Generalised Rate Control Approach For Stream Processing Applications

Ziren Xiao

REVIEW 4 major objections 5 minor 14 references

Generalised Rate Control Approach For Stream Processing Applications

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims a graph-attention reinforcement learning agent can control source emission rates across multiple stream processing topologies, improving throughput by up to 13.5% and cutting latency by up to 30% compared with default…

desk verdict Plausible idea, but the 'generalised' claim is untested: the same three topologies are used for training and testing, so the paper shows multi-task fitting, not transfer to unseen topologies. read the letter →

arxiv 2506.11710 v1 pith:7LFDG5ZA submitted 2025-06-13 cs.NI

Ziren Xiao This is my paper · ORCID

classification cs.NI

keywords streamprocessingratecontroldeepreinforcementlearninggraphneuralnetworkproximalpolicyoptimizationbackpressuretopologygeneralization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a graph neural network-based deep reinforcement learning controller that decides how fast each stream source is allowed to emit data, throttling the rate before queues overflow. It argues that replacing the usual MLP state encoder with a graph attention network lets one agent control multiple stream-processing topologies without retraining and removes two manual parameters: the historical state size and the metrics collection interval that earlier rate control required. The author reports experiments on three applications where the learned controller outperforms the default back-pressure scheme, raising throughput by up to 13.5% and reducing end-to-end latency by up to 30%. If correct, the result makes proactive rate control more deployable because operators no longer need topology-specific tuning to use it.

What carries the argument

The load-bearing mechanism is the author's GPPO design, a graph-based proximal policy optimisation agent in which a Graph Attention Network (GAT) with 128 hidden units serves as the feature extractor and feeds the actor and critic through max-pooling and mean-pooling, respectively. The state is a DAG over source, operator, and sink components with per-node metrics such as arrival rate, queue capacities, back-pressure time, and sink latency; the action space is ten discrete fractions of the maximum generation rate, and the reward is the min-max normalised throughput. This structure is what lets the agent see the whole topology at once, drop the stored history of states, and keep a small fixed metrics collection interval, because the current state already reflects the system-wide effect of earlier actions.

What would settle it

Train the same agent on the word count and log stream processing topologies only, then test it on the randomly generated topology; if its throughput and latency then fall to the baseline's level or worse, the cross-topology generalisation claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that the data emission rate at the source of a distributed stream processing system can be controlled collaboratively by a single deep reinforcement learning agent whose policy is built on graph neural network feature extraction. System metrics from every component are assembled into a directed acyclic graph whose nodes carry per-component metrics and whose edges follow data flow; a graph attention network encodes the graph, and actor and critic networks select a discretized throttle action that scales the source's generation rate. Because the graph encoder accepts variable-size graphs, the agent can be trained simultaneously on several topologies and then applied to any of them, and the state need not include historical observations, so the training interval can be short and fixed. The paper reports that this all-in-one agent improves throughput by up to 8.14%, 4.29%, and 13.5% over the default back-pressure scheme on the word count, log stream processing, and randomly generated topologies, and reduces end-to-end latency by 30%, 26%, and 13% respectively.

Load-bearing premise

The claim that the agent generalises across topologies rests on the assumption that a graph attention network trained on the three evaluated topology shapes will transfer to unseen shapes; since all three topologies are used in training and testing, this assumption is never tested.

