Case Study: Scaling a Next.js Site to 10K+ Daily Users with Minimal Ops
In eight weeks, we took a content-and-commerce Next.js app from 800 to 12,600 daily users without adding a single full-time SRE. This deep dive shows the decisions, tradeoffs, and measurements that mattered, framed for leaders who care about speed, stability, and cost.
Baseline and constraints
We inherited a monorepo with Next.js 13, a Vercel deployment, Postgres for core data, and a headless CMS. Builds took 19 minutes, p95 TTFB was 1.2s, and cache hit rates hovered at 43%. Traffic spiked during influencer campaigns, creating cold starts, blocked event loops, and confused dashboards.
- Targets: 10K+ daily users, 99.95% SLO, p95 TTFB sub-400ms on cached paths, sub-800ms uncached.
- Ops budget: near-zero. Prefer serverless primitives, managed queues, and strict observability.
- Marketing need: predictable Core Web Vitals for SEO and paid media landing pages.
Architecture decisions that moved the needle
- Adopted Incremental Static Regeneration and per-path revalidation. 78% of pages shifted to ISR with 60s stale-while-revalidate, cutting origin reads by 71%.
- Split rendering: Edge runtime for read-heavy routes; Node runtime for write and checkout. This avoided slow crypto and image libs at the edge.
- Streaming React Server Components for product grids. Largest Contentful Paint dropped 28% with skeletons and partial hydration budgets.
- Introduced a token-bucket rate limiter in middleware to apply backpressure during campaign bursts, keeping p99 under 1.2s.
- Queued background revalidation via a lightweight queue. Rebuild hot pages after stock or price changes without hammering the database.
- Image optimization: moved originals to R2, served via Vercel Image Optimization with deterministic widths and WebP/AVIF. CDN egress fell 34%.
- Database: Read replicas for analytics-heavy queries; added prepared statements and page-level caching for faceted search.
Performance audits for web apps: our playbook
- Instrument first. RUM for Core Web Vitals, tracing for server components, logs with high-cardinality fields like tenant, feature flag, and cache status.
- Define error budgets and SLOs per route group. Landing pages received 70% of the budget; admin got the rest.
- Create a dependable test matrix: 3 device classes, 2 networks, and 2 locales. Automate via synthetic checks gating deploys.
- Audit hydration cost. We removed a carousel and 19kB of client JS from the homepage; conversions rose 4.2% as CLS stabilized.
- Measure build and deploy friction. We split the repo, enabled Turborepo remote caching, and cut builds to 6 minutes; canaries every 15 minutes.
- Set cache observability. Dashboards show hit/miss by path, TTL, and revalidate triggers; marketing sees real cache heatmaps.
AI agent development in the runtime loop
We introduced small, focused agents rather than a monolith. Each agent owned a narrow decision: routing anomalies, schema drift detection, and content freshness. Agents consume traces, logs, and business events, propose actions with confidence scores, and write to an approvals topic.
- Incident triage agent suggests rollbacks when p95 degrades beyond SLO and correlates with a feature flag. Median time to detect fell from 14 to 3 minutes.
- SEO content freshness agent predicts which pages to revalidate ahead of spikes. Cache hit rate for campaign landers climbed to 86% during launches.
- Schema sentinel flags breaking API diffs and generates migration PRs with test fixtures.
Team model: Gigster managed teams, plus partners
Delivery ran through Gigster managed teams for predictable velocity and governance. We paired that with specialized partners: slashdev.io provided remote engineers for burst capacity on frontend polish and data pipelines, while our in-house staff owned domain rules and approvals. Clear swimlanes minimized meetings and allowed weekly ship cadences.
Results and what to steal
After week eight, the site handled 12,600 daily users with headroom. Cached route p95 TTFB averaged 280ms; uncached 650ms. Core Web Vitals passed at 93% of sessions, and infra cost per 1K visits fell 41%.
- Bias to ISR and smart revalidation before scaling databases.
- Guardrails beat heroics: SLOs, error budgets, and canary gates keep launches boring.
- Push intelligence to the edge: rate limits, cache hints, and lightweight agents.
- Design for observability from day one; dashboards your marketers actually read.
- Keep ops thin by choosing managed primitives and automating the noisy parts.
Final note: schedule quarterly performance audits for web apps, revisit cache TTLs, and rotate load tests against real traffic shapes; calibrations compound, keeping spend predictable and your roadmap focused on growth.
If you're planning a push to 10K+ daily users, start with a performance audit, identify the three slowest routes, and turn them into ISR candidates. Add a queue for revalidation, set a clear SLO, and wire a small AI agent to watch your trace stream. Then keep shipping.
