AdaptLSTM: Efficient Adaptive Online Learning for Cloud Workload Forecasting under Distribution Drift

Xinhua Miao; Bowei Yang; Zhengong Cai

cs.LGOct 8, 2026

AdaptLSTM: Efficient Adaptive Online Learning for Cloud Workload Forecasting under Distribution Drift

Authors: Xinhua Miao, Bowei Yang, Zhengong Cai

Organizations: Zhejiang University Ningbo, China · Zhejiang University Hangzhou, China

Abstract

Accurate workload forecasting is critical for elastic resource provisioning in web-scale cloud services, where distribution shifts driven by viral content, product launches, and user behavior degrade offline-trained models rapidly. Naive online learning recovers accuracy but incurs prohibitive per-step compute cost. We propose AdaptLSTM, an adaptive online framework that detects drift via validation-calibrated thresholds and applies selective, targeted updates. On the Alibaba Machine Trace, AdaptLSTM recovers 54% of Naive Online's improvement at 20% cost ( $2.7\times$ efficiency, $p=0.002$ over 10 seeds). On the more volatile Container Trace, it achieves 96% at 20% cost ( $4.8\times$ efficiency, $+75\%$ MAE reduction over Static). Unlike classical drift detectors (ADWIN, DDM, Page-Hinkley) which fail to trigger on regression-scale error streams, AdaptLSTM fires 42 times over 301 steps and outperforms matched-budget baselines. Wall-clock profiling shows $1.33\times$ throughput gain and 45% update-time reduction. The framework is model-agnostic: identical Pareto patterns hold for LSTM, GRU, and Transformer backbones.

Figures & tables

Figure 1. Motivation: MAE under sudden drift on the Alibaba Machine Trace. Static suffers persistent post-drift error, while Naive Online recovers accuracy by updating at every step (100% cost). AdaptLSTM selectively updates under drift (20% budget), closing over half of the Static

\to

Naive gap while approaching Naive’s MAE.

Figure 2. AdaptLSTM architecture. (Top) The base LSTM $f_{\theta}$ predicts $\hat{\mathbf{Y}}_{t}$ from a sliding window $\mathbf{X}_{t}$ and, once ground truth arrives, yields per-node MAE $e_{t}$ . (Middle) A FIFO window $Q$ of recent errors is compared against a validation-calibrated reference $\mu_{\text{ref}}$ ; drift fires when $\overline{e}^{\text{recent}}\!>\!\alpha\,\mu_{\text{ref}}$ . (Bottom) On a trigger, a priority score $p_{n}\!=\!\tilde{e}_{n}/(1\!+\!\beta u_{n})$ selects the top- $K$ nodes, and a single Adam step updates the shared parameters.

Scenario	MAE			Improvement vs Static		Update Cost
Scenario	Static	Naive	AdaptLSTM	Naive	AdaptLSTM	Naive	AdaptLSTM
Gradual Drift	0.1699	0.1639	0.1678	+3.56%	+1.27%	100%	20%
Sudden Drift	0.1776	0.1641	0.1681	+7.62%	+5.33%	100%	20%
Recurring Drift	0.1737	0.1645	0.1691	+5.31%	+2.66%	100%	20%
Mixed Drift	0.1733	0.1639	0.1679	+5.42%	+3.07%	100%	20%
Average	0.1736	0.1641	0.1682	+5.50%	+3.10%	100%	20%

Table 1. Main results on the Alibaba Machine Trace. AdaptLSTM captures 54% of Naive Online’s absolute improvement using only 20% of computational cost (

2.7\times

efficiency per unit compute).

Figure 3. Pareto frontier on sudden drift. AdaptLSTM sweeps

\alpha\in[0.85,1.30]

, while fixed-period baselines sweep

k\in[3,100]

. Classical detectors (ADWIN, DDM, PH) never trigger and collapse to Static. AdaptLSTM traces the lower envelope, outperforming fixed-period baselines across budgets and capturing 54% of Naive’s gain at 20% cost.

Variant	MAE	Updates	Cost
Full AdaptLSTM	0.1674	42	20.0%
w/o Drift Detection	0.1695	43	100.0%
w/o Selective Update	0.1674	42	100.0%
w/o Adaptive Threshold	0.1722	23	20.0%

Table 2. Ablation study on the sudden-drift scenario.

Figure 4. Threshold factor

\alpha

sweep on sudden drift. Lower

\alpha

produces more triggers and slightly better MAE but at higher effective cost.

\alpha{=}1.0

minimizes MAE while keeping the trigger count near AdaptLSTM’s default operating point. Increasing

\alpha

beyond

1.10

starves the update budget and lets MAE climb back toward Static.

