Relay v4.2 ships Adaptive Canary — variance-aware rollouts that tune their own thresholds.
v4.2 — Adaptive Canary is live

Ship to production without holding your breath.

Relay is the progressive-delivery control plane. Promote every change through Build, Test, Canary and Production with adaptive analysis and one-second rollback.

31,000+
deploys / day
1.1s
median rollback
0.04%
change-fail rate
relay · deploy #4821 — main@a91f2c4
Build
Test
Canary
Promote
Production
p99
41ms
errors
0.01%
canary
0.98
rollback
armed
one-second revert standing by
healthy

Trusted by platform teams shipping at scale

Northwind
Hyperplane
Cobalt
Atlas Foundry
Meridian
Quanta
Lumen
Drift
The deploy pipeline

Five stages. One safe path to production.

Every change rides the same rails — built, gated, canaried, promoted and guarded. Relay deals them out one at a time, and never advances a stage that isn't healthy.

01Artifact

Build

Every commit becomes an immutable, signed artifact. Relay fingerprints the build, captures provenance, and stages it the instant CI goes green.

  • Signed provenance (SLSA-3)
  • Reproducible artifacts
  • Auto-stage on green
relay deploy --watch
  • [build]fetch source @ a91f2c4
  • [build]compile 1,284 modules
  • [ok]artifact signed · slsa-3
  • [test]612 gates running
  • [ok]gates green 612/612
  • [canary]route 5% → candidate
02Gate

Test

Quality gates run in parallel and report back as a single verdict. A failed gate never reaches a user — the artifact simply waits.

  • Parallel gate execution
  • Single roll-up verdict
  • No flaky-test promotions
quality gates · parallel
  • Unit
    248 passed
  • Types
    1 passed
  • Lint
    1 passed
  • Integration
    96 passed
  • End-to-end
    34 passed
  • Security scan
    1 passed
verdict: 612/612 green
03Analysis

Canary

Relay routes a sliver of live traffic to the new version and watches latency, errors and saturation against the incumbent. Adaptive thresholds tune themselves to each service.

  • Variance-aware scoring
  • Live traffic shadowing
  • Auto-abort on regression
adaptive canary · live
baseline candidate
score 0.98
p99
41ms
errors
0.01%
cpu
38%
04Rollout

Promote

A healthy canary promotes itself on a curve you control — 5%, 25%, 50%, 100% — pausing automatically the moment a signal drifts out of band.

  • Progressive traffic shift
  • Policy-gated promotion
  • Pause on drift
progressive rollout
traffic to candidate100%
5%
25%
50%
100%
auto-pause armedpromoted in 4m 12s
05Steady state

Production

Once at 100%, Relay keeps watching. Post-deploy guardrails hold for the full bake window, and a single keystroke — or a single anomaly — reverts in about a second.

  • Post-deploy guardrails
  • One-second revert
  • Always-on health
production · steady state
live · all green
uptime
0.00%
req / s
0.0k
p99
0ms
errors
0.00%
post-deploy guardrailsholding 5m

Below: zoom into the stage that does the watching — Adaptive Canary.

The 2am problem

Deploys shouldn't require a hero.

Most teams ship by gathering the bravest engineer, watching dashboards by hand, and hoping the rollback runbook still works. It doesn't scale, it burns people out, and it makes Friday a no-deploy day.

01

Manual watching

Someone babysits Grafana for twenty minutes after every release, correlating spikes by eye.

02

Slow, scary rollbacks

By the time a human notices, redeploys the old version and waits for it to roll out, the incident is already minutes old.

03

Blast radius

A bad change hits 100% of traffic at once because there's no safe way to test it on a fraction of users.

04

Deploy anxiety

Fear of breaking prod slows everyone down — releases pile up, batch sizes grow, and risk compounds.

Adaptive Canary

It scores the release so you don't have to.

Relay shadows a fraction of real traffic onto the new version and compares it to the incumbent across latency, error rate and saturation. Thresholds aren't static — they adapt to each service's own variance, so a noisy endpoint isn't held to the same bar as a quiet one.

  • Compares candidate vs. baseline on live, mirrored traffic
  • Variance-aware scoring tuned per service, automatically
  • Aborts and reverts before a regression reaches the majority
  • Every decision is explained, logged and replayable

Median time from regression detected to full revert: 1.1 seconds.

adaptive canary · live
baseline candidate
score 0.98
p99
41ms
errors
0.01%
cpu
38%
0.00%
Fleet uptime across customers
trailing 90 days
0.0s
Median automated rollback
detect → revert
0K
Deploys orchestrated daily
and climbing
0%
Drop in change-failure rate
after 30 days
How it works

Live in an afternoon. Watching forever.

Relay sits beside the tools you already run. There's nothing to rip out — point it at your pipeline and describe a safe rollout.

