Research overview - kmosoti/FabricO11y GitHub Wiki
Research overview
FabricO11y's research has concentrated on bounded ingestion, efficient queries, durable delivery and recoverable operation. Measurements below describe particular workloads and revisions. They are not a deployment capacity guarantee.
| Area | Observed result | Engineering consequence |
|---|---|---|
| Sealing memory | The aligned-input/disk-PageStore campaign passed eight workload cells with three pairs each. Peak builder heap was 37.3 MiB for steady256 and 39.5 MiB for bigrows64. Seven cells were 1.7–11.9% slower; bigrows64 was 2.7% faster on the shared host. | Bounded representation movement became the normal writer. Builder heap is only one component of server memory. |
| Selective queries | One tail workload improved from 614 to 28 ms with the Walk approach. Other workloads and pruning structures have different costs. | Storage-owned maps, hierarchical pruning and delayed payload work remain useful; Scan is still the default. |
| Sustained operation | The companion-inclusive R2 soak passed its ten gates, with 546,000 simulated batches and 825 companion batches accounted for exactly. Query p99 was about 216 ms. | There is finite integrated evidence for the tested snapshot; release artifacts still need their own acceptance. |
| Investigations | Fourteen synthetic journal/restart RCA cells recovered 350 exact complete chains and rejected 30 corruptions. | Cross-signal evidence can support investigations. This does not establish automatic causal diagnosis or real application integration. |
| Reference systems | Comparisons and source dissections produced useful sort, ownership and representation ideas, along with rejected timing and memory candidates. | Keep the demonstrated mechanisms and counterexamples; no alternative storage engine was selected by these trials. |
Read the full records before comparing numbers:
- Bounded writer campaign
- Default adoption
- Readiness continuation
- Walk queries
- Companion soak
- Synthetic RCA
- Cross-system synthesis
- Coupled performance model review
Heap, RSS and cgroup memory measure different things. Isolated performance gains cannot be multiplied into an end-to-end capacity claim. Failed, interrupted and inconclusive runs stay in the historical reports.
The selected release direction is a bounded development/small-deployment product: a central server with its dedicated Spindle, remote edge Spindles, Debian/Fedora packaging and configurable 100 GB retained telemetry. These are release-plan requirements, not a declaration that Fedora packaging or the new default already exists. Product promises and the release gate remain versioned with the code.