Profiling and Testing Loqui

[howardbaik]

Performance workflow

Estimated peak traffic: 10~20 concurrent users per R process.

Is Loqui fast enough?

Using shinyloadtest, we can generate large amounts of realistic traffic to Loqui, then analyze latency. I ran a load test by inputting a Google Slides URL and rendering video with 10 workers and eventually got this Event Waterfall plot:

and these boxplots for Event Duration:

Key takeaways

The example videos on Loqui and the HTML display of the Loqui-rendered video are taking up lots of time.

What's making Loqui slow?

Using profvis, I found that httr::GET(url, httr::write_disk(tmp)) and the preceding code to create a tempfile, tempfile(fileext = paste0(".", type)) is taking up the most time.

cc @cansavvy

Resources:

1. Talk on shinyloadtest and profvis in Shiny: Joe Cheng | Shiny in production: Principles, practices, and tools | RStudio (2019)
2. The Optimization Loop

[cansavvy]

This is great investigation work. Thanks for putting this down in an issue @howardbaek ! Have you done an initial search to see if there's a more efficient version of httr::GET(url, httr::write_disk(tmp)) we could use? Even if that might me using Python?

[howardbaik]

I realized that I was unnecessarily repeating calls to gsplyr::download(), which runs httr::GET(url, httr::write_disk(tmp)) under the hood, so I dealt with that in this commit 71a6c93