EventFlowSentry
A reproducible fault-injection and metamorphic-testing approach for finding event-time failures in streaming pipelines.
01
Problem
Event-time systems can appear correct in ordinary tests while producing different outcomes under late data, reordering, replay, watermark movement, and recovery. The hard part is not generating random failures; it is producing faults whose expected invariants and evidence remain reproducible.
Constraints that shaped the design
- Keep experiments repeatable enough for independent inspection.
- Exercise timing and recovery behavior without coupling the method to one production stack.
- Separate observed behavior from claims that require broader empirical validation.
02
Architecture
- 01
Controlled fault model
Represent event-time perturbations as explicit, repeatable experimental inputs.
- 02
Metamorphic oracle
Compare related executions through invariants when a single expected output is insufficient.
- 03
Evidence trail
Retain the execution conditions and resulting observations for analysis and reproduction.
03
Design decisions
- Treat reproducibility as part of the test design, not an after-the-fact reporting step.
- Use metamorphic relations to test semantics that are difficult to express as fixed expected outputs.
- Describe the artifact conservatively as a preprint until peer review and external replication change its status.
04
Alternatives considered
- Purely random chaos tests without replayable fault definitions.
- Happy-path examples that do not vary event-time or recovery conditions.
- Presenting repository activity or manuscript submission as proof of independent impact.
05
Testing and results
- The methodology, motivation, and reported evaluation are documented in the public SSRN manuscript.
- The DOI provides a stable external identifier for the current preprint version.
- Claims on this page are intentionally bounded by the public artifact; external adoption is not asserted.
Outcome
The work turns timing faults into an inspectable systems-testing problem and provides a public research artifact that readers can evaluate directly.
06
Limitations
- The current public evidence is a preprint, not an accepted publication.
- Results should not be generalized beyond the evaluated systems and fault models without replication.
07
Lessons learned
- A fault is useful only when the resulting behavior can be explained and replayed.
- Event-time correctness needs relationships between executions, not only example-by-example assertions.