diff --git a/domains/internet-finance/hysteresis-in-autoscaling-prevents-oscillation-by-using-asymmetric-thresholds-for-scale-up-and-scale-down.md b/domains/internet-finance/hysteresis-in-autoscaling-prevents-oscillation-by-using-asymmetric-thresholds-for-scale-up-and-scale-down.md new file mode 100644 index 00000000..1bd4949d --- /dev/null +++ b/domains/internet-finance/hysteresis-in-autoscaling-prevents-oscillation-by-using-asymmetric-thresholds-for-scale-up-and-scale-down.md @@ -0,0 +1,42 @@ +--- +type: claim +domain: internet-finance +description: "Different thresholds for adding versus removing resources prevent rapid oscillation in auto-scaling systems" +confidence: proven +source: "Tournaire et al., 'Optimal Control Policies for Resource Allocation in the Cloud' (2021); established operations research principle" +created: 2026-03-11 +--- + +# Hysteresis in autoscaling prevents oscillation by using asymmetric thresholds for scale-up and scale-down + +Hysteresis in auto-scaling systems—using different thresholds for scaling up versus scaling down—prevents oscillation where resources are rapidly added and removed in response to workload fluctuations near a single threshold. + +For example, a system might scale up when queue length reaches 10 but only scale down when queue length drops to 3. This asymmetry creates a "dead zone" between thresholds that absorbs short-term fluctuations without triggering scaling actions. + +Tournaire et al. (2021) demonstrate this principle in cloud VM provisioning, where MDP-based optimal control policies automatically discover the optimal hysteresis gap given cost structure (energy + SLA violations). The principle is well-established in operations research and control theory more broadly. + +## Why Hysteresis Works + +Without hysteresis, a system operating near a single threshold (e.g., scale at queue=5) will constantly add and remove resources as the queue fluctuates around that value. Each scaling action has overhead cost (VM startup time, worker initialization, context switching), making oscillation expensive. + +Hysteresis trades increased resource utilization during the dead zone (queue between 3-10 in the example) for reduced scaling overhead and more stable operation. + +## Application to Pipeline Management + +For autonomous pipeline workers: +- Scale up threshold: unprocessed queue > N sources +- Scale down threshold: unprocessed queue < M sources (where M < N) +- Dead zone width (N-M) should be tuned to workload volatility and worker startup cost + +The optimal gap depends on: +- Worker initialization time (longer startup → wider gap) +- Cost per worker-minute (higher cost → narrower gap, more aggressive scaling down) +- Workload volatility (higher variance → wider gap to avoid thrashing) + +--- + +Relevant Notes: +- [[mdp-based-autoscaling-with-hysteresis-outperforms-simple-threshold-heuristics-for-cloud-resource-allocation]] + +Topics: +- domains/internet-finance/_map diff --git a/domains/internet-finance/mdp-based-autoscaling-with-hysteresis-outperforms-simple-threshold-heuristics-for-cloud-resource-allocation.md b/domains/internet-finance/mdp-based-autoscaling-with-hysteresis-outperforms-simple-threshold-heuristics-for-cloud-resource-allocation.md new file mode 100644 index 00000000..b1f53cd3 --- /dev/null +++ b/domains/internet-finance/mdp-based-autoscaling-with-hysteresis-outperforms-simple-threshold-heuristics-for-cloud-resource-allocation.md @@ -0,0 +1,38 @@ +--- +type: claim +domain: internet-finance +description: "Structured MDP algorithms that incorporate hysteresis properties achieve better performance and faster execution than simple threshold heuristics in cloud VM provisioning" +confidence: likely +source: "Tournaire et al., 'Optimal Control Policies for Resource Allocation in the Cloud' (2021)" +created: 2026-03-11 +--- + +# MDP-based autoscaling with hysteresis outperforms simple threshold heuristics for cloud resource allocation + +Markov Decision Process formulations that incorporate hysteresis properties (different thresholds for scaling up versus scaling down) outperform simple threshold heuristics in both execution time and accuracy for cloud auto-scaling problems. The