Architecture and performance
Where the pixel runs, what latency it adds, how it scales. No promises — verifiable numbers.
Anyone can promise "complete, real-time measurement at scale." What follows is what it takes to back that sentence with numbers. This chapter describes where the SealMetrics pixel runs, what latency it introduces, how it scales, and what we guarantee.
It's written for technical leads — CTO, head of data, SRE — who need to validate before they install. If you find an outdated number, write to us: we publish them so they can be corrected.
Where the pixel runs
All event processing runs in owned infrastructure in Dublin, Ireland — European Union, no traffic transiting the United States, no sub-processors outside the EEA. This includes pixel ingestion, validation, attribution, and storage.
The reason is operational before it is legal: a single region means a single surface to audit and maintain. The legal reason — Schrems II — follows from it.
Latency and client footprint
The pixel weighs under 4 KB gzipped and loads asynchronously — it doesn't block document render, doesn't compete for main thread with critical scripts, and doesn't consume cookies or browser storage.
"Measuring more data cannot come at the cost of the user. If measurement degrades experience, we have broken the most important part of the product."
Figures measured in production over the last quarter. Next revision: August 2026.
Capacity and scaling
The ingestion backend is a horizontally scalable Go service backed by ClickHouse. Each node processes events in batches with disk confirmation before returning a response — no silent event loss during traffic spikes.
Today we process traffic peaks for customers handling billions of events per month. The architecture is sized to absorb 5x without structural changes.
Availability and SLAs
Scale+ plans include a 99.9% SLA on ingestion and dashboard availability. Twelve-month history is published at status.sealmetrics.com.
What we don't measure
By architecture, there are things the pixel cannot capture — not in this chapter, not in any other. We do not identify individuals, we do not reconstruct cross-device sessions, and we do not fingerprint. That boundary is covered in chapter 09.
Was this chapter useful? Tell us →