Skip to content

13. Execution cache and resume

The Nextflow caching mechanism works by assigning a unique ID to each task which is used to create a separate execution directory where the tasks are executed and the results stored.

The task unique ID is generated as a 128-bit hash value composing the task input values, files and command string.

The workflow work directory is organized as shown below:

work/
├── 12
│   └── 1adacb582d2198cd32db0e6f808bce
│       ├── genome.fa -> /data/../genome.fa
│       └── index
│           ├── hash.bin
│           ├── header.json
│           ├── indexing.log
│           ├── quasi_index.log
│           ├── refInfo.json
│           ├── rsd.bin
│           ├── sa.bin
│           ├── txpInfo.bin
│           └── versionInfo.json
├── 19
│   └── 663679d1d87bfeafacf30c1deaf81b
│       ├── ggal_gut
│       │   ├── aux_info
│       │   │   ├── ambig_info.tsv
│       │   │   ├── expected_bias.gz
│       │   │   ├── fld.gz
│       │   │   ├── meta_info.json
│       │   │   ├── observed_bias.gz
│       │   │   └── observed_bias_3p.gz
│       │   ├── cmd_info.json
│       │   ├── libParams
│       │   │   └── flenDist.txt
│       │   ├── lib_format_counts.json
│       │   ├── logs
│       │   │   └── salmon_quant.log
│       │   └── quant.sf
│       ├── ggal_gut_1.fq -> /data/../ggal_gut_1.fq
│       ├── ggal_gut_2.fq -> /data/../ggal_gut_2.fq
│       └── index -> /data/../asciidocs/day2/work/12/1adacb582d2198cd32db0e6f808bce/index

Info

You can create these plots using the tree function if you have it installed.

13.0.1 How resume works

The -resume command-line option allows the continuation of a workflow execution from the last step that was completed successfully:

nextflow run <script> -resume

In practical terms, the workflow is executed from the beginning. However, before launching the execution of a process, Nextflow uses the task unique ID to check if the work directory already exists and that it contains a valid command exit state with the expected output files.

If this condition is satisfied the task execution is skipped and previously computed results are used as the process results.

The first task for which a new output is computed invalidates all downstream executions in the remaining DAG.

13.0.2 Work directory

The task work directories are created in the folder work in the launching path by default. It is recommended that this is a scratch storage area that can be cleaned up once the computation is completed.

Note

It is recommended that you store the final output(s) in a different location using one or more publishDir directives.

Warning

Make sure to delete your work directory occasionally, else your machine/environment may be filled with unused files.

A different location for the execution work directory can be specified using the command line option -w:

nextflow run <script> -w /some/scratch/dir

Warning

If you delete or move the workflow work directory, it will prevent the use of the resume feature in subsequent runs.

The hash code for input files is computed using:

  • The complete file path
  • The file size
  • The last modified timestamp

Therefore, just touching a file will invalidate the related task execution.

13.0.3 How to organize in-silico experiments

It’s good practice to organize each experiment in its own folder. The main experiment input parameters should be specified using a Nextflow config file. This makes it simple to track and replicate an experiment over time.

Note

In the same experiment, the same workflow can be executed multiple times, however, launching two (or more) Nextflow instances in the same directory concurrently should be avoided.

The nextflow log command lists the executions run in the current folder:

nextflow log
Output
TIMESTAMP            DURATION  RUN NAME          STATUS  REVISION ID  SESSION ID                            COMMAND
2019-05-06 12:07:32  1.2s      focused_carson    ERR     a9012339ce   7363b3f0-09ac-495b-a947-28cf430d0b85  nextflow run hello
2019-05-06 12:08:33  21.1s     mighty_boyd       OK      a9012339ce   7363b3f0-09ac-495b-a947-28cf430d0b85  nextflow run rnaseq-nf -with-docker
2019-05-06 12:31:15  1.2s      insane_celsius    ERR     b9aefc67b4   4dc656d2-c410-44c8-bc32-7dd0ea87bebf  nextflow run rnaseq-nf
2019-05-06 12:31:24  17s       stupefied_euclid  OK      b9aefc67b4   4dc656d2-c410-44c8-bc32-7dd0ea87bebf  nextflow run rnaseq-nf -resume -with-docker

You can use either the session ID or the run name to recover a specific execution:

nextflow run rnaseq-nf -resume mighty_boyd

13.0.4 Execution provenance

The log command, when provided with a run name or session ID, can return many useful bits of information about a workflow execution that can be used to create a provenance report.

