GHA: ccache not efficient?

[pitrou]

Bug description:

1. If you look at the ccache stats in ccache-enabled Github Actions jobs, you'll notice that the cache seems hardly useful most of the time, as there are few hits and many misses.

Example here: https://github.com/python/cpython/actions/runs/8332366775/job/22801352882

ccache stats
  /usr/bin/ccache -s
  Summary:
    Hits:              52 /  341 (15.25 %)
      Direct:          30 /  343 (8.75 %)
      Preprocessed:    22 /  313 (7.03 %)
    Misses:           289
      Direct:         313
      Preprocessed:   291
    Uncacheable:      104
  Primary storage:
    Hits:             387 /  685 (56.50 %)
    Misses:           298
    Cache size (GB): 0.18 / 0.20 (90.22 %)
    Cleanups:           8

Most likely this is because all those builds clash for the same cache keys even though they use different compiler options. The solution would be to use per-build cache keys.

2. The "Post Configure ccache action" step itself takes ~2 minutes (this seems unexpected), which is as long as it takes to compile CPython. So it's not obvious there is a benefit at all, even if the problem above was fixed.

[pitrou]

@hugovk

[hugovk]

cc @encukou and @erlend-aasland who've also been looking into ccache (#113858).

Time in seconds for each step:

Step	with ccache	without ccache
Set up job	1	1
Run actions/checkout@v4	5	5
Runner image version	0	0
Restore config.cache	0	0
Install Dependencies	18	80
TSAN Option Setup	0	0
Add ccache to PATH	0	0
Configure ccache action	2	n/a
Configure CPython	4	79
Build CPython	97	93
Display build info	5	7
Tests	130	125
Post Configure ccache action	136	n/a
Post Restore config.cache	0	0
Post Run actions/checkout@v4	0	0
Complete job	0	0
Total	398	390

The "Install Dependencies" 18s vs 80s looks unrelated, so the total without ccache could be seen here as ~330s instead of 390s?

Let's just look at ccache/configure steps:

• With ccache, "Configure ccache action" + "Post Configure ccache action" = 2+136 = +138s.
• Without ccache, "Configure CPython" is +75s.

So with ccache costs an extra ~63s over without cccache.

[pitrou]

By the way, the slowness of the "Post Configure ccache action" doesn't match our experience on Apache Arrow, where saving a 1GB ccache can take 20 seconds. Example here with the "Post Cache Docker Volumes" step:
https://github.com/apache/arrow/actions/runs/8329904897/job/22793369592

Instead of using hendrikmuhs/ccache-action, you might look into using the cache action directly. Setting up ccache is not difficult.