MDP approach automatically discovers optimal hysteresis thresholds rather than requiring manual tuning. + +The problem formulation treats VM provisioning as a sequential decision problem where: +- States = queue lengths + active VMs +- Actions = add/remove VMs +- Rewards = negative cost (energy + SLA violations) + +Value iteration and policy iteration algorithms find optimal threshold policies that prevent oscillation by using different thresholds for scaling up (e.g., queue=10) versus scaling down (e.g., queue=3). + +Tournaire et al. (2021) demonstrate that structured MDP algorithms incorporating hysteresis properties outperform heuristic approaches in both execution time and accuracy. The key insight is that hysteresis—different thresholds for scaling up versus scaling down—prevents oscillation, and MDP algorithms can discover these optimal thresholds automatically rather than through manual tuning. + +## Relevance to Pipeline Architecture + +This formulation maps directly to autonomous pipeline management: +- States = (unprocessed queue, in-flight extractions, open PRs, active workers) +- Actions = (spawn worker, kill worker, wait) +- Cost = (Claude compute cost per worker-minute + delay cost per queued source) + +The hysteresis insight is particularly valuable for preventing worker thrashing in variable-load scenarios. Simple threshold policies (scale up at queue=N, scale down at queue=M where M < N) provide reasonable baseline performance, but MDP optimization can find better thresholds given cost structure and workload patterns. + +--- + +Relevant Notes: +- domains/internet-finance/_map + +Topics: +- domains/internet-finance/_map diff --git a/inbox/archive/2021-04-00-tournaire-optimal-control-cloud-resource-allocation-mdp.md b/inbox/archive/2021-04-00-tournaire-optimal-control-cloud-resource-allocation-mdp.md index cc38341b..ee0edb3b 100644 --- a/inbox/archive/2021-04-00-tournaire-optimal-control-cloud-resource-allocation-mdp.md +++ b/inbox/archive/2021-04-00-tournaire-optimal-control-cloud-resource-allocation-mdp.md @@ -6,8 +6,13 @@ url: https://arxiv.org/abs/2104.14879 date: 2021-04-30 domain: internet-finance format: paper -status: unprocessed +status: processed tags: [pipeline-architecture, operations-research, markov-decision-process, cloud-autoscaling, optimal-control] +processed_by: rio +processed_date: 2026-03-11 +claims_extracted: ["mdp-based-autoscaling-with-hysteresis-outperforms-simple-threshold-heuristics-for-cloud-resource-allocation.md", "hysteresis-in-autoscaling-prevents-oscillation-by-using-asymmetric-thresholds-for-scale-up-and-scale-down.md"] +extraction_model: "anthropic/claude-sonnet-4.5" +extraction_notes: "Extracted two claims about MDP-based autoscaling and hysteresis principles. The source is an operations research paper on cloud resource allocation that maps directly to pipeline worker management. The hysteresis insight is particularly valuable for preventing worker thrashing. No entities to create (academic paper, no companies/products/markets). No enrichments to existing claims (this is a novel technical domain for the KB)." --- # Optimal Control Policies for Resource Allocation in the Cloud @@ -28,3 +33,9 @@ Compares MDP-based optimal scaling policies against heuristic approaches for clo The MDP formulation maps directly: states = (unprocessed queue, in-flight extractions, open PRs, active workers), actions = (spawn worker, kill worker, wait), cost = (Claude compute cost per worker-minute + delay cost per queued source). The hysteresis insight is particularly valuable — we should have different thresholds for spinning up vs. spinning down workers to prevent oscillation. Key finding: structured MDP with hysteresis outperforms simple threshold heuristics. But even simple threshold policies (scale up at queue=N, scale down at queue=M where M < N) perform reasonably well. + + +## Key Facts +- MDP formulation for cloud autoscaling: states = queue lengths + active VMs, actions = add/remove VMs, rewards = negative cost (energy + SLA violations) +- Value iteration and policy iteration algorithms used to find optimal threshold policies +- Example hysteresis thresholds: scale up at queue=10, scale down at queue=3