Editorial extensions

If this is right

Operators can deploy one trained controller across several stream applications without per-topology retraining, if the graph encoder truly transfers.
Rate control can be added without changing application source code, since the agent only communicates through a metrics reporter and a rate-adjustment socket.
Avoiding overloading proactively reduces the oscillatory behaviour of back-pressure mechanisms, yielding more stable throughput and latency.
The min-max normalised reward makes multi-topology training feasible, allowing simple topologies to accelerate learning on more complex ones.
Removing manual historical size and collection interval lowers the deployment barrier for rate control in real stream processing systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the decisive test the paper does not run is training on two topologies and testing on the third; until that is done, 'generalised' should be read as a design goal rather than a demonstrated property.
Editorial inference: because the parameter study appears to rely on a single training run, the reported convergence differences may reflect stochastic variation; repeating with several seeds would separate tuning effects from luck.
Editorial inference: a natural stress test is to change queue capacities, link latencies, or operator processing times after deployment and check whether the same checkpoint still tracks the new back-pressure threshold.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper has a nice idea: use a GNN-based PPO agent to control source emission rates in distributed stream processing, and train one agent across multiple topologies at once. The motivation is clear and practical: the earlier MLP-based controller [4] required manual tuning of metrics interval and history size, and retraining per topology. The all-in-one training strategy is a genuine extension, and the reported gains over the default back-pressure baseline (up to 13.5% throughput, 30% latency) are plausible given the mechanism. The setup is described in enough detail that the simulation could be reproduced.

The main soft spot is exactly what the stress-test note flags: Section V.C.1 claims the agent 'should be able to learn new topologies or patterns that had not yet been seen', but all three topologies used in training also serve as test topologies. That is multi-task fitting, not inductive generalization. This gap directly undermines the word 'generalised' in the title and in the central claim. Without at least one held-out topology, the paper does not support its headline contribution.

Other issues are real but secondary. There are no error bars or multiple seeds, so the reported percentages could shift with a different random draw. There is no direct comparison to the prior MLP-based rate controller [4], which is a missed opportunity: it would actually test whether the GNN representation helps. The 'parameter-free' claim also sits awkwardly next to the interval parameter study in Fig. 4, which shows the WCT agent is sensitive to the metrics-collection interval. And with no code or data, the results are not independently checkable.

None of these flaws disprove the core mechanism. The approach is reasonable and the results are directionally consistent. But the evidence as presented supports 'a multi-task rate controller for three specific topologies', not a 'generalised' one. This is a solid workshop paper or a modest conference paper after revision.

I would send it to serious peer review, but the decision should be conditional: the authors need to add a held-out topology test, seed statistics, and a comparison to [4]. If those are added, it becomes a useful incremental contribution. I would not cite it in its current form, but I would read the revision if it appears.

Referee Report

4 major / 5 minor

Summary. The paper proposes a graph-neural-network-based deep reinforcement learning rate controller for distributed stream processing systems. The approach models a stream topology as a DAG whose node features are system metrics, uses a GAT-based feature extractor inside a PPO agent (GPPO), and selects a source emission rate as a discretized fraction of the maximum generation rate. The reward is a min-max normalized throughput. The authors claim that, unlike an earlier MLP-based controller, the GNN agent avoids state stacking and manual tuning of the metrics collection interval, and can be trained jointly on multiple topologies. Experiments in an OMNeT++ simulation compare the all-in-one agent against a default back-pressure-only scheme on three topologies (WCT, LSPT, RGT) and report throughput improvements and latency reductions, plus a parameter study of the collection interval.

Significance. If the central claims were fully supported, the paper would be a useful step toward reusable DRL-based rate control for stream processing: the GNN representation is a natural fit for variable-size topologies, and the pluggable environment design is practical. The paper deserves credit for formulating the problem as an MDP on a DAG, for building on the author's prior GPPO method, and for reporting an all-in-one training scheme. However, the experimental evidence as presented supports only per-topology fitting, not the paper's headline 'generalised' transfer claim. A held-out topology experiment, direct comparison with the prior MLP-based controller, and multiple-seed statistics are needed to justify the main conclusions. The numerical results may be reproducible, but their significance is currently limited by the experimental protocol.

major comments (4)

