When p95 moves, the commit that shipped 24 seconds earlier is already next to it.
brew install signalhq/tap/sgl8.4 MBfree for solo devs · no card · self-host the collector
$ sgl watch --env prod --since 15m prod · us-east-1, eu-west-1 · 3 services ok api-gateway p95 184ms err 0.02% ▁▂▁▁▂▁ ok checkout p95 221ms err 0.05% ▁▁▂▁▁▁ !! payments-worker p95 912ms err 1.31% ▁▂▄▆██ ▲ regression payments-worker · p95 +611ms baseline 301ms (24h rolling) onset 14:02:11Z · 11m ago deploy 8f21c4e "retry budget for webhooks" @nadia · 24s before onset → sgl trace payments-worker --deploy 8f21c4e
No proprietary agent: anything that speaks OTLP works the day you install it, and the rest pipes through sgl ingest.
What you get
Signal is a CLI first: everything you can click, you can also review in a pull request.
Every release lands on the timeline with its sha, author and diff. When a regression starts inside a deploy window, Signal attributes it before you open a dashboard.
# runs after your existing deploy step
- name: Mark deploy in Signal
uses: signalhq/mark-deploy@v2
with:
service: payments-worker
env: prod
sha: ${{ github.sha }}
by: ${{ github.actor }}
fail-on-regression: trueOne import at the top of your entrypoint. The SDK wraps HTTP, gRPC and your database driver, and adds nothing to the request path but a buffer write.
import { init } from "@signal/node";
init({
service: "payments-worker",
env: process.env.SIGNAL_ENV,
// all errors, half the slow ones, 1% of the rest
sample: { errors: 1, slow: 0.5, baseline: 0.01 },
endpoint: process.env.SIGNAL_ENDPOINT,
});Thresholds are code, reviewed in pull requests, versioned with the service they watch. No click-ops, no drift between staging and prod.
import { alert, p95 } from "@signal/config";
export default [
alert({
when: p95("checkout.submit").above("400ms"),
for: "5m",
notify: ["#oncall-payments", "escalate:tier1"],
}),
];The same query language in the CLI, the dashboard and the API. Pipe it to jq, commit it as a saved view, or page it into an incident doc.
$ sgl query 'p95(latency) by route' \
--service checkout --since 24h
route p95 n 24h
/checkout/submit 412ms 18.2k +38%
/checkout/quote 121ms 44.9k -2%
/checkout/receipt 88ms 12.1k +1%
/checkout/webhook 118ms 6.4k -1%
/checkout/refund 204ms 2.1k +4%“Time to attribution went from forty minutes to under three.”
Head sampling throws away exactly the requests you needed. Signal decides after the fact: every error, every p99 outlier, a slice of the boring remainder.
sgl config sampling
The collector is a single static binary: no queue, no cluster to babysit. On the self-hosted plan raw traces never leave your VPC.
sgl collector verify
Priced per ingested gigabyte after sampling, not per host and not per seat. Estimate it against your current volume before you commit.
sgl cost --estimateestimator
Changelog
216 releases since 2024, each written by the person who shipped it.
One repository can now emit a deploy event per service, so a forty-package monorepo stops blaming every regression on the same commit.
Processes that never restart drifted above their configured baseline rate, up to 4x on workers older than a week. Thanks to @tvoss for the reproduction.
An alert now arrives carrying the three slowest traces that triggered it, so the notification is the first step of the investigation rather than a reason to go find one.
Install the CLI, point one service at the collector, and the next deploy annotates itself.
sgl init --service payments-worker