relay: Allocation for compression goes very high

My relay just OOMed some circuits with filled up queue (legacy/trac#25226 (moved)) but then a useful log was printed showing that the compress total allocation is huge.

Feb 27 20:02:55.718 [notice] We're low on memory (cell queues total alloc: 232279872 buffer total alloc: 1937408, tor compress total alloc: 878586075 rendezvous cache total alloc: 4684497). Killing circuits withover-long queues. (This behavior is controlled by MaxMemInQueues.)

That 878586075 = ~838MB. My relay is hovering around 1.4GB of RAM right now which means ~60% of the RAM used is in the compression subsystem.

I'm not sure where it all comes, the relay is serving directory data but I have my doubt that compressed, it comes down to 800+ MB...

Datapoint:

$ du -sh diff-cache/
131M	diff-cache/