A picture is better than words here:

I realized this because I had some CI job erroneously generate 250MB of logs, on workers that have 1GB of RAM. And those workers just stopped uploading logs because /tmp was full (which is another bug tbh).