sharedcache: Guidelines for sharing files in a team data analysis

Description Cache sharing strategy

Description

For these purposes, we refer to files stored in a shared location as existing on a 'shared cache'. Although the term 'cache' highlights the changeable nature of the files in this location, shared files could include raw data, munged intermediate data products, model outputs, or even final reports. They could be stored on Amazon S3, on a shared network drive, in a project directory on a computing cluster, or on a local computer to which all team members have file access.

Cache sharing strategy

Every file (cached and uncached) is assigned a corresponding lightweight indicator file of the same name but with the additional extension '.ind'.

remake.yml targets should declare dependencies on the indicator files rather than among the data files. The data files can be included as targets in the remake but should only depend on their own indicator files, and the target command should not generate or update the data file from other data files; instead, it should simply pull an existing cached version of the file into local storage.

Behind the scenes, remake creates an additional build status file for each target that gets made. Most of these stay local, but calling scmake() instead of remake::make() will create git-committable versions of certain build status files (exactly those corresponding to indicator files) in the build/status subdirectory of your project.

Always git commit the indicator files and the build/status files, but not the data files.

Your collaborators will generally be unaware of any changes to the cache until you commit and they pull a changed indicator file. The only exception is when they need to download a file from the cache (with need based on the files already on their computers and the targets they're trying to build) - if you've changed what's in the cache, they'll get the new version even if you haven't yet announced the change via git. In the probably-rare cases where this inconsistency will create problems, it's a good idea to communicate with teammates by other channels (email, talking, etc.) while the cache-indicator mismatch persists.

As a team member, it's your responsibility to git pull and push as often as is practical, especially when you might be obtaining or altering cached files. An up-to-date shared understanding of what's in the cache will reduce the amount of redundant processing that's required on everybody's computers.


USGS-R/scipiper documentation built on Aug. 14, 2018, 5:23 a.m.