Figure 5. Update ratio

\rho

sweep on sudden drift. Because the gradient is computed on all nodes, MAE stays essentially flat across

\rho\in[0.1,1.0]

;

\rho

acts as a pure reporting knob for effective compute cost. Operators can dial

\rho

down to match a strict infrastructure budget without sacrificing accuracy.

Detector	Gradual		Sudden		Recurring		Mixed
Detector	MAE	Trig.	MAE	Trig.	MAE	Trig.	MAE	Trig.
Static	0.1699	0	0.1776	0	0.1737	0	0.1733	0
ADWIN ( $\delta{=}0.05$ )	0.1699	0	0.1776	0	0.1737	0	0.1733	0
DDM	0.1699	0	0.1776	0	0.1737	0	0.1733	0
Page–Hinkley	0.1700	1	0.1772	1	0.1728	1	0.1730	1
Fixed period ( $k{=}8$ )	0.1665	37	0.1684	37	0.1689	37	0.1675	37

Table 3. Comparison with classical drift detectors across all four drift scenarios. All rows share the same LSTM backbone and single-step Adam update; only the detector varies. Trigger counts are on the 301-step test stream. ADWIN and DDM never fire on any of the four scenarios; PH fires exactly once. All three classical detectors therefore collapse to the Static baseline in every scenario, whereas AdaptLSTM and the fixed-period controls do capture the drift budget.

Scenario	Static	Naive (100%)	AdaptLSTM (20%)
Gradual	$0.16992$	$0.16386\pm 1.5{\times}10^{-5}$	$\mathbf{0.16775\pm 3.0{\times}10^{-5}}$
Sudden	$0.17762$	$0.16412\pm 1.4{\times}10^{-5}$	$\mathbf{0.16803\pm 1.2{\times}10^{-4}}$
Recurring	$0.17373$	$0.16450\pm 1.7{\times}10^{-5}$	$\mathbf{0.16913\pm 3.9{\times}10^{-5}}$
Mixed	$0.17325$	$0.16388\pm 1.2{\times}10^{-5}$	$\mathbf{0.16795\pm 3.4{\times}10^{-5}}$

Table 4. Multi-seed MAE (mean

\pm

std over

N{=}10

seeds) on the Machine Trace. All 10 seeds preserve the ordering Static

>

Adapt

>

Naive per scenario, giving paired Wilcoxon two-sided

p{=}0.00195

for both Static vs Adapt and Naive vs Adapt (one-sided

p{=}0.00098

). Static has

\sigma{=}0

because the frozen model is deterministic w.r.t. the drift stream.

Scenario	Backbone	Static	Naive	Adapt	Trig. / Eff. cost
Gradual	LSTM	0.1772	0.1631	0.1695	18 / 1.20%
	GRU	0.1685	0.1629	0.1671	22 / 1.46%
	Transformer	0.1718	0.1603	0.1679	18 / 1.20%
Sudden	LSTM	0.1866	0.1635	0.1728	22 / 1.46%
	GRU	0.1746	0.1634	0.1698	26 / 1.73%
	Transformer	0.1777	0.1607	0.1690	22 / 1.46%

Table 5. Model-agnostic evaluation across all four drift scenarios. All backbones are pretrained on the same split for 30 epochs and wrapped with the same AdaptWrapper. Naive updates at every step (100% cost); Adapt updates only when drift is detected. Trig. is the number of drift triggers out of

T{=}301

; effective cost

=(\text{Trig.}/T)\cdot\rho

with

\rho{=}0.2

Metric	Static	Naive	AdaptLSTM
Total time (s)	1.16	3.37	2.53
Throughput (samp/s)	258.2	89.0	118.4
Predict latency (ms)	0.30	1.02	3.17
Update latency, avg (ms)	–	9.14	40.78
# Updates	0	300	37
Cumulative update time (s)	0	2.74	1.51

Table 6. Wall-clock efficiency profile on sudden drift (300 samples, NVIDIA A30).

Dataset	Scenario	MAE			Improvement over Static		Cost
Dataset	Scenario	Static	Naive	AdaptLSTM	Naive (100%)	AdaptLSTM (20%)	Adapt/Naive
Machine	Gradual	0.1699	0.1639	0.1677	+3.56%	+1.28%	35.8%
	Sudden	0.1776	0.1641	0.1680	+7.60%	+5.40%	71.1%
	Recurring	0.1737	0.1645	0.1691	+5.32%	+2.65%	49.9%
	Mixed	0.1733	0.1639	0.1680	+5.41%	+3.06%	56.5%
	Average	0.1736	0.1641	0.1682	+5.47%	+3.10%	53.3%

Table 7. Cross-dataset validation (mean over 10 seeds for Machine, 5 seeds for Container). Container exhibits more severe drift, and AdaptLSTM correspondingly captures a larger fraction of Naive’s absolute improvement (avg

96.3\%

53.3\%

on Machine). On the Container Mixed scenario Adapt slightly edges Naive despite using only

20\%

of its update budget.

