ci: switch to GitHub Actions

metrumresearchgroup / mrgsolve

Simulate from ODE-based population PK/PD and QSP models in R

https://mrgsolve.org

129 stars 36 forks source link

ci: switch to GitHub Actions #1189

Closed kyleam closed 4 months ago

kyleam commented 4 months ago

This series switches the CI from Drone to GitHub Actions. Summary of keys differences:

OS
- new: Ubuntu 20.04 and 22.04
- old: custom image based on Ubuntu 18.04
R
- new: 3.6, 4.0, 4.1, and latest R (currently 4.4.0)
- old: 3.6, 4.0, and 4.1

The fourth commit makes the main change (and its commit message has more details about the differences compared to the Drone builds).

[x] Drop tip commit that tests dry-run s3 sync.
[ ] Coordinate with open pull requests. Once this is merged, these PRs should bring this in (by merge or rebase) to move away from Drone.

kyleam commented 4 months ago

Test s3 auth and dry-run upload look good:

I'll drop the tip commit.

kyleam commented 4 months ago

The uncached vs cached time for the R 3.6.3 build is unfortunate: ~25 min (link) vs ~7 min (link).

That discrepancy isn't there for the other R versions. I suspect the difference is that, for the builds other than 3.6.3, most things are coming as binary packages from RSPM. For R 3.6.3, an older RSPM is used to get a compatible pmxTools. Given the upgrade=TRUE, I think pak is then preferring newer packages from CRAN, leading to the slower times. (It's possible also possible that there are no RSPM binaries available for that R/date/OS combination.)

Anyway, I'll try to disable upgrade=TRUE for the 3.6 build to see if it improves things.

kyleam commented 4 months ago

I said:

I'll try to disable upgrade=TRUE for the 3.6 build to see if it improves things.

That did the trick. Brought the uncached 3.6 run down from ~25m to ~7m: https://github.com/metrumresearchgroup/mrgsolve/actions/runs/8941856925/job/24563805652

kylebaron commented 4 months ago

Brought the uncached 3.6 run down from ~25m to ~7m

Nice.