By default, it will list the work directories used to compute each task:

nextflow log tiny_fermat
Output
/data/.../work/7b/3753ff13b1fa5348d2d9b6f512153a
/data/.../work/c1/56a36d8f498c99ac6cba31e85b3e0c
/data/.../work/f7/659c65ef60582d9713252bcfbcc310
/data/.../work/82/ba67e3175bd9e6479d4310e5a92f99
/data/.../work/e5/2816b9d4e7b402bfdd6597c2c2403d
/data/.../work/3b/3485d00b0115f89e4c202eacf82eba

The -f (fields) option can be used to specify which metadata should be printed by the log command:

nextflow log tiny_fermat -f 'process,exit,hash,duration'
Output
index    0   7b/3753ff  2.0s
fastqc   0   c1/56a36d  9.3s
fastqc   0   f7/659c65  9.1s
quant    0   82/ba67e3  2.7s
quant    0   e5/2816b9  3.2s
multiqc  0   3b/3485d0  6.3s

The complete list of available fields can be retrieved with the command:

nextflow log -l

The -F option allows the specification of filtering criteria to print only a subset of tasks:

nextflow log tiny_fermat -F 'process =~ /fastqc/'
Output
/data/.../work/c1/56a36d8f498c99ac6cba31e85b3e0c
/data/.../work/f7/659c65ef60582d9713252bcfbcc310

This can be useful to locate specific task work directories.

Finally, the -t option enables the creation of a basic custom provenance report, showing a template file in any format of your choice:

<div>
    <h2>${name}</h2>
    <div>
        Script:
        <pre>${script}</pre>
    </div>

    <ul>
        <li>Exit: ${exit}</li>
        <li>Status: ${status}</li>
        <li>Work dir: ${workdir}</li>
        <li>Container: ${container}</li>
    </ul>
</div>

Exercise

Save the above snippet in a file named template.html. Then run this command (using the correct id for your run, e.g., tiny_fermat):

nextflow log tiny_fermat -t template.html > prov.html

Finally, open the prov.html file with a browser.

Summary

In this step you have learned:

  1. How the workflow execution cache works
  2. How to use the -resume command line option
  3. How to organize in-silico experiments in different folders
  4. How to create and customize a basic provenance report

13.1 Troubleshooting resume

Being able to resume workflows is a key feature of Nextflow, but it doesn't always work as you expect. In this section you will learn common reasons why Nextflow may be ignoring your cached results.

Tip

To learn more details about the resume mechanism and how to troubleshoot please refer to the following three blog posts:

  1. Demystifying Nextflow resume
  2. Troubleshooting Nextflow resume
  3. Analyzing caching behavior of pipelines

13.1.1 Input file changed

Make sure that there’s no change in your input file(s). Don’t forget the task unique hash is computed by taking into account the complete file path, the last modified timestamp and the file size. If any of this information has changed, the workflow will be re-executed even if the input content is the same.

13.1.2 A process modifies an input

A process should never alter input files, otherwise the resume for future executions will be invalidated for the same reason explained in the previous point.

13.1.3 Inconsistent file attributes

Some shared file systems, such as NFS, may report an inconsistent file timestamp (i.e. a different timestamp for the same file) even if it has not been modified. To prevent this problem use the lenient cache strategy.

13.1.4 Race condition in global variable

Nextflow is designed to simplify parallel programming without taking care about race conditions and the access to shared resources. One of the few cases in which a race condition can arise is when using a global variable with two (or more) operators

snippet.nf
1
2
3
4
5
6
7
8
9
Channel
    .of(1, 2, 3)
    .map { it -> X = it; X += 2 }
    .view { "ch1 = $it" }

Channel
    .of(1, 2, 3)
    .map { it -> X = it; X *= 2 }
    .view { "ch2 = $it" }

The problem in this snippet is that the X variable in the closure definition is defined in the global scope. Therefore, since operators are executed in parallel, the X value can be overwritten by the other map invocation.

