Example: Technical Talk

A conference-style talk demonstrating the code layout, Mermaid diagrams, annotated code blocks, and the split and full-bleed layouts. Uses the forge theme for a technical aesthetic.

Copy the source below into a new .md file and open it in Kova.

---

Source

---
title: "Zero-downtime deploys at scale"
author: Marcus Lee
date: 2026
theme: forge
aspect_ratio: "16:9"
---

# Zero-downtime deploys at scale

How we ship 40× a day without waking anyone up.

Marcus Lee · SRE Summit 2026

---

## Agenda

- The old way and its costs
- Blue-green at scale
- Our canary pipeline
- Observability hooks
- What we got wrong first

---

## The old way

---

## Deploy night

- Feature freeze two weeks before release
- All-hands on a Friday evening
- 45-minute maintenance window
- Rollback plan: pray

> It worked in staging.
> — Every engineer, ever

???

Moment of recognition from the audience — most rooms have been there.
Hold for a beat before moving on.

---

## The cost

!progress[Engineering time in deploy coordination](38)
!progress[Incidents traced to deploy window](61)
!progress[Mean time to recover from bad deploy](79)

38% of our SRE time · 61% of P0 incidents · 4.5h avg MTTR

---

## The new approach

---

## Architecture overview

```mermaid
graph TD
    A[Feature branch] --> B[CI — tests + lint]
    B --> C[Build artefact]
    C --> D{Canary gate}
    D -->|Pass| E[Blue-green swap]
    D -->|Fail| F[Automatic rollback]
    E --> G[Traffic shifted 100%]
    F --> H[Alert + post-mortem]
```

???

Walk through left to right. The canary gate is the key innovation —
the audience will ask about it. Save the detail for the next slide.

---

## The canary gate

```rust
pub struct CanaryGate {
    error_threshold: f64,   // (1)
    latency_p99_ms:  u64,   // (2)
    sample_window:   Duration,
}

impl CanaryGate {
    pub fn should_proceed(&self, metrics: &Metrics) -> Decision {
        if metrics.error_rate > self.error_threshold {
            return Decision::Rollback;         // (3)
        }
        if metrics.latency_p99 > self.latency_p99_ms {
            return Decision::Rollback;
        }
        Decision::Proceed
    }
}
```

???

(1) Defaults to 0.1% error rate
(2) 200ms p99 — tighter than our SLA to give headroom
(3) Rollback is synchronous and completes in < 30s

---

## Traffic shifting

```bash
# Shift 5% to canary
./shift.sh --canary 5

# Wait for gate evaluation (default: 10 min)
./gate.sh --wait

# On pass: shift 100% and retire blue
./shift.sh --canary 100 && ./retire.sh --blue
```

---

## Observability hooks

---

## What we instrument

- **Canary error rate** — compared against rolling 7-day baseline, not absolute threshold
- **p50 / p95 / p99 latency** — per-endpoint, not aggregate
- **DB query duration** — catches schema migration regressions
- **Memory growth rate** — catches slow leaks before they become OOMs

---

## Dashboard

![Deploy health dashboard](./assets/dashboard-deploy.png)

Each deploy gets a unique ID surfaced in traces, logs, and the dashboard.
One click shows every error that occurred during the canary window.

???

We open-sourced the dashboard config — Grafana + Prometheus. Link in the
talk notes.

---

## What we got wrong first

---

## Mistake 1: comparing against a fixed threshold

**Wrong:** roll back if error rate > 0.1%

**Right:** roll back if error rate > baseline × 1.5

Traffic spikes make absolute thresholds fire constantly on perfectly
healthy deploys.

---

## Mistake 2: ignoring cold-start latency

<!-- layout:split -->

![Cold-start spike graph](./assets/cold-start.png)

The first 60 seconds after a swap show artificially high latency due to
JIT warmup and cache priming. We now exclude that window from gate
evaluation.

---

## Mistake 3: insufficient canary traffic

| Traffic to canary | Time to detect P0 bug |
|:-----------------:|:---------------------:|
| 1% | ~40 min |
| 5% | ~8 min |
| 10% | ~4 min |
| 25% | ~90 sec |

We now send **10%** to canary by default — enough signal, not enough
blast radius.

---

## Results

!progress[Deploy frequency (40× / day)](100)
!progress[P0 incidents from deploys](15)
!progress[MTTR improvement](88)
!progress[SRE time freed from coordination](74)

40× / day · P0s from deploys down 85% · MTTR < 30 min · 74% coordination time saved

---

## Resources

- github.com/acme/canary-gate — open-source gate library
- grafana.com/dashboards/acme-deploy — dashboard config
- talk slides: acme.io/sre-summit-2026

Thank you.

???

Take questions. Common ones:
- "What's your rollback time?" — < 30 seconds, fully automated
- "What about DB migrations?" — expand-contract pattern, separate talk
- "How do you handle stateful services?" — good question, we don't (yet)

---

What this demonstrates

Slide	Layout	Feature
Title	title	Theme and frontmatter
"The cost"	bsp	Multiple progress bars
"Architecture overview"	code	Mermaid flowchart
"The canary gate"	code	Rust syntax highlighting
"Traffic shifting"	code	Bash code block
"Dashboard"	title-image	Image-heavy slide
"Mistake 2"	split	Manual <!-- layout:split --> override
"Mistake 3"	title-content	GFM table
"Results"	bsp	Progress bar group
Speaker notes	—	??? throughout