The Pipe Transformer for Parallelizing Command-Line Bioinformatics Tools

Some single-node tools take a long time to run. To accelerate them, Glow includes a utility called the Pipe Transformer to process Spark DataFrames with command-line tools.

The tool supports vcf and txt / csv formatted Spark DataFrames as inputs. And it returns a Spark DataFrame. You can specify a quarantine location for partitions of the DataFrame that error when processed by the bioinformatics tool. This is analagous to how liftOver handles failures caused by edge cases.


Consider a minimal case with a DataFrame containing a single column of strings. You can use the Pipe Transformer to reverse each of the strings in the input DataFrame using the rev Linux command:

Provide options through the arg_map argument or as keyword args.

# Create a text-only DataFrame
df = spark.createDataFrame([['foo'], ['bar'], ['baz']], ['value'])
rev_df = glow.transform('pipe', df, cmd=['rev'], input_formatter='text', output_formatter='text')

The options in this example demonstrate how to control the basic behavior of the transformer:

  • cmd is a JSON-encoded array that contains the command to invoke the program

  • input_formatter defines how each input row should be passed to the program

  • output_formatter defines how the program output should be converted into a new DataFrame

The input DataFrame can come from any Spark data source — Delta, Parquet, VCF, BGEN, and so on.

Integrating with bioinformatics tools

To integrate with tools for genomic data, you can configure the Pipe Transformer to write each partition of the input DataFrame as VCF by choosing vcf as the input and output formatter. Here is an example using bedtools.


The bioinformatics tool must be installed on each virtual machine of the Spark cluster.

df ="vcf").load(path)

intersection_df = glow.transform(
    cmd=['bedtools', 'intersect', '-a', 'stdin', '-b', bed, '-header', '-wa'],

You must specify a method to determine the VCF header when using the VCF input formatter. The option infer instructs the Pipe Transformer to derive a VCF header from the DataFrame schema. Alternately, you can provide the header as a blob, or you can point to the filesystem path for an existing VCF file with the correct header. For a more complex example using The Variant Effect Predictor (VEP) see the notebook example below.


Option keys and values are always strings. You can specify option names in snake or camel case; for example inputFormatter, input_formatter, and InputFormatter are all equivalent.




The command, specified as an array of strings, to invoke the piped program. The program’s stdin receives the formatted contents of the input DataFrame, and the output DataFrame is constructed from its stdout. The stderr stream will appear in the executor logs.


Converts the input DataFrame to a format that the piped program understands. Built-in input formatters are text, csv, and vcf.


Converts the output of the piped program back into a DataFrame. Built-in output formatters are text, csv, and vcf.


Spark SQL table to write partitions in the dataframe that throw an error.


File type for quarantined output. Built-in output formatters are csv and delta.


Options beginning with env_ are interpreted as environment variables. Like other options, the environment variable name is converted to lower snake case. For example, providing the option env_aniMal=MONKEY results in an environment variable with key ani_mal and value MONKEY being provided to the piped program.

Some of the input and output formatters take additional options.

VCF input formatter




How to determine a VCF header from the input DataFrame. Possible values:

  • infer: Derive a VCF header from the DataFrame schema. The inference behavior matches that of the sharded VCF writer.

  • The complete contents of a VCF header starting with ##

  • A Hadoop filesystem path to a VCF file. The header from this file is used as the VCF header for each partition.

The CSV input and output formatters accept most of the same options as the CSV data source. You must prefix options to the input formatter with in_, and options to the output formatter with out_. For example, in_quote sets the quote character when writing the input DataFrame to the piped program.

The following options are not supported:

  • path options are ignored

  • The parserLib option is ignored. univocity is always used as the CSV parsing library.


The pipe transformer uses RDD caching to optimize performance. Spark automatically drops old data partitions in a least-recently-used (LRU) fashion. If you would like to manually clean up the RDDs cached by the pipe transformer instead of waiting for them to fall out of the cache, use the pipe cleanup transformer on any DataFrame. Do not perform cleanup until the pipe transformer results have been materialized, such as by being written to a Delta Lake table.

glow.transform('pipe_cleanup', df)


The examples below show how to parallelize Bedtools, Plink and VEP.


Please troubleshoot pipe transformer errors by inspecting the stderr logs for failed tasks via: Spark UI -> Stages -> Failed Stages -> Description -> Logs -> stderr


Bedtools shuffle and intersect are two bedtools commands suited to the pipe transformer.


The VEP example shows how to quarantine corrupted records. This functionality was introduced from Glow v1.1.2.

Pipe Transformer bedtools example notebook

Pipe Transformer Variant Effect Predictor (VEP) example notebook