Dockerfile image size #98
Description
This is forked from issue #96.
Before the recent update to R 4.5.2, I did some playing with local builds of images:
REPOSITORY TAG IMAGE ID CREATED SIZE
rtr-452 latest c2d4dfb04bef 4 hours ago 2.87GB
rtr-current latest 2bb37d8512fc 4 hours ago 2.39GB
rocker/tidyverse 4.5.2 fa0e5da2d9ea 4 weeks ago 2.86GB
rocker/r-ver 4.5.2 44e5a15c1df9 4 weeks ago 947MB
rhub/r-minimal latest e92270504c84 3 months ago 47.3MB
In the table, rtr-current was the 4.3.1 image we've been using for the previous 3 years, rtr-452 is very close to the build that went live on Exercism earlier today.
These are huge! And cost Exercism money every month!
We're currently based on rocker/tidyverse, which is where the hugeness comes from. Anything else in our Dockerfile is a rounding error.
Exercism really can't use all of Tidyverse. In particular, ggplot2 is a great graphing package, but the test runner isn't going to do anything useful with it.
These are the listed imports in the tidyverse package:
Imports: broom (>= 1.0.3), conflicted (>= 1.2.0), cli (>= 3.6.0),
dbplyr (>= 2.3.0), dplyr (>= 1.1.0), dtplyr (>= 1.2.2), forcats
(>= 1.0.0), ggplot2 (>= 3.4.1), googledrive (>= 2.0.0),
googlesheets4 (>= 1.0.1), haven (>= 2.5.1), hms (>= 1.1.2),
httr (>= 1.4.4), jsonlite (>= 1.8.4), lubridate (>= 1.9.2),
magrittr (>= 2.0.3), modelr (>= 0.1.10), pillar (>= 1.8.1),
purrr (>= 1.0.1), ragg (>= 1.2.5), readr (>= 2.1.4), readxl (>=
1.4.2), reprex (>= 2.0.2), rlang (>= 1.0.6), rstudioapi (>=
0.14), rvest (>= 1.0.3), stringr (>= 1.5.0), tibble (>= 3.1.8),
tidyr (>= 1.3.0), xml2 (>= 1.3.3)
I think I need to:
- Work out which are useful to us and which are just bloat.
- Try some test builds from a smaller base image, installing just the packages we need on top.
I'm not totally optimistic. Installing Tidyverse locally on a Linux system (Mint 22.3 in my case) is quite slow and painful: everything is compiled from C++, and there are a strange variety of system dependencies.
Whatever, we don't finish unless we start...