Development miscellany

Third-party use

We ask that third parties who build upon the codebase to do so from a versioned release. This will help them determine when bug fixes apply and generally make it easier to collaborate. If more intensive modifications happen then we request that the repository is forked, again preferably from a version tag.

Build framework

make should build everything, including tests and "development" code.

git and GitHub strategies

Still in the works, but read this.

Make a branch on dib-lab (preferred so others can contribute) or fork the repository and make a branch there.

Each piece or fix you are working on should have its own branch; make a pull request to dib-lab/master to aid in code review, testing, and feedback.

If you want your code integrated then it needs to be mergeable.

Code coverage

Travis and CodeCov calculate code coverage for every build, and post changes in code coverage to every pull request thread after a successful build.

Code coverage should never go down and new functionality needs to be tested.


All khmer scripts used by a published recommended analysis pipeline must be included in scripts/ and meet the standards therein implied.

Command line scripts

Python command-line scripts should use '-' instead of '_' in the name. (Only filenames containing code for import should use _.)

Please follow the command-line conventions used in scripts/, as described in the scripts and sandbox documentation.

Command line thoughts:

If a input filename is required, typically UNIX commands don't use a flag to specify it.

Also, positional arguments typically aren't used with multiple files.

CTB's overall philosophy is that new files, with new names, should be created as the result of filtering etc.; this allows easy chaining of commands. We're thinking about how best to allow override of this, e.g. <ct file> <filename> [ -o <filename.keep> ]

All code in scripts/ must have automated tests; see tests/ Otherwise it belongs in sandbox/.

When files are overwritten, they should only be opened to be overwritten after the input files have been shown to exist. That prevents stupid command line mistakes from trashing important files.

A general error should be signaled by exit code 1 and success by 0. Linux supports exit codes from 0 to 255 where the value 1 means a general error. An exit code of -1 will get converted to 255.

CLI reading:

Python / C integration

The Python extension that wraps the C++ core of khmer lives in khmer/

This wrapper code is tedious and annoying so we use a static analysis tool to check for correctness.

Developers using Ubuntu Precise will want to install the gcc-4.6-plugin-dev package

Example usage:

--maxtrans=512" python build_ext 2>&1 | less

False positives abound: ignore errors about the C++ standard library. This tool is primarily useful for reference count checking, error-handling checking, and format string checking.

Errors to ignore: "Unhandled Python exception raised calling 'execute' method", "AttributeError: 'NoneType' object has no attribute 'file'"

Warnings to address:

khmer/ note: this function is too complicated
for the reference-count checker to fully analyze: not all paths were

Adjust --maxtrans and re-run.

khmer/ warning: Mismatching type in call to
Py_BuildValue with format code "i" [enabled by default]
  argument 2 ("D.68937") had type
    "long long unsigned int"
  but was expecting
  for format code "i"

See below for a format string cheat sheet One also benefits by matching C type with the function signature used later.

"I" for unsigned int "K" for unsigned long long a.k.a khmer::HashIntoType.

Read handling

Several bugs have gone unnoticed due to inconsistencies in read handling. On the C++ side, there are an abundance of consume functions for loading Fasta/Fastq sequences. On the Python side, read handling is sometimes delegated to the C++ library, and sometimes handled in Python using screed.

In an attempt to normalize read handling in Python, the functions in khmer/ should be used whenever possible. Here, broken_paired_reader in khmer/ should be used to do all paired-end sequence handling, and sequence loading should go through khmer.utils.clean_input_reads(iter); this is a generator that wraps the iterator produced by, and it adds a cleaned_seq attribute to screed Record objects. This attribute should be used for any k-mer or graph operations, while the normal sequence attribute is what should be written out. write_record and write_record_pair should be used to output records. All of these functions are aware of FASTA and FASTQ records, too.

For applying operations to collections of reads, the ReadBundle class is available. This is used to wrap a collection of reads for examination and processing in situations where (for example) something should be done to either both reads in a pair, or neither.

Some basic rules of sequence handling in khmer are:

  • consume and produce "broken paired" format, such that pairs of sequences always stay together; see khmer.utils.broken_paired_reader.
  • when looking at the coverage of reads (for trimming or digital normalization) always consider pairs; see khmer.utils.ReadBundle(...).
  • only apply graph or k-mer operations to sequences consisting only of ATCG; typically this will be record.cleaned_seq. See khmer.utils.clean_input_read(...).
comments powered by Disqus