targets VS fastverse

Is this it https://github.com/ropensci/targets?

I'll chime in with others to say that using targets can help with the memory load as well. If you partition your data adequately (e.g. grouping by subjects), you can take advantage of the way targets maps data so it only loads what it needs to. Moreover, if you use the memory = "transient" option, it will unload objects between steps -- adding a little bit of time overhead but saving you on memory. targets and tidytable together have enabled me to work on pretty sizeable datasets while rarely running into memory issues. In fact, the only time I ran into a data memory hog was because I didn't adequately partition the data across worker nodes.

targets

I highly recommend the holy grail of workflow orchestrators / executors in the R ecosystem: targets.

That said, you might want to check out the ‘targets’ package, which provides a DSL for specifying complex workflow descriptions in R. When repeatedly running the same jobs on changing data, this package helps ensure that only necessary work is performed (suitable intermediate results are reused), and scripts are run reproducibly. This might help with sceduling.

It may be worth it to put these individual steps into a targets pipeline. targets is designed to support parallelization with future and make it easier to visualize downstream dependencies.

Given your requirements, I recommend you check out ‘targets’, which specifically addresses the needs of reusable workflows in R, and it seems like it fits your requirements to a T.

The targets package is the official successor to Drake, and has the same primary author (Will Landau). He has explained why he created targets, which includes stronger guardrails for users and better UX.

I use it for ETL. I use targets as the workflow management software, and, like others, have a cron job set up to run nightly builds.

For writing faster code, the first thing you want to try is making sure it's properly vectorised. See The R Inferno for this. Some problems are more difficult than others to vectorise. When vectorisation is impossible, you probably want to interface with C++. First, check if there's already a fast package that serves your needs. If your problem is too specific, consider writing your own C++ code with Rcpp.

I understand the sentiment and I'm not trying to convince you to start writing optimised code to save ~2ms. There's a ton of optimised tools that I don't use myself because the time benefit is immaterial for what I do.