Reliability program
Announcing the Streaming-Reliability Benchmark Program

A vendor-neutral
reliability signal
the streaming category hasn't had.

Today we're formally opening the Streamwake Streaming-Reliability Benchmark Program — a quarterly, editorially-reviewed percentile breakdown of streaming uptime, mean time to detect, and mean time to remediate. Each report should state its dataset, observation window, methodology, and provenance so on-calls can compare their own SLOs against a published baseline instead of guessing.

Vendor-neutral·Editorially reviewed·One email per quarter

Illustrative data disclaimer

Illustrative data.

The figures on this page — cohort size, uptime percentiles, mean time to detect, and mean time to remediation — are sample methodology numbers used to explain how the report is structured. They are not a record of verified customer results.

Metric definitions

Three measurements, no hidden denominator.

The benchmark separates health, detection, and recovery. Every claim below carries its current evidence status so an illustrative number cannot read like a production result.

Uptime

Percent of session time classified healthy for each stream. A published report may show stream-level p50, p90, and p99 uptime percentiles after its cohort and window are approved.

The figures currently shown on /benchmark are illustrative methodology numbers, not verified customer results.

AlertingTESTED

Detection latency

Elapsed time from the first failing five-second probe to the alert reaching the operator channel. The report should publish a median and a tail percentile such as p95.

The definition is exercised in the controlled illustrative report path; that test coverage does not establish a production cohort result.

Time to remediation

Elapsed time from the operator alert to a typed, verified remediation event, followed by a clear next probe cycle. Recovery is not inferred from infrastructure health alone.

Recovery gates are documented in the sample report, but no approved customer result is published for this program yet.

Run definition

What counts as one benchmark run?

A run is more than a green dashboard interval. It needs a declared window, a known cohort, traceable signals, and enough evidence to explain both detected incidents and unresolved ones.

One benchmark run

PLANNED

A run starts when the declared cohort, probe mix, and observation window begin. It ends when the window closes and each incident has a detection and remediation verdict, or is explicitly retained as unresolved.

Evidence gate

A qualifying report must name its dataset, observation window, probe cadence, signal provenance, and review decision. Sandbox, pre-production, synthetic, and illustrative data stay labeled as such.

Scoring method

Percentiles first.
Weights in the open.

The score is intended to make the underlying method inspectable, not to hide it behind a single green number. Percentiles are published from the declared stream observations; the composite then applies the documented operational weights.

Percentiles and weighted score

How the score is calculated

Stream-level uptime, detection, and remediation observations are summarized at published percentiles. The proposed 0–100 composite weights verified restore-time at ×1.0, detection-time at ×0.75, false-positive cost at ×−0.40, and vendor grading at ×0.00.

Restore-time wins

× 1.0 anchor

Detection-time wins

× 0.75

False-positive cost

× −0.40

Vendor grading

× 0.00

Those weights are the methodology rule for a future approved report, not a current score or promise that an illustrative value reflects production performance.

Evidence status

Every measurement claim carries its status.

The labels distinguish a live method from a tested definition, a limited pilot, future work, and a claim that has no direct evidence. They are evidence labels, not operational health states.

PRODUCTION

Live, supported measurement behavior. No metric on this page carries this label today.

TESTED

Exercised in a repeatable fixture or controlled evaluation with its limits stated.

BETA

Reserved for a limited pilot while feedback and reliability are still being evaluated.

PLANNED

Defined future work that is not yet a published production measurement.

UNSUBSTANTIATED

No direct evidence supports a production claim, so the page does not make one.

Claim status labels: see definitions.

How to subscribe

Get the next snapshot
the day it goes out.

Subscribe and the next quarterly benchmark lands in your inbox the day it publishes. One email a quarter, the percentile breakdown above, and the editor's notes on what changed since this report. Methodology and the per-stream breakdown are available under NDA on request.

No marketing replies, just the report. Easy to leave.

QuarterlyEditorially reviewedOne-click unsubscribe
Subscribe to next quarter's benchmark

Drops Q1 2027. Same channel as the rest of the marketing list — easy to leave.

Read the methodologySee a sample report

Next report lands Q1 2027

Already published

Read the
latest numbers now.

Subscription unlocks the next report, but the most recent quarterly snapshot and the annual editor's round-up are already live. Read the data without waiting for the next publish.