1

Connect your pipeline

Point Relay at your existing CI and cluster. No rip-and-replace — it sits beside what you already run.

2

Describe a safe rollout

Declare promotion steps, bake windows and guardrail metrics in a few lines of policy. Version it like code.

3

Ship and let it watch

Merge. Relay builds, gates, canaries, promotes and guards — paging a human only when a decision genuinely needs one.

Integrations

Sits beside everything you already run.

Relay consumes your artifacts and signals from the tools your team already uses. No rip-and-replace, no agent sprawl.

GitHub
Source
GitLab
Source
Kubernetes
Runtime
Argo CD
Delivery
Datadog
Signals
Prometheus
Signals
AWS
Cloud
Vercel
Edge
PagerDuty
On-call
Slack
Notify
OpenTelemetry
Traces
Terraform
Infra
Trust & control

Production-grade by construction.

Relay runs in your account, touches the minimum, and proves what it did. Every promotion and revert is signed, attributed and exportable.

SOC 2 Type II

Independently audited controls, renewed annually.

Runs in your VPC

The control plane never holds your data or your secrets.

Signed provenance

SLSA-3 artifacts with a verifiable chain of custody.

Scoped access

Least-privilege by default, with full RBAC and SSO.

Immutable audit log

Every decision is recorded, attributed and exportable.

Private by design

No telemetry leaves your boundary without consent.

Relay vs. hand-rolled

The difference is who's awake at 2am.

Capability
Hand-rolled scripts
Relay
Rollback
Manual redeploy, minutes
Automatic, ~1 second
Canary analysis
Eyeballing dashboards
Variance-aware scoring
Blast radius
100% of traffic at once
Progressive 5 → 100%
Promotion gates
Tribal knowledge
Versioned policy
Audit trail
Scattered logs
Signed, immutable
On-call load
Every deploy is a watch
Paged only when needed
Loved by on-call teams

The graphs watch themselves now.

We went from a no-deploy Friday culture to shipping forty times a day. Relay reverts faster than a human can read the alert.

PN
Priya Nair
VP Engineering, Hyperplane
Pricing

Start free. Scale when production does.

Every plan ships canary analysis and automatic rollback. Pay only when you outgrow a single service.

Solo

$0/ forever

For a single service and a small team finding their footing.

  • 1 service, 1 cluster
  • Canary + automatic rollback
  • 7-day audit retention
  • Community support

Team

Most popular
$890/ month

For engineering orgs running real, frequent production traffic.

  • Unlimited services
  • Adaptive Canary scoring
  • Policy-as-code promotions
  • 90-day audit retention
  • SSO + RBAC
  • Priority support

Enterprise

Custom

For regulated fleets that need it in their own account.

  • Runs in your VPC
  • SOC 2 + SLSA-3 provenance
  • Custom guardrail metrics
  • Immutable, exportable audit
  • Dedicated solutions engineer
  • 99.99% SLA
Changelog

Shipped, naturally, one stage at a time.

Relay ships itself through Relay. Here's the recent cascade of releases.

v4.2LatestJun 2026

Adaptive Canary

Canary thresholds now tune to each service's own variance. Noisy endpoints stop crying wolf; quiet ones get held to a tighter bar.

v4.1StableApr 2026

Policy-as-code v2

Promotion policies gained reusable fragments, dry-run previews and a typed schema with editor completion.

v4.0StableFeb 2026

One-second revert

Reworked the data plane so a revert is a traffic-table flip, not a redeploy. Median detect-to-revert is now 1.1s.

v3.6StableNov 2025

Run-in-your-VPC

The control plane can now run entirely inside your account, holding none of your data or secrets.

FAQ

Questions, answered straight.

No. Relay sits beside your existing CI and cluster. It consumes your build artifacts and orchestrates the rollout — you keep the tools you already use.

Relay keeps the previous version warm and shifts traffic with a routing-table flip rather than a redeploy. Reverting is a state change, not a rebuild, so it lands in about a second.

Latency, error rate and saturation out of the box, plus any custom metric you expose through Prometheus, Datadog or OpenTelemetry. Scoring is variance-aware per service.

On Team it's hosted; on Enterprise the control plane runs inside your own VPC and never holds your data or secrets. Either way, every action is signed and logged.

Kubernetes today, with first-class support for Argo CD, AWS and Vercel edge. Anything that can split traffic by weight can be a Relay target.

Yes — Solo covers one service with canary and automatic rollback, free forever. Upgrade to Team when you outgrow a single service.

Ship with confidence

Make your next deploy a non-event.

Join the teams who let Relay watch production so their engineers can sleep. Request access and we'll have you shipping safely this week.

Runs in your VPC · SOC 2 Type II · 99.99% SLA