	Gradual		Sudden		Recurring		Mixed
Detector	MAE	Trig.	MAE	Trig.	MAE	Trig.	MAE	Trig.
Static	0.0329	0	0.1289	0	0.0964	0	0.0953	0
ADWIN ( $\delta{=}0.05$ )	0.0329	0	0.1289	0	0.0964	0	0.0953	0
DDM	0.0329	0	0.1289	0	0.0964	0	0.0953	0
Page–Hinkley	0.0442	2	0.0722	5	0.0964	0	0.0579	4
Fixed period ( $k{=}8$ )	0.0235	37	0.0383	37	0.0314	37	0.0201	37

Table 8. Comparison with classical drift detectors on the Alibaba Container Trace. Same protocol as Table 3 : identical LSTM backbone and single-step update, only detector varies. ADWIN and DDM never trigger, while PH fires 0–5 times. Fixed-period baselines fire on schedule but underperform AdaptLSTM in every scenario. AdaptLSTM values are 5-seed means.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Detector	Parameters
ADWIN	$\delta{=}0.05$ (permissive), window cap 500, min samples 20, values rescaled to $[0,1]$ with saturation at $2\mu_{\text{ref}}$
DDM	$\tau=\mu_{\text{ref}}$ (mistake threshold), drift level $p{+}s>p_{\min}{+}3s_{\min}$
Page–Hinkley	$\lambda{=}2.0$ , $\delta{=}0.005$ , warm-started with reference mean
Fixed-period	$k\in\{5,8\}$ ; $k{=}8$ approximately matches AdaptLSTM’s trigger count, $k{=}5$ approximately matches AdaptLSTM’s effective cost

Appendix

Table 9. Classical detector settings. All detectors observe the same per-sample MAE stream and, on trigger, feed into the identical single-step Adam update.

Parameter	Value
Base LSTM
Hidden dimension	128
Number of layers	2
Dropout	0.2
Optimizer (offline)	Adam
Learning rate (offline)	$10^{-3}$

Appendix

Table 10. Full hyperparameters.

Scenario	Variant	MAE	Updates	Cost
Sudden	Full	0.1674	42	20.0%
	w/o Drift Detection	0.1695	43	100.0%
	w/o Selective	0.1674	42	100.0%
	w/o Adaptive Threshold	0.1722	23	20.0%
Gradual	Full	0.1678	36	20.0%
Gradual	w/o Drift Detection	0.1687	36	100.0%

Appendix

Table 11. Ablation across all four drift scenarios (Alibaba Machine Trace). “w/o Drift Detection” triggers on a fixed period matching Full’s average trigger count; “w/o Selective” sets

\rho{=}1.0

; “w/o Adaptive Threshold” hardcodes

\mu_{\text{ref}}=0.17

$\alpha$	MAE	# Updates	Effective cost
0.85	0.1670	53	3.52%
0.90	0.1688	49	3.26%
0.95	0.1673	41	2.72%
1.00	0.1679	36	2.39%
1.05	0.1686	29	1.93%
1.10	0.1708	24	1.60%

Appendix

Table 12. Threshold factor

\alpha

sweep on sudden drift. # Updates is measured; effective cost

=(\text{Updates}/T)\times\rho

with

T{=}301

\rho{=}0.2

$\rho$	MAE	# Updates	Cost
0.1	0.1675	42	10.0%
0.2	0.1674	42	20.0%
0.3	0.1673	42	30.0%
0.5	0.1673	42	50.0%
1.0	0.1673	42	100.0%

Appendix

Table 13. Update ratio

\rho

sweep on sudden drift. MAE is essentially flat because the gradient uses all nodes;

\rho

only re-scales reported cost.

$w$	MAE	# Updates	Cost
10	0.1681	71	33.8%
20	0.1674	42	20.0%
50	0.1687	18	8.6%
100	0.1712	5	2.4%

Appendix

Table 14. Detection window

w

sweep on sudden drift.

$\eta$	MAE	# Updates	Cost
$5\times 10^{-5}$	0.1691	42	20.0%
$1\times 10^{-4}$	0.1674	42	20.0%
$3\times 10^{-4}$	0.1685	42	20.0%
$5\times 10^{-4}$	0.1732	42	20.0%

Appendix

Table 15. Online learning rate

\eta

sweep on sudden drift.

