Here is the postgresql service on a 4 GB limit, straight from its cgroup:

anon              112 MB
file             3915 MB   (inactive_file 2974 MB, reclaimable)
oom_kill            0

docker stats reports 27% for that same container at that same moment, because it subtracts inactive page cache.

The tricky part is that 99% looks identical whether the cache is comfortably full or the container is thrashing. Here it happened to be the latter: the cgroup hit its ceiling 154 times per minute and the container read 117 MB/min from disk. After raising the limit those dropped to 0 and 44 kB/min. Nothing in the percentage told us which of the two situations we were in.

Would it make sense to base the figure on memory.current minus inactive_file, or to show the reclaim and oom counters from memory.events next to it?