Differences from CWL

There is no Directory type in CWL. The one instance of this, vep_cache_dir, has been changed to vep_cache_dir_zip and expects a zipped file of the dir, where the zip contains the contents of the dir not the dir itself.

As far as I can tell there's no equivalent of "InitialWorkDirRequirement" from CWL in WDL. Where we use this to create a script to execute, that scripts contents have been moved to the command block of their task. No difference to the caller.

Input files must prefix each argument with the name of the workflow they're going to run, because a WDL file can contain multiple workflows or pass inputs over a layer if they aren't propagated through in the definition. e.g. to call workflow somaticExome with input foo, yaml key must be somaticExome.foo

WDL does not allow relative imports on the root file, so you'll need a local copy that strips the leading ../ from each import. In CWL a similar issue happened but had a workaround of zipping the deps in a tricky way to allow it. WDL bans it outright.

TODO:

Various tools need space allocation for their outputs

tools/vep.wdl

Actually use custom annotations that are passed in

tools/cnvkit_batch.wdl

Docker version update from 0.9.5 to 0.9.8 causes a regression and we need a solution because it enables CRAM support. Specifically permissions to write are denied for commands like mv, touch, and generated index files within the script. It does work locally, so troubleshooting is probably required before filing a report to CNVkit owners.

Runtime on cloud is slower than cluster, why?

Timing diagrams for the same run between cloud and cluster show that the cloud runs about 50% longer than the cluster. Localization accounts for a portion of this, but task runtimes are also slower on the cloud.