Telemetry Placement Evaluation for Microservice Failure Drills
Ryan C. Okafor · PhD Candidate, Khoury College of Computer Sciences, Northeastern University · Spring 2026 research log

What this project does

I am comparing my own telemetry-placement heuristic ("flowrank") against a betweenness-centrality baseline, using the CloudLab public profile tcts-fault-localization as the evaluation harness. The profile generates identifiability verdicts (fully / partially / ambiguous / unidentifiable + abstention) for its registered graph families; my heuristic is scored against the same frozen endpoints as the profile's own baseline.

Why this harness

The profile is the first public harness I found that (a) freezes its acceptance thresholds ahead of time and (b) refuses to answer when the configuration is unidentifiable. The abstention behavior is the reason I trust its verdicts at all: an evaluator that says "I don't know" at B=2 does not quietly lie to you at B=16.

Results (v1.0/v1.0.1, Jan–Mar 2026)

Familybetweenness baseline (C3 ratio)flowrank (mine)delta
H1 (mesh)0.310.27better
H2 (geo)0.290.26better
H3 (mesh)0.330.30better
H4 (fan-in)0.410.46worse
H5 (chain)0.360.32better
H6 (tree)0.380.34better
H7 (mesh)0.300.28better
H8 (public overlay)0.340.31better

Takeaway: flowrank beats the baseline on 5 of 8 families and is clearly worse on fan-in — which matches the profile's own documented boundary (leaf-side observations are mandatory there). That failure taught me more than the wins: my heuristic assumes information concentrates near high-flow nodes, which is false in fan-in topologies.

Run manifests

All 42 runs (cloudlab instantiation records, seeds, hashes) are linked from the manifest table below. Two runs were invalid (allocation loss) and are logged, not silently dropped.

BatchWindowProfile versionRunsManifest hash (first 12)
FP-012026-01-19 → 01-30v1.018f3a91c04d7e2
FP-022026-02-09 → 02-27v1.014b844e17a02c9
FP-032026-03-09 → 03-20v1.0.1107d20bf55e8a1

Relationship disclosure

I found the profile by browsing the CloudLab public profile collection in January 2026. I have no relationship with the TCTS project team, their PI, or Weber State University; nothing was provided to me for this use. Contact: okafor.r@northeastern.edu.