Per-client resource allocation?

[malcolmstill]

As it stands:

• all foxwhale resource allocation (juxtaposed to the allocation of object ids) occur at a global level
• if we out-of-memory (OOM) the client that happens to hit the OOM will be killed.

I was thinking that perhaps resources should be allocated on a per-client basis. This would mean the greedy client would get killed, with the added complexity (is it actually more complex) would get killed. Which would be nice.

[malcolmstill]

With the above said, I am wondering whether to keep the global allocation, but rather than just killing the client that see the OOM, could we go out of our way and find the client that is misbehaving and kill that instead.

That seems nice, but that would introduce the complexity of having the client that innocently hit OOM somehow rerun whatever was attempting to allocate memory in the first place.