This paper applies the longstanding notion of caching to the new field of serverless computing. Moving computation to the data, that is, functions to the nodes where libraries are cached, is advantageous to function-as-a-service (FaaS) functions that are ephemeral in nature. However, there are certain limitations to the process.

A few concerns:

(1) The authors fail to declare that this idea is only favorable to functions that encounter a cold start.

(2) The scheduling concepts have been described only for the pull-based model with favorable results.

(3) There are inherent security concerns with scheduling functions with similar packages on the same nodes and sharing processes between containers.

(4) In the pull mechanism, it is stated that whenever the node has “spare capacity” it can pull work from the respective queues. What this spare capacity means is not clearly explained.

(5) Another limitation of this model is the choice of language. It only works for Python.

(6) The results only show the impact of scheduling on the end-to-end latency of the task.

Some shortcomings:

(1) The queue assignment algorithm does not consider if the target is overloaded.

(2) First in, first out (FIFO) ordering loses the ability to cater to prioritized functions.

(3) The preliminary evaluation in section 4 outlines all the configuration parameters, but fails to describe the function or computation tasks.

I would recommend the paper. Its target audience is primarily cloud service providers who are evaluating FaaS for possible inclusion in their service models. Academics and developers should also benefit from this.