---
title: "The engineering floor: billions of impressions"
url: https://tanghoong.com/work/adtech-scale-marketplace/
locale: en
description: "Eleven years in AdTech: a platform conversion, a self-serve advertising marketplace built end to end, and backends carrying tens of thousands of requests a second."
---
# The engineering floor: billions of impressions, and a marketplace that ran itself

> Eleven years across two stints at the same AdTech company: four as the sole front-end engineer on a platform conversion, then a self-serve advertising marketplace built end to end, then backend services carrying tens of thousands of requests a second.

- **Role:** Software Engineer, then Senior Software Engineer
- **Capabilities:** Scale · Platform · Hands-on engineering
- **Domain:** Advertising technology — publisher, sales and campaign operations
- **Span:** Two stints, eleven years

## Context

An advertising technology company running publisher, sales and campaign-operations platforms. I joined in 2011, left in 2017 to start my own company, and was asked back twice over the following two years; I returned in 2019 at a higher grade and stayed through the pandemic.

## Problem and ambiguity

Three different problems over eleven years, but one recurring shape: something that worked at one scale did not work at the next, and the fix was never in the part everyone was looking at.

## My role

Years one to four: the sole front-end engineer on a legacy-to-new conversion of the platform’s operator surfaces, plus QA. Year five: I asked my manager to move me to the back end and got it — I was not reassigned, I requested it. Then I built the company’s B2B self-serve advertising marketplace, reporting directly to the CTO. In the second stint I led backend and API improvements and the migration onto a new API surface.

## Discovery

The operator surfaces were not a design exercise, they were a usability problem with a commercial edge: a sales pipeline that carries opportunities through confidence stages to a won deal and on into a campaign lifecycle, a campaign manager over the country, advertiser, campaign, flight, placement and creative hierarchy, and a publisher portal covering ad-slot configuration, earnings accrual and payout requests. Getting those right meant sitting with ad operations and sales rather than with a specification.

## Architecture and trade-offs

The scale work was a migration of a legacy architecture onto a new API surface, and both halves came down to taking work out of the request path. Ad telemetry — every impression, click, and which element inside an ad was clicked — would have flooded the servers if each event fired its own call, so events were accumulated and shipped on a fixed interval to a separate bookkeeping service, processed outside the main system, and written back. Reporting had been generated synchronously, so a page blocked while its report ran; that moved onto worker batches, with the page loading immediately and the job reporting back when done.

## Implementation and integration

The marketplace was a complete commercial surface: advertising bundled into packages and deals, a deliberately simple membership layer over a complex core, wired to the company’s API engine, with customer self-ordering and self-management, creative upload, prepaid points top-up and deduction on the buyer side, and earnings accrual and payout requests on the publisher side. Publishers could list packages, set their placements and sell.

## Production controls

Report submission was made idempotent — refreshing the page could not re-trigger a job that was already running, which is the failure every user finds within a week. Multi-layer caching, database query tuning and Node.js refactoring on the hot paths.

## Impact

Backend services processing tens of thousands of requests per second and billions of monthly ad impressions, with selected p95 response times brought into the sub-100 millisecond range.

## Reusable learning

Both of the scale wins came from the same move — get the work out of the request path — and neither required a rewrite. The throughput number is a consequence of that decision, not an achievement in its own right, and I would rather be asked about the mechanism than the number.

## Evidence pending

This decade is best carried by a slide deck rather than a repository — project P8 — because the work is too old and too client-owned to rebuild honestly.