[§V.C.1 and Fig. 5] The central generalization claim is not tested. The all-in-one agent is trained on WCT, LSPT, and RGT simultaneously, and then evaluated on exactly those same three topologies. This demonstrates multi-task fitting, not the ability to 'learn new topologies or patterns that had not yet been seen in the previous training' as stated in §V.C.1. To support the 'generalised' label, the authors should include a held-out topology setting, for example training on two topologies and testing on the third, or testing on a new DAG shape never used in training, and report whether the agent transfers without retraining.
[§V.C.2 and Fig. 4] The claim that the method is 'parameter-free' is contradicted by the metrics collection interval K. Figure 4 shows that the choice of K affects convergence speed and training time, and the text recommends a smaller interval (1s). This means K remains a manually selected parameter. Other parameters, such as the action discretization set and the PPO hyperparameters, are also manually fixed. The authors should either soften the 'parameter-free' claim or provide a sensitivity analysis showing that the method's final performance is robust to K and to reasonable variations of the other settings.
[§V.C.3 and Fig. 5] The claimed improvements over the baseline rely on single runs without error bars or seed statistics. The paper itself attributes the slower convergence of the 5s interval in Fig. 4 to 'unlucky' neural network initialization, which acknowledges that PPO results are seed-dependent. Without multiple independent runs and reported mean/standard deviation or confidence intervals, the specific numbers 'up to 8.14%, 4.29% and 13.5%' and the latency reductions of 30%, 26%, and 13% cannot be assessed for statistical significance. The authors should rerun each experiment with several seeds and report variability.
[§V and §II] The paper's motivation is built on the shortcomings of the prior MLP-based rate controller [4], yet no experimental comparison with [4] is provided. The abstract and introduction claim that GPPO removes the need for historical state stacking and manual interval tuning compared with [4], but the experiments only compare with the default back-pressure scheme. A direct comparison on the same topologies, reporting throughput, latency, training time, and sensitivity to K and historical size, is necessary to support the claimed advantages over the earlier method.

minor comments (5)

[§IV.A.b] The phrase 'discredited action space' should be 'discretized action space'.
[§III] The throughput formula 'thr = PK skp / K' is rendered unclearly; please define the summation indexing and the tuple-count variable explicitly.
[§V.C.2] The phrase 'converge trending' should be 'convergence trend', and the use of 'unlucky' is informal for a journal report; consider describing this as seed-dependent initialization.
[§V.C.3] 'the baseline unstablise the system' should be 'the baseline destabilizes the system'.
[Fig. 4] The interval parameter study shows only one curve per setting; adding multiple seeds would make the convergence comparison more informative, consistent with the major comment on statistical robustness.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: reward/evaluation alignment is the intended RL objective; the untested unseen-topology claim is an evidence gap, not a circular step.

full rationale

The paper's claimed improvements (up to 13.5% throughput, 30% latency) are evaluated against a fixed external baseline — the default back-pressure scheme — on the same three topologies used for training. This is a standard RL evaluation of a trained controller, not a prediction from an independently fitted parameter: the reward is the min-max normalized throughput (Sec. IV.A.c), so using throughput as the evaluation metric is the intended objective, not a disguised restatement of the input. The 'generalised' claim in Sec. V.C.1 that the agent 'should be able to learn new topologies or patterns that had not yet been seen' is not demonstrated by any held-out topology experiment, and all three topologies are used for both training and testing (Fig. 3, Fig. 5); however, that is an empirical-support gap, not a circularity, because no equation or parameter is defined in terms of the claimed result. The use of GPPO from the author's prior work [5] is a normal citation to an independently published method; the variable-size-graph property of GAT is an externally established property, and the paper adds its own all-in-one training evidence. The interval parameter study (Fig. 4) qualifies the 'parameter-free' claim as a consistency concern, but it is a hyperparameter choice, not a fitted input renamed as a prediction. No load-bearing step reduces, by the paper's own equations or by definition, to its own inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the graph representation is a reformulation of existing topology data and GPPO is adopted from prior work [5]. The free parameters are mostly environment and training choices that the paper claims to eliminate but still relies on.

free parameters (6)

