dmk

README

dmk

This is a simplified, automated build tool for data projects.

The idea behind dmk is to support build files that are easy to read and write, and to support automating a build system for data artifacts. make, scons, and fac all provided inspiration for dmk.

Who should use this?

This is a tool for data flows and simple projects. Often these projects involve one-time steps for getting the data used for analysis. Often the user manually places that data in the directory after downloading it from Amazon S3 or a research server or whatever. Scripts or programs run in a pipeline fashion: handling cleaning, transformation, analysis, model building, figure production, presentation building, etc.

Pipelines like this are often written Python or R, partially automated with shell scripts, and then tied together with a Makefile (or a SConstruct file if you're an scons fan).

Protip: if you're looking for a command to handle building reports from .tex files (including handling metapost and biblatex), look into rubber.

What is this NOT for?

This is not mean to replace a real automated build tool for a software project. As a general rule:

• If you're building Go software, use the Go tools (and optionally make)
• If you're building Java/Scala use sbt, gradle, mvn, ant, etc
• If you're building .NET, erm, I'm not sure
• There are great tools like scons that understand how to build lots of artifacts (including LaTeX docs)
• If you're not sure, at least understand why you wouldn't use make

For instance, this project (written in Go) is actually built with make + the standard Go tools.

Using

When running dmk, you may specify -h on the command line for information on command line parameters. Specify -v for "verbose" mode to get output you may want for debugging or understanding what is going on.

For each command in a pipeline, you need to supply:

• A name
• The actual command to run (in the shell)
• The inputs required
• The outputs generated

This list is not exhaustive; see below for everything you can specify for a build step.

The file is generally named Pipeline or pipeline.yaml. If you do not specify a pipeline file with the -f command line parameter, dmk looks for the following names in the current directory (in order):

• Pipeline
• Pipeline.yaml
• pipeline
• pipeline.yaml
• .Pipeline.yaml
• .pipeline.yaml

You may also supply a custom name with the -f command line flag. If the pipeline file is in a different directory, dmk will change to that directory before parsing the config file.

All build steps run in parallel, but each step waits until other steps build its dependencies. A single build step executes the following steps:

1. The step is "Started"
2. If any of the required inputs are another step's outputs, then wait for a built message.
3. Check to see if any outputs are older than any of the inputs. If not, then the step is "Completed"!
4. If not done, set status to "Executing" and run the command.
5. If the command returns an error code or if any outputs are missing or older than any inputs, the step is "Failed".
6. If not done, send notification messages for each output for any waiting steps.
7. The step is now "Completed"

The outputs for a step must be unique to that step: you can't have two steps both list foo.data as an output.

dmk provides an automatic "clean" mode that deletes all outputs. To use it, specify -c on the command line. dmk will delete all the outputs for all steps. If you have files to clean not specified as outputs, you can specified them in the clean list for a build step (see the Pipeline file format below). Good candidates for the clean section are intermediate files (such as logs) generated as part of a build process that are not dependencies and should not determine if a build step is up to date.

Pipeline file format

The file is in YAML format where each build step is a named hash. Each build step should specify:

• command - The command to run as part of the build. dmk uses bash to run the command, so it can rely on bash shell niceties (like using ~ for the home directory)
• inputs - a list of inputs needed for the build. These are also the dependencies that must exist before the step can run. An entry can be a glob pattern (like *.txt)
• outputs - a list of outputs generated by the step. Outputs decide if the step must run, and the clean phase deletes them. Glob patterns are ignored for outputs.
• clean - A list of files to clean. These and outputs are the files deleted during a clean. You may use glob patterns for these.
• explicit - Optional, defaults to false. If set to true, the step will run if you specify it on the command line. It will not run by default. Any steps required by steps specified on the command line will also run, regardless of their explicit setting.
• delOnFail - Optional, defaults to false. If set to true and the step fails, then dmk will delete all the step's output files.
• direct - Optional, default to false. If set to true, both stdout and stderr from the stepis written to the dmk process standard streams. If set to false (the default), stdout and stderr are written in single blocks after the step completes (stdout is only written if dmk is running in verbose mode). Note in direct mode (direct=True), step output may be intereaved with "normal" output when steps are running in parallel!

