`NewDataModule` performance when loading large amounts of data values

[emh-jump]

Hello!

We use YTT for some templating and given the size of our templates and the number of times we generate with different inputs, YTT rendering performance is something we pay attention to. I did some quick profiling, and I believe I’ve identified some big performance wins.

I noticed the CPU use for NewDataModule was unusually high (close to 90% of all samples in an internal benchmark), mostly spent converting the yamlmeta.Document argument into a starlark.Value. Since NewDataModule is called for each file evaluation, this adds up quickly when many files are evaluated for a template even though the document doesn't change.

I believe it is safe to do this conversion once within the TemplateLoader and then reuse the result for each new data module. A quick ~10 line proof of concept of this change passes repo tests and our internal tests.

With GOGC=off I see the following stats on my dev machine when running an internal benchmark:

• Mainline develop:
  5.26user 0.11system 0:05.49elapsed 97%CPU (0avgtext+0avgdata 353880maxresident)k
• Convert in TemplateLoader:
  0.55user 0.03system 0:00.60elapsed 97%CPU (0avgtext+0avgdata 247556maxresident)k

I'll plan to open a PR with the proposed change shortly, but wanted to check for any concerns

[emh-jump]

I've been able to reproduce this slowdown with a large amount of data values when loading many starlark files. I created a quick python script to generate a reproduction (didn't think it made sense to share a values file full of random text, but would be happy to do that if prefered) https://gist.github.com/emh-jump/a6fe44aa1bcb7ee613161c0affcb860f

The reproduction is slightly contrived to demonstrate maximum impact, but it does bear similarities to the production rendering mentioned in the original issue.

Mainline develop:
2.11user 0.07system 0:02.20elapsed 99%CPU (0avgtext+0avgdata 415620maxresident)k
#954 :
0.12user 0.02system 0:00.15elapsed 98%CPU (0avgtext+0avgdata 101152maxresident)k