Benchmarks
Explore the latest retained regular benchmark recordings by the operation you care about. Each section identifies the measured workload and units; source links and downloadable JSON let you check the observations behind a chart.
Image performance
Boot readiness, fork and prepared-image size.
Three recorded trials per image profile. See the downloadable records for their observed ranges.
Data table
| Category / date | Mean (ms) |
|---|---|
| lifecycle/data_0.7.2_hosted.json | 484.2 |
Local network performance
VM-to-VM private network latency and throughput.
The retained regular bundle does not contain eligible measurements for this section. Existing diagnostic or exploratory recordings are not substituted.
External network path
Controlled HTTP fixtures exercise the request path; these are not Internet speed tests.
50,000 successful requests; 0 failures. Fixture source: local.
Data table
| Category / date | Latency (ms) |
|---|---|
| p50 | 13.6 |
| p95 | 21 |
| p99 | 576.2 |
Hypervisor I/O
Sequential and random reads and writes inside the guest.
Data table
| Category / date | Guest disk (MiB/s) |
|---|---|
| Read | 4,808.5 |
| Write | 230.1 |
Hypervisor CPU and memory
Loaded CPU, memory and VM-count scaling require matching recorded resource samples.
The retained regular bundle does not contain eligible measurements for this section. Existing diagnostic or exploratory recordings are not substituted.
Logging and query performance
Service reads during concurrent mutation; distinct from event ingestion and flush throughput.
160 samples. Read/API latency; not an ingestion or flush benchmark.
Data table
| Category / date | Service UDS (ms) |
|---|---|
| p50 | 0.131 |
| p95 | 0.214 |
| p99 | 0.242 |
Security inspection
Model and credential response paths; distinct from isolated plugin, reconstruction or rule evaluation cost.
End-to-end observed paths, not isolated reconstruction, plugin or rule cost. Compatible payloads and concurrency are required for comparison.
Data table
| Category / date | Host direct (requests/s) | Through Capsem (requests/s) |
|---|---|---|
| model json response | 2,879.8 | 3,470.7 |
| credential response | 2,075.7 | 2,329.1 |
Reading these results
These charts show existing measurements from one recorded source, not historical performance trends. Select a measurement to keep units and workloads separate, and download the recorded JSON for sample counts and observed ranges.
Provenance, units and coverage
Source: hosted release benchmark artifact, commit f2b2a6d9b28d. The snapshot identifies x86_64; it does not record a complete CPU-model inventory. Comparisons with other machines or architectures are not implied.
The legacy throughput_mbps field is produced from bytes divided by 1024² and seconds, so it is labeled MiB/s here. Format-version fields in individual JSON files are not the Capsem release version.
Image timings have three recorded trials per profile. Model and credential response scenarios have matching recorded request counts and concurrency. Their end-to-end times do not isolate plugin execution, reconstruction or rule evaluation, and differences between paths do not establish a statistically significant overhead claim.
Logging currently shows service read latency during profile mutation. It is not an ingestion or flush-throughput result. Local Criterion outputs without cohort identity and groups marked exploratory or failed doctor fitness remain outside the published charts.
Docs CI imports newer eligible hosted recordings when available. Reviewed source JSON is retained so artifact expiry does not erase these measurements. The importer never starts a VM or runs a benchmark.
For code growth, test counts and dependency relationships, explore Statistics.