The database systems research community has a reproducibility problem, and it may be hiding in plain sight. A growing body of work uses eBPF, the Linux mechanism for running verified programs at kernel hook points, to claim dramatic performance gains for database systems. However, according to a published peer-reviewed paper, many of these claims omit the details needed to verify whether the results would hold outside the original lab.
Sheriff Adepoju, a large-scale automation engineer whose work spans kernel-level systems and infrastructure at scale, is the author of this challenge. His paper introduces MR-1, a minimum reporting standard with a quantitative compliance score, designed to make database–kernel performance claims interpretable, comparable, and, critically,independently verifiable by other researchers and practitioners.
This work arrives at a moment when the gap between benchmark headlines and real-world deployment outcomes has become a recurring frustration for both researchers attempting to build on prior work and engineering teams evaluating architectural decisions based on published evidence.
The Problem: Invisible Variables That Decide Real-World Outcomes
Adepoju's central argument is precise: when papers omit kernel versions, hardware topology, eBPF attachment points, export paths, and measurement controls, readers cannot separate the contribution of a proposed technique from the influence of unreported system variables. In practice, this means that a result reported as "X% faster" may be impossible to replicate or even interpret outside the specific environment in which it was produced.
This is not an accusation of bad faith. Adepoju's critique targets the experimental condition and not the intentions of researchers. His paper demonstrates that under-specified conditions can make headline numbers unreliable guides for both follow-on research and enterprise adoption, where infrastructure teams routinely rely on academic and industry publications when evaluating database architectures, observability stacks, and performance trade-offs.
A Taxonomy That Prevents Category Errors
Before proposing his standard, Adepoju establishes a framework that is lacking in the field: a structured taxonomy that separates three fundamentally different modes of database–kernel integration via eBPF, each with distinct feasibility constraints, failure modes, and reporting requirements.
Mode 1 - Observability: Using eBPF to derive database-relevant signals from kernel and user-space hooks without altering the kernel policy. The risk here is subtle: the measurement stack itself can distort latency and throughput, and export mechanisms can silently drop events under load, unless the loss behavior is explicitly measured and reported.
Mode 2 - Policy injection: Installing workload-specific cache or networking policies at kernel choke points. Adepoju identified this category as operationally riskier because it modifies shared kernel behavior and can produce externalities under mixed workloads that go unreported.
Mode 3–kernel-resident state: maintains a structured, semantics-bearing state in the kernel context for fast-path decisions that require coordinated updates. Adepoju argues that this mode expands the correctness and security boundaries and, therefore, demands the most rigorous reporting of assumptions and failure modes.
The taxonomy's purpose is not to rank these modes but to prevent a specific analytical error that currently distorts the literature: evaluating a tracing pipeline and a kernel fast path as though they represent the same kind of "optimization," when they carry fundamentally different constraints.
MR-1: What Must Be Disclosed - and What Need Not Be
MR-1 defines the minimum set of experimental disclosures that should accompany any database, kernel integration paper-making performance, or feasibility claims involving eBPF. Notably, MR-1 does not require researchers to release proprietary codes, internal datasets, or customer workloads. It focuses strictly on system variables that determine whether results are transferred across environments:
Kernel and platform: kernel version, relevant configuration flags, CPU model and topology (including NUMA layout), and key device details affecting the measured path.
Program and Attachment Specifics: Exact hook points and attachment types; program build toolchain; map types and resource footprints in which the kernel-resident state is part of the claim.
Export and Backpressure: Data movement out of the kernel context, dropping of events, detection and reporting ofloss or backpressure.
Runtime Controls: CPU pinning and affinity settings; interrupt or networking affinity where relevant; sampling and aggregation windows for observability work.
Workloads and Baselines: Workload configuration, including dataset size, skew, scan behavior, and request mix; baseline settings; warm-up procedures; and steady-state criteria.
Metrics and Variability: Definitions for throughput and tail latency; number of repetitions and variability reporting; and overhead accounting when the eBPF path meaningfully consumes the CPU or memory.
Adepoju argues that these are not administrative details but experimental conditions. Without them, quantitative comparisons between papers become unreliable because the reader cannot distinguish the effects produced by the proposed technique from those produced by unreported kernel behavior, topology, or measurement stack choices.
A Quantitative Compliance Score: Making Omissions Visible
To prevent MR-1 from remaining a qualitative checklist, Adepoju introduced a compliance score ranging from 0 to 10, computed strictly from what a paper explicitly discloses. Points are assigned across the six disclosure categories: kernel/platform, attachment details, state/export behavior, runtime controls, workloads/baselines, and metrics/variability.
The score functions as a disclosure index — not a verdict on scientific quality. A higher score signals that a paper's results are easier for other researchers to interpret, compare, and build upon. A lower score indicates that readers and practitioners should treat quantitative comparisons with caution, because key variables remain unspecified.
This distinction is central to Adepoju's contribution: missing disclosures do not prove that a result is wrong. They make it difficult to validate or implement research findings. The scoring mechanism provides reviewers, conference program committees, and journal editors with a concrete, auditable tool for evaluating the completeness of experimental reporting — a function that extends Adepoju's influence beyond individual research into the evaluation standards that govern the field.
Implications for Industry and Enterprise Adoption
Adepoju connects the reproducibility debate directly to enterprise reality. Database and infrastructure teams regularly evaluate architectures, observability approaches, and performance trade-offs based on published research. When performance evidence is under-specified, organizations can spend months pursuing results that do not transfer to their kernels, fleet configurations, workload profiles, or operational constraints.
The paper further argues that systematic under-reporting distorts the research record over time, creating an environment in which "best results" appear comparable when they were produced under fundamentally incompatible conditions. For practitioners making infrastructure investments based on this evidence, the cost of that distortion is measured in wasted engineering effort and misallocated resources.
Limitations and Anticipated Criticism
Adepoju directly addresses the objections that his proposal is likely to face, enhancing the work’s credibility through transparency.
First, not every laboratory has access to identical hardware, and strict reporting standards could raise the barrier to publishing performance work. Adepoju responded that MR-1 requires explicit reporting, not identical resources, enabling honest comparison rather than gatekeeping.
Second, some researchers and industry authors may resist disclosing configurations that they consider proprietary. MR-1 threads this needle by targeting minimal technical conditions rather than implementation details.
Third, checklists can be gamed. A paper can disclose an exhaustive list of parameters while still making questionable comparisons. Adepoju positions MR-1 as a floor — it prevents invisible variables from hiding in plain sight but does not replace careful evaluation design.
Fourth, the compliance score measures disclosure, not replicability. A fully disclosed paper can still be difficult to reproduce, and an under-disclosed paper can still be correct. The claim is deliberately narrow: disclosure is a prerequisite for independent verification and responsible comparison.
What Comes Next
Adepoju has proposed MR-1 as a minimum standard that journals, conferences, and peer reviewers can adopt for database–kernel integration work making performance claims involving eBPF. He has also outlined a path to reduce the compliance burden on authors: machine-checkable reporting templates and tooling that can automatically collect the kernel and attachment metadata MR-1 requires.
His broader objective is to shift performance discourse in database systems research away from results that read as universally transferable and toward results that clearly state the conditions under which they hold. The goal is to establish a research environment in which practitioners can confidently decide when to adopt, when to test further, and when to set aside a headline number that was never reproducible outside a specific laboratory configuration.