The res subdirectory contains sample Pipeline files (used for testing), but a quick example would look like:

# You can have comments in a file
step1:                                # first step
    command: "xformxyz i{1,2,3}.txt"  # command with some shell magic
    inputs:                           # 3 inputs (read by our imaginary command)
        - i1.txt                  
        - i2.txt
        - i3.txt
    outputs:                          # 3 outputs
        - o1.txt
        - o2.txt
        - o3.txt
    clean: [a.aux, b.log]             # two extra clean targets, specified in
                                      # an alternate syntax for YAML lists

step2:                                # second step
    command: cmd1xyz                  # note the lack of inputs - this means
    outputs:                          # the step will run without waiting for
        - output.bin                  # other steps to complete.

depstep:                              # third/final step: it won't run until the
    command: cmd2xyz                  # previous steps finish because their
    inputs:                           # outputs are in the this step's inputs.
        - o3.txt                      
        - output.bin
    outputs:
        - combination.output
    clean:
        - need-cleaning.*             # An example of using a glob pattern
    delOnFail: true

extrastep:
    command: special-command
    inputs:
        - some-script-file.txt
    outputs:
        - my-special-file.extra
    explicit: true                    # Run if specified on command line (and not by default)

If you were to run dmk -c then it would deleted the following files:

• o1.txt, o2.txt, o3.txt, a.aux, and b.log because of step1
• because of the clean list in step1
• output.bin because of step2
• combination.output and any files matching the pattern need-cleaning.* because of depstep

Note that my-special-file.extra from extrastep is not deleted unless you specify extrastep on the command line.

After cleaning, if you run dmk the following steps would occur:

• The commands from step1 (xformxyz i{1,2,3}.txt) and step2 (cmd1xyz) would run
• When they were both finished, depstep would start and cmd2xyz would run.
• As before, extrastep would NOT run.
• If the depstep command (cmd2xyz) fails, then dmk will delete combination.output (if it exists).
• If all the steps succeed, running dmk again would not cause any command to run (because all outputs are newer than their steps' inputs).

If you were to run dmk extrastep then the command special-command would run. Nothing else would run.

If you were to run dmk extrastep depstep then all steps would run (because step1 and step2 are depstep dependencies).

Building

godep manages dependencies in the vendor directory. You shouldn't need to worry about this if you are building with the Makefile. Also note the fact that we use make to build dmk. We are serious about using the correct build tool for the job.

You should also have Python 3 installed (for script/update and for the test script res/slow).

make dist will build cross-platform binaries in ./dist. Yes, we commit them to the repo. Deal with it, they're small.

make release handles tagging and pushing to GitHub.

Build step environment

11 When a build step runs, dmk sets environment variables in the step command's process:

• DMK_VERSION - version string for dmk
• DMK_PIPELINE - absolute path to the pipeline file running
• DMK_STEPNAME - the name of the current step
• DMK_INPUTS - a colon (":") delimited list of inputs for this step
• DMK_OUTPUTS - a colon (":") delimited list of outputs for this step
• DMK_CLEAN - a colon (":") delimited list of extra clean files for this step

Also note that because bash evaluates the command, you can use the environment variables in the command itself. E.g. mycmd --inputs $DMK_INPUTS

Some helpful hints to remember

A pipeline file is a YAML document, and a JSON document is valid YAML. For instance, res/slowbuild.yaml and res/slowbuild.json are semantically identical pipeline files. If you need a customized build, you can generate the pipeline file in the language of your choice in JSON or YAML and then call dmk.

Commands run in a new bash shell (which also means you need bash).

dmk changes to the directory of the Pipeline file, so you can specify file names relative to the Pipeline file's directory.

You may use globbing patterns for the inputs and clean.