Figure 6. Drift detection timeline under sudden drift (Machine Trace, steps 100–200). After drift at step 150, MAE rises from

\approx 0.164

0.21

. The detector first fires at step 152, reducing error to

0.173

, followed by three additional triggers for further adaptation. The gray dashed line denotes

\mu_{\text{ref}}=0.1639

. Total effective cost is 0.8%, versus Naive’s 100%.

Policy	Total GFLOPs (24h)	Cost (24h)	Annual cost
Static	3,600	$0.79	$289
Naive Online	14,400	$3.19	$1,165
AdaptLSTM	5,112	$1.13	$412
Savings vs Naive	9,288 (64.5%)	$2.06	$753

Appendix

Table 16. Projected 24-hour compute cost for a 10,000-machine deployment at 1-minute cadence. GFLOPs and cost assume LSTM forward = 2.5 GFLOPs, update = 7.5 GFLOPs; A30 GPU at $2.50/hour, 300 TFLOP/s sustained. AdaptLSTM triggers 14% of steps (from sudden-drift empirical ratio).

Dataset	Static MAE	Naive MAE	AdaptLSTM MAE	Triggers
Machine	0.1652	0.1588	0.1629	35 / 301 (11.6%)
Container	0.0312	0.0139	0.0312	0 / 301 (0%)

Appendix

Table 17. Natural drift evaluation on held-out test splits. Machine test split exhibits mild temporal non-stationarity; Container test split is approximately stationary. Detector conservatism prevents over-triggering in stable regimes.

Explore similar work

When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting
Sep 1, 2026Takumi Fujimoto, Hiroaki NishiMultivariate Time Series Forecasting Learning Rate Scheduling
School of Science for Open and Environmental Systems, Graduate School of Science and Technology, Keio University, Yokohama, Japan
SWIFT: Spatio-temporal Wavelet Integrated Forecasting Framework for Workload Traces
May 7, 2026Zeyuan Ding, Lingfeng Zheng, Dian Ding +1Multivariate Time Series Forecasting Time Series Forecasting
1Shanghai Jiao Tong University, Shanghai, China · 2Shaanxi Normal University, Xi’an, China
ADAPT: A Self-Calibrating Proactive Autoscaler for Container Orchestration
May 15, 2026Himanshu Singh BaghelAdaptive Control
Department of Computer Engineering [J.C. Bose University of Science and Technology] Faridabad, India

Sep 1, 2026, cs.LGJ/K move · Enter open · S save

When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting

Takumi Fujimoto, Hiroaki Nishi

School of Science for Open and Environmental Systems, Graduate School of Science and Technology, Keio University, Yokohama, Japan

Multivariate Time Series Forecasting Learning Rate Scheduling Continual Test-Time Adaptation Edge Computing Temporal Distribution Shift

Open paper AskPDF arXiv

Online adaptation can help edge time-series forecasting under distribution drift, but its measured benefit is sensitive to evaluation choices. We study six public multivariate streams, including building-sensor and smart-meter data, under a leakage-free streaming protocol. We identify two additional sources of comparison bias. First, the warmup budget of the static baseline has a two-sided effect: insufficient warmup undertrains the baseline, whereas excessive warmup can degrade its pre-drift generalization. Across six dataset-backbone settings, the estimated adaptation benefit changes by 3.0 to 18.8 percentage points (pp) over the 1,000-20,000-step warmup range. Second, comparing SGD with momentum (SGD+m) and Adam at a shared default learning rate conflates optimizer quality with rate sensitivity. We select both the warmup budget and each optimizer's online rate using a held-out pre-drift validation slice without accessing test data. Under this validation-only procedure, Adam outperforms SGD+m in 310 of 360 evaluated cells, while 4 Adam cells remain below the static baseline. We further characterize accuracy against adaptation-state memory and A100-measured per-update latency for full, head-only, and calibration-based adaptation. In the evaluated PatchTST frontier settings, several parameter-efficient variants are nondominated on the adaptation-state-memory axis. Smart-meter analyses also show that reported gains depend on meter-selection rules. These findings support a validation-only commissioning procedure, while target-device latency and energy remain to be measured. Code, data, and all reported numbers: https://github.com/keiotakmin/tsf-edge-adaptation.

Abstract

Explore similar work

When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting

SWIFT: Spatio-temporal Wavelet Integrated Forecasting Framework for Workload Traces

ADAPT: A Self-Calibrating Proactive Autoscaler for Container Orchestration