the problem.

When dealing with ambiguous rules, it is best practice to first try to solve the ambiguity by using a proper file structure, for example, by separating the output files of different steps in different directories.