Speed improvements in manifest writing/deployment

[slodge]

In our environment we're experiencing a lot of user complaints about how slow RStudio Connect is.

Their current issues mainly seem to be linke with "deployment time" - which I think is mainly down to packrat.

We've made some investigations locally and I've made one suggestion for packrat - see rstudio/packrat#603

I'd also be interested in seeing #404 - as this might help us improve deployment times too.

Are there other things we can try in order to improve performance?

[kevinushey]

I think the best thing we could do here is switch to using renv for deployments. (We could have renv create a Packrat-compatible lockfile to help ease the change)

[slodge]

Thanks. I agree that renv feels like the right direction for a proper fix/improvement.

However, I 'm assuming that renv integration is going to take a little longer to achieve, so was wondering if there are other measures I can take (like the PR above) to make things better in the short term for my complaining users.

[kevinushey]

You're correct. I took a quick look at your PR, and given that it doesn't change the default behavior I'm happy to take it. Thanks!

I think the main caveat is that users who take advantage of this feature may have to explicitly list their recursive package dependencies in addition to their top-level dependencies, though (e.g. in a file dependencies.R).

[aronatkins]

@slodge - With rstudio/packrat#615, packrat does a much better job at avoiding repeatedly processing the same package dependency. Are you able to test performance against the latest GitHub version of packrat?

[slodge]

I did a quick test on a sample app in our Azure Devops CI pipeline: A build/deployment yesterday included 2021-01-26 16:47:24: Writing manifest for dir apps/ForecastDashboard with primary: 2021-01-26 16:48:46: Writing manifest complete The same build today with packrat head is.... 2021-01-27 17:48:47: Writing manifest for dir apps/ForecastDashboard with primary: 2021-01-27 17:48:53: Writing manifest complete So - 82 seconds down to 6 The app deployed fine (and runs) The deployment is against RStudio Connect(1.8.0.4-21) I'll do some more testing tomorrow - I need to do a little fiddling to quickly update all our build pipelines to the same packrat version. All our apps are similar - shiny+tidyverse+internal packages - so I'd expect them to be similar in performance. - but I'll see if I can spot any anomalies. Thanks! Stuart

[slodge]

More experiments today.

Similar sort of scale of results
e.g. 1m40 down to 9 seconds

No functional change observed.

These builds are done within "clean" Docker container images so feel quite a solid platform.

[aronatkins]

@slodge Thanks so much for those experiments! This feels like enough of an improvement that I'm going to close this issue tracking the underlying performance problem with the packrat dependency analysis. Use of renv by rsconnect is tracked separately (including #471).