MEASURED · RSS
Process memory samples
local baseline57,920–77,496 KiB
E2E resourced59,924–104,236 KiB
RSS is process resident memory, not Go heap. No heap profile was collected; no universal sub-50MB claim is supported.
// audited performance evidence · 2026-09-21
Maximum throughput.
Microscopic footprint.
A measured view of Logstrm’s CGO-free Go Data Plane: up to 9,512.090 accepted EPS at a 10,000 EPS request, with the E2E matrix reaching 2,454.933 EPS at 2,500 EPS.
MEASURED · 15/15 localMEASURED · 9/9 E2ETARGETS ≠ RESULTS
accepted_evidence.jsonCHECKSUMMED
local / requested10,000 EPS
9,512.090 EPS
e2e / requested2,500 EPS
2,454.933 EPS
validator: PASS
raw artifacts: SHA-256 verified
raw artifacts: SHA-256 verified
01 / resource efficiency
Measured footprint. Explicit targets.
The green bars are observed process samples. Dashed bars are product targets or design illustrations only; they are not represented as benchmark results.
ARTIFACT
TARGET / ILLUSTRATIVE
Efficiency envelope
memory target<50 MB
illustrative under-load target~40 MB
cold-start target<1 second
These bars define future measurement criteria. Cold-start timing and the target memory envelope are not present in the accepted artifacts.
CPU · MEASURED
3.9–6.3%
local baseline and burst samples
THREADS · MEASURED
15–18
local process sample range
COLD START · TARGET
<1s
illustrative target; not measured
02 / architecture boundary
Small runtime. Fewer moving parts.
Logstrm is built and benchmarked as a CGO-free Go service. The comparison below describes architectural shape, not a third-party performance test.
Comparison boundary
Cribl, Logstash and JVM-router figures were not collected under the same harness. Their numeric resource and throughput fields are therefore intentionally marked
Cribl, Logstash and JVM-router figures were not collected under the same harness. Their numeric resource and throughput fields are therefore intentionally marked
not measured.| Dimension | Logstrm | Heavy enterprise routers |
|---|---|---|
| Runtime model | Go, CGO-free build | varies by product |
| Memory comparison | RSS measured | not measured |
| CPU under 10k EPS | local samples: 3.9–6.3% | not measured |
| Cold start | target <1s; not measured | not measured |
| External runtime dependency | none implied by CGO-free binary | product-dependent |
03 / event flow
One stream in. Many destinations out.
Logstrm receives telemetry, applies declarative routing and splits matching events to the destinations that need them. Retries, metrics and bounded DLQ behavior stay inside the Data Plane.
01 / LOG SOURCES
Collect
HTTPSyslogAzure Event HubsAWS S3 / SQSGCP GCS / Pub/Sub
02 / LOGSTRM DATA PLANE
Normalize → route → split
pipeline transformsdeclarative routesbatch + retrybounded DLQ
SEARCH / SIEM
ELK / Elasticsearch
Splunk HEC · Sentinel DCR
HTTP / API
HTTP JSON
destination-specific delivery
STORAGE
S3 / GCS
JSONL archive · batch output
DATA SYSTEMS
SQL / queues
SQS · Pub/Sub · SQL
same event pipeline · route by content · deliver to one or many emitters · observe every handoff
Architecture diagram: it shows supported connector paths, not a claim that every source-to-destination combination was included in the published benchmark matrix.
04 / throughput + chaos
Every number has a test boundary.
Local acceptance and end-to-end delivery are separate measurements. The tables preserve that distinction instead of collapsing them into one headline number.
LOCAL DATA PLANE
HTTP acceptance matrix
| Requested | Average achieved | Rows |
|---|---|---|
| 500 EPS | 496.646 | 3/3 |
| 1,000 EPS | 989.985 | 3/3 |
| 2,500 EPS | 2,459.968 | 3/3 |
| 5,000 EPS | 4,847.743 | 3/3 |
| 10,000 EPS | 9,512.090 | 3/3 |
Measures benchmark HTTP ingestor acceptance, not downstream sink delivery.
E2E CHAOS MATRIX
Delivery and recovery
| Requested | Average achieved | Replay / drain |
|---|---|---|
| 500 EPS | 494.976 | 3/3 |
| 1,000 EPS | 989.964 | 3/3 |
| 2,500 EPS | 2,454.933 | 3/3 |
The isolated capacity matrix verified replay and full drain. The historical resourced failure run is separate and retained 29 DLQ lines / 27,979 bytes.
DELIVERY CONTRACT
At-least-once path
Emitter handoff, sink counters and DLQ evidence are reported separately.
FAILURE MODE
Controlled HTTP 503
Retries and persisted failed events were observed during the sink outage.
UNIVERSAL CLAIM
Not asserted
“100% deterministic recovery” is not claimed beyond the verified run directories.
04 / terminal proof
Proof over promises.
A compact view of the recorded process samples and validator outcome. The terminal panel mirrors accepted evidence; it does not invent a 40MB or negligible-CPU run.
benchmark@logstrm:~
$ python3 validate_run.py results/capacity-e2e-isolated-20260921
[checksums] SHA256SUMS ............. PASS
[matrix] validated rows ........ 9/9
[replay] replay_verified ........ true
[drain] drain_verified .......... true
[result] validate_run.py ........ SUCCESS
$ cat resource-samples.txt
local cpu .................. 3.9–6.3%
local rss .................. 57,920–77,496 KiB
e2e rss .................... 59,924–104,236 KiB
profiles ................... not collected05 / methodology boundary
What is measured. What remains open.
Measured:
throughput matrices, resource.csv samples, sink counters, DLQ snapshots, Prometheus snapshots and checksums.
Not measured:
direct Cribl/JVM comparison, cold-start duration, Go heap profiles and 50,000 EPS capacity.
Target:
the <50MB, ~40MB and <1s bars are future acceptance criteria, not current results.
Scope:
local results are acceptance observations; E2E results cover the tested emitter, sink and fixture.
audit trail
Inspect the methodology and raw evidence.
The technical documentation defines the acceptance contract, validator rules, drain semantics and reproducibility commands. Target language is kept separate from measured evidence.