To efficiently manage serverless computing platforms, a key aspect is the auto-scaling of services, i.e., the set of computational resources allocated to a service adapts over time as a function of the traffic demand. The objective is to find a compromise between user-perceived performance and energy consumption. In this paper, we consider the scale-per-request auto-scaling pattern and investigate how many function instances (or servers) should be spawned each time an unfortunate job arrives, i.e., a job that finds all servers busy upon its arrival. We address this problem by following a stochastic optimization approach: we develop a stochastic gradient descent scheme of the Kiefer-Wolfowitz type that applies over a single run of the state evolution. At each iteration, the proposed scheme computes an estimate of the number of servers to spawn each time an unfortunate job arrives to minimize some cost function. Under natural assumptions, we show that the sequence of estimates produced by our scheme is asymptotically optimal almost surely. In addition, we prove that its convergence rate is$O(n^{-2/3})$where n is the number of iterations. From a mathematical point of view, the stochastic optimization framework induced by auto-scaling exhibits non-standard aspects that we approach from a general point of view. We consider the setting where a controller can only get samples of the transient – rather than stationary – behavior of the underlying stochastic system. To handle this difficulty, we develop arguments that exploit properties of the mixing time of the underlying Markov chain. By means of numerical simulations, we validate the proposed approach and quantify its gain with respect to common existing scale-up rules.

In this paper, we revisit the regret of undiscounted reinforcement learning in MDPs with a birth and death structure. Specifically, we consider a controlled queue with impatient jobs and the main objective is to optimize a trade-off between energy consumption and user-perceived performance. Within this setting, the diameter $D$ of the MDP is $\Omega(S^S)$, where $S$ is the number of states. Therefore, the existing lower and upper bounds on the regret at time $T$, of order $O (\sqrt{DSAT})$ for MDPs with $S$ states and $A$ actions, may suggest that reinforcement learning is inefficient here. In our main result however, we exploit the structure of our MDPs to show that the regret of a slightly-tweaked version of the classical learning algorithm UCRL2 is in fact upper bounded by $\tilde{\mathcal{O}} (\sqrt{E_2AT})$ where $E_2$ is a weighted second moment of the stationary measure of a reference policy. Importantly, $E_2$ is bounded independently of $S$. Thus, our bound is asymptotically independent of the number of states and of the diameter. This result is based on a careful study of the number of visits performed by the learning algorithm to the states of the MDP, which is highly non-uniform.