The correct implementation requires the use of the def keyword to declare the variable local.

snippet.nf
1
2
3
4
5
6
7
8
9
Channel
    .of(1, 2, 3)
    .map { it -> def X = it; X += 2 }
    .println { "ch1 = $it" }

Channel
    .of(1, 2, 3)
    .map { it -> def X = it; X *= 2 }
    .println { "ch2 = $it" }

13.1.5 Non-deterministic input channels

While dataflow channel ordering is guaranteed – data is read in the same order in which it’s written in the channel – be aware that there is no guarantee that the elements will maintain their order in the process output channel. This is because a process may spawn multiple tasks, which can run in parallel. For example, the operation on the second element may end sooner than the operation on the first element, changing the output channel order.

In practical terms, consider the following snippet:

snippet.nf
process FOO {
    input:
    val x

    output:
    stdout

    script:
    """
    sleep \$((RANDOM % 3))
    echo -n "$x"
    """
}

process BAR {
    input:
    val x

    output:
    stdout

    script:
    """
    echo -n "$x" | tr '[:upper:]' '[:lower:]'
    """
}

process FOOBAR {
    input:
    val foo
    val bar

    output:
    stdout

    script:
    """
    echo $foo - $bar
    """
}

workflow {
    ch_letters = channel.of('A', 'B', 'C', 'D')
    FOO(ch_letters)
    BAR(ch_letters)
    FOOBAR(FOO.out, BAR.out).view()

}

Processes FOO and BAR receive the same inputs, and return something on standard output. Somebody wrote process FOOBAR to receive those processed outputs and generate combined output, from matched inputs. But even though we set the order of the input values, the FOO and BAR processes output whenever their tasks are complete, not respecting input order. So FOOBAR receiving the outputs of those channels, will not receive them in the same order, as you can see from the output:

Output
...
B - c

D - a

C - d

A - b

So D is matched with 'a' here, which was not the intention. That order will likely be different every time the workflow is run, meaning that the processing will not be deterministic, and caching will also not work, since the inputs to FOOBAR will vary constantly.

Exercise

Re-run the above code a couple of times using -resume, and determine if the FOOBAR process reruns, or uses cached results.

Solution

You should see that while FOO and BAR reliably re-use their cache, FOOBAR will re-run at least a subset of its tasks due to differences in the combinations of inputs it recieves.

The output will look like this:

Output
 [58/f117ed] FOO (4)    [100%] 4 of 4, cached: 4 ✔
 [84/e88fd9] BAR (4)    [100%] 4 of 4, cached: 4 ✔
 [6f/d3f672] FOOBAR (1) [100%] 4 of 4, cached: 2 ✔
 D - c

 A - d

 C - a

 B - b

A common solution for this is to use what is commonly referred to as a meta map. A groovy object with sample information is passed out together with the file results within an output channel as a tuple. This can then be used to pair samples from separate channels together for downstream use.

To illustrate, here is a change to the above workflow, with meta maps added:

snippet.nf
process FOO {
    input:
    tuple val(meta), val(x)

    output:
    tuple val(meta), stdout

    script:
    """
    sleep \$((RANDOM % 3))
    echo -n "$x"
    """
}

process BAR {
    input:
    tuple val(meta), val(x)

    output:
    tuple val(meta), stdout

    script:
    """
    echo -n "$x" | tr '[:upper:]' '[:lower:]'
    """
}

process FOOBAR {
    input:
    tuple val(meta), val(foo), val(bar)

    output:
    stdout

    script:
    """
    echo $foo - $bar
    """
}

workflow {
    ch_letters = channel.of(
        [[id: 'A'], 'A'],
        [[id: 'B'], 'B'],
        [[id: 'C'], 'C'],
        [[id: 'D'], 'D']
    )
    FOO(ch_letters)
    BAR(ch_letters)
    FOOBAR(FOO.out.join(BAR.out)).view()

}

Now, we define ch_letters with a meta map (e.g. [id: 'A']). Both FOO and BAR pass the meta through and attach it to their outputs. Then, in our call to FOOBAR we can use a join operation to ensure that only matched values are passed. Running this code provides us with matched processes, as we'd expect:

Output
...
D - d

B - b

A - a

C - c

If meta maps are not possible, an alternative is to use the fair process directive. When this directive is specified, Nextflow will guarantee that the order of outputs will match the order of inputs (not the order in which the tasks run, only the order of the output channel).

Warning

Depending on your situation, using the fair directive will lead to a decrease in performance.

Summary

In this step you have learned:

  1. Common reasons why Nextflow may be ignoring your cached results