Metrics collection interval K = 1s recommended; 5s and 10s explored in Fig. 4
The interval for aggregating system metrics is manually chosen. The paper claims to be parameter-free, but this parameter still needs to be set and is studied in Section V.C.2.
Action space discretization = {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0}
The discrete set of rate multiplier actions is hand-designed in Section IV.A.b and controls the granularity of throttling.
Min-max reward normalization bounds = thrmin and thrmax observed so far during training
The reward in Section IV.A.c normalizes throughput using the best and worst values observed in the agent's own history, making the reward non-stationary during training.
PPO hyperparameters = lambda=0.99, lambda_gae=0.95, clip=0.2, entropy_coef=0, value_coef=0.5, learning_rate=0.0003
These are borrowed from stable-baselines3 defaults [13] and are not tuned, but they are still part of the method and affect convergence.
Back-pressure check interval and delay = 0.1ms check interval, 0.05ms message delay
Simulation settings copied from Apache Storm documentation in Section V.A.2; they directly affect the system dynamics the agent experiences.
Rate fluctuation multiplier = uniform random in [0.7, 1.3] every 100 time steps
Chosen in Section V.A.1 to simulate variability in the source rate; it shapes the training distribution and is not derived from measured data.

assumptions (5)

domain assumption Apache Storm back-pressure semantics are faithfully captured in the OMNeT++ simulation.
Section V.A.2 describes a simulation based on Storm's back pressure, but no validation against a real cluster is provided.
ad hoc to paper Graph states retain enough information to act as a Markov state without stacking history.
Section IV.A.a claims historical metrics can be omitted because the graph state 'collaboratively reflects' the overall system status; this is asserted, not proven.
ad hoc to paper GNN inductive node and edge representations generalize to unseen topology shapes.
Section V.C.1 claims adaptability to new topologies, but no held-out topology is evaluated in the experiments.
domain assumption Profiled input rates and processing latencies represent real workloads.
Section V.B says WCT and LSPT parameter settings are profiled from real Apache Storm implementations, but no data or code confirms this.
standard math The MDP formulation and PPO update rules are standard and applicable to this environment.
Section IV.A formulates the problem as an MDP and Section IV.C uses PPO updates from stable-baselines3 [13], following established RL practice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Generalised Rate Control Approach For Stream Processing Applications." pith.science (2026). https://pith.science/paper/7LFDG5ZA

@misc{pith2026250611710,
  author       = {Pith},
  title        = {Pith review of: Generalised Rate Control Approach For Stream Processing Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7LFDG5ZA}},
  note         = {Machine review of arXiv:2506.11710}
}

read the original abstract

Distributed stream processing systems are widely deployed to process real-time data generated by various devices, such as sensors and software systems. A key challenge in the system is overloading, which leads to an unstable system status and consumes additional system resources. In this paper, we use a graph neural network-based deep reinforcement learning to collaboratively control the data emission rate at which the data is generated in the stream source to proactively avoid overloading scenarios. Instead of using a traditional multi-layer perceptron-styled network to control the rate, the graph neural network is used to process system metrics collected from the stream processing engine. Consequently, the learning agent (i) avoids storing past states where previous actions may affect the current state, (ii) is without waiting a long interval until the current action has been fully effective and reflected in the system's specific metrics, and more importantly, (iii) is able to adapt multiple stream applications in multiple scenarios. We deploy the rate control approach on three applications, and the experimental results demonstrate that the throughput and end-to-end latency are improved by up to 13.5% and 30%, respectively.

Figures

Figures reproduced from arXiv: 2506.11710 by the authors.

