Intent-driven scheduling of backup jobs

TL;DR

The paper addresses integrating new backup jobs into existing schedules under time-dependent constraints while preserving ongoing operations. It introduces an intent-driven framework that parses admin goals into scheduling parameters and uses a kernel density estimate $F_h(t)$ of current schedules to construct a sampling distribution $G(\alpha,t)$ that guides placement of new windows. A greedy sampling algorithm then selects $k$ new windows while enforcing constraints through domain expansion, edge correction, and affinity adjustments. Evaluated on thousands of NetBackup-like policies with $P$-periodicity and real workload patterns, the approach improves reliability and reduces disruptions, demonstrating practical value for enterprise backup management.

Abstract

Job scheduling under various constraints to achieve global optimization is a well-studied problem. However, in scenarios that involve time-dependent constraints, such as scheduling backup jobs, achieving global optimization may not always be desirable. This paper presents a framework for scheduling new backup jobs in the presence of existing job schedules, focusing on satisfying intent-based constraints without disrupting current schedules. The proposed method accommodates various scheduling intents and constraints, and its effectiveness is validated through extensive testing against a variety of backup scenarios on real-world data from Veritas Netbackup customer policies.

Figures (6)

• Figure 1: Left: Backup client-server topology with $c$ clients; if the number of simultaneous backups is more than the concurrency limit of the server, backup jobs for some clients may fail. Right: Backup schedule for a customer, where 72% backup jobs are clustered around midnight (from a Veritas customer's Netbackup policy configuration.
• Figure 2: From the intent, a small language model extracts parameters for the scheduler, which in turn integrates with the existing job windows, to create a new schedule that is designed to match the user intent.
• Figure 3: Sample $F_h(t)$ with (solid blue) and without (dashed black) resolving edge effects and periodicity. The time axis is hours starting Mon 12:00 PM.
• Figure 4: The original KDE $F_h(t)$ and the distributions $G(\alpha, t)$ for various values of the overlap parameter $\alpha \in [0, 1]$. The time axis is hours starting Mon 12:00 PM.
• Figure 5: Example evolution of $G(\alpha=1,t)$ after two sampling rounds with $\Delta = 15$ and reinforcement $\omega = 0.2$. The time axis is hours starting Mon 12:00 PM.