**Figure 2.** Graph representations of the evaluated distributed [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** All-in-one model training performance. performance of our approach under multi-level applications. Both parameter settings from WCT and LSPT, such as the processing latency and the maximum data generation rate, are profiled by real topology implementations in Apache Storm. C. Experiment Results 1) Multiple Topology Adaptability: The main advantage of using graph neural networks is the adaptability of multiple topol… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Parameter study of the metrics collection interval in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Testing performance three topologies in a single DRL agent. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

[4]

Intelligent data source emission rate control for optimising the per- formance of streaming applications,

Z. Xiao, L. Christopher, and R. Maria, “Intelligent data source emission rate control for optimising the per- formance of streaming applications,” in International Symposium on Cluster , Cloud and Internet Computing (CCGrid). IEEE/ACM, 2024, pp. 266–272

work page 2024
[1]

Real-time calculus for scheduling hard real-time systems,

L. Thiele, S. Chakraborty, and M. Naedele, “Real-time calculus for scheduling hard real-time systems,” in 2000 IEEE international symposium on circuits and systems (ISCAS), vol. 4. IEEE, 2000, pp. 101–104

work page 2000
[2]

Optimal operator placement for distributed stream pro- cessing applications,

V . Cardellini, V . Grassi, F. Lo Presti, and M. Nardelli, “Optimal operator placement for distributed stream pro- cessing applications,” in Proceedings of the 10th ACM International Conference on Distributed and Event-based Systems, 2016, pp. 69–80

work page 2016
[3]

Efficient operator placement for distributed data stream processing applications,

M. Nardelli, V . Cardellini, V . Grassi, and F. L. Presti, “Efficient operator placement for distributed data stream processing applications,” IEEE Transactions on Parallel and Distributed Systems , vol. 30, no. 8, pp. 1753–1767, 2019

work page 2019
[5]

Macns: A generic graph neural network integrated deep reinforce- ment learning based multi-agent collaborative navigation system for dynamic trajectory planning,

Z. Xiao, P. Li, C. Liu, H. Gao, and X. Wang, “Macns: A generic graph neural network integrated deep reinforce- ment learning based multi-agent collaborative navigation system for dynamic trajectory planning,” Information Fusion, vol. 105, p. 102250, 2024

work page 2024
[6]

Storm@ twitter,

A. Toshniwal, S. Taneja, A. Shukla, K. Ramasamy, J. M. Patel, S. Kulkarni, J. Jackson, K. Gade, M. Fu, J. Donham et al. , “Storm@ twitter,” in Proceedings of the 2014 ACM SIGMOD international conference on Management of data , 2014, pp. 147–156

work page 2014
[7]

Placeto: Learning Generalizable Device Placement Algorithms for Distributed Machine Learning

R. Addanki, S. B. Venkatakrishnan, S. Gupta, H. Mao, and M. Alizadeh, “Placeto: Learning generalizable device placement algorithms for distributed machine learning,” arXiv preprint arXiv:1906.08879 , 2019

work page Pith review arXiv 1906
[8]

GDP: Generalized Device Placement for Dataflow Graphs

Y . Zhou, S. Roy, A. Abdolrashidi, D. Wong, P. C. Ma, Q. Xu, M. Zhong, H. Liu, A. Goldie, A. Mirhoseini et al. , “Gdp: Generalized device placement for dataflow graphs,” arXiv preprint arXiv:1910.01578 , 2019

work page Pith review arXiv 1910

Show all 14 references

[9]

M. R. Garey and D. S. Johnson, Computers and in- tractability. freeman San Francisco, 1979, vol. 174

1979
[10]

Tata: Throughput-aware task placement in heterogeneous stream processing with deep reinforcement learning,

X. Huang, Y . Jiang, H. Fan, H. Tang, Y . Wang, J. Jin, H. Wan, and X. Zhao, “Tata: Throughput-aware task placement in heterogeneous stream processing with deep reinforcement learning,” in 2021 IEEE Intl Conf on Parallel & Distributed Processing with Applications, Big Data & C...

2021
[11]

Readys: A reinforcement learning based strategy for heterogeneous dynamic scheduling,

N. Grinsztajn, O. Beaumont, E. Jeannot, and P. Preux, “Readys: A reinforcement learning based strategy for heterogeneous dynamic scheduling,” in 2021 IEEE Inter- national Conference on Cluster Computing (CLUSTER) . IEEE, 2021, pp. 70–81

2021
[12]

[Online]

(2024) Parameter settings of ppo. [Online]. Avail- able: https://stable-baselines3.readthedocs.io/en/master/ modules/ppo.html#parameters

2024
[13]

Stable-baselines3: Reliable reinforcement learning implementations,

A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-baselines3: Reliable reinforcement learning implementations,” Journal of Machine Learning Research , vol. 22, no. 268, pp. 1–8, 2021. [Online]. Available: http://jmlr.org/papers/v22/20-1364.html

2021
[14]

[Online]

(2023) Performance tuning in apache storm 2.1.1. [Online]. Available: https://storm.apache.org/releases/2. 1.1/Performance.html

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [4]

Intelligent data source emission rate control for optimising the per- formance of streaming applications,

Z. Xiao, L. Christopher, and R. Maria, “Intelligent data source emission rate control for optimising the per- formance of streaming applications,” in International Symposium on Cluster , Cloud and Internet Computing (CCGrid). IEEE/ACM, 2024, pp. 266–272

work page 2024

[2] [1]

Real-time calculus for scheduling hard real-time systems,

L. Thiele, S. Chakraborty, and M. Naedele, “Real-time calculus for scheduling hard real-time systems,” in 2000 IEEE international symposium on circuits and systems (ISCAS), vol. 4. IEEE, 2000, pp. 101–104

work page 2000

[3] [2]

Optimal operator placement for distributed stream pro- cessing applications,

V . Cardellini, V . Grassi, F. Lo Presti, and M. Nardelli, “Optimal operator placement for distributed stream pro- cessing applications,” in Proceedings of the 10th ACM International Conference on Distributed and Event-based Systems, 2016, pp. 69–80

work page 2016

[4] [3]

Efficient operator placement for distributed data stream processing applications,

M. Nardelli, V . Cardellini, V . Grassi, and F. L. Presti, “Efficient operator placement for distributed data stream processing applications,” IEEE Transactions on Parallel and Distributed Systems , vol. 30, no. 8, pp. 1753–1767, 2019

work page 2019

[5] [5]

Macns: A generic graph neural network integrated deep reinforce- ment learning based multi-agent collaborative navigation system for dynamic trajectory planning,

Z. Xiao, P. Li, C. Liu, H. Gao, and X. Wang, “Macns: A generic graph neural network integrated deep reinforce- ment learning based multi-agent collaborative navigation system for dynamic trajectory planning,” Information Fusion, vol. 105, p. 102250, 2024

work page 2024

[6] [6]

Storm@ twitter,

A. Toshniwal, S. Taneja, A. Shukla, K. Ramasamy, J. M. Patel, S. Kulkarni, J. Jackson, K. Gade, M. Fu, J. Donham et al. , “Storm@ twitter,” in Proceedings of the 2014 ACM SIGMOD international conference on Management of data , 2014, pp. 147–156

work page 2014

[7] [7]

Placeto: Learning Generalizable Device Placement Algorithms for Distributed Machine Learning

R. Addanki, S. B. Venkatakrishnan, S. Gupta, H. Mao, and M. Alizadeh, “Placeto: Learning generalizable device placement algorithms for distributed machine learning,” arXiv preprint arXiv:1906.08879 , 2019

work page Pith review arXiv 1906

[8] [8]

GDP: Generalized Device Placement for Dataflow Graphs

Y . Zhou, S. Roy, A. Abdolrashidi, D. Wong, P. C. Ma, Q. Xu, M. Zhong, H. Liu, A. Goldie, A. Mirhoseini et al. , “Gdp: Generalized device placement for dataflow graphs,” arXiv preprint arXiv:1910.01578 , 2019

work page Pith review arXiv 1910