TL;DR
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Pragmatic Engineer published an interview with Maude Lemaire about scaling infrastructure at Slack and GitHub, based on her experience at both companies. The discussion covers Slack’s response to enterprise load, a lightweight tool for testing hundreds of thousands of WebSocket connections, GitHub’s database-sharding constraints, and Lemaire’s approach to verifying AI-generated work.
The Pragmatic Engineer has published an interview with Maude Lemaire on scaling Slack and GitHub, detailing how the services addressed growing infrastructure demands and what limits emerged in their systems. Lemaire, a principal engineer who spent nearly eight years at Slack before joining GitHub in 2024, discusses enterprise load, performance testing and database architecture in a conversation relevant to teams operating large software products.
At Slack, Lemaire led the backend performance infrastructure team. The interview describes efforts to prepare for a prospective customer with 500,000 users. A test using a separate headless Chrome instance for each simulated user was estimated to cost about $150,000 for a five-hour run. Instead, Lemaire built Koi Pond, a tool that creates slimmed-down Slack clients in Go to establish WebSocket connections. The account presents this as a less costly way to test a large number of connections; it does not provide a benchmark comparing the tool’s results with production traffic.
The discussion also explains two kinds of load problems Slack had to manage. Custom emojis posted in a large workspace can generate substantial fan-out traffic because each item must reach users. The source says Slack capped custom emojis at 23 to limit this effect. Separately, an event flood can overwhelm an individual client when a user belongs to many busy channels and receives a surge of messages and related activity at once.
At GitHub, Lemaire worked on database-scaling challenges after joining in 2024, as Copilot was growing. The report describes how GitHub’s Rails monolith and its use of Active Record could obscure the database work triggered by application features. It gives the example of home-page reloads issuing roughly 100 SQL queries across three database clusters. Many queries without a sharding key had to run across shards, which the report says constrained GitHub’s ability to expand beyond 16 database shards.
Managing Load Without Hiding Its Cost
The examples show why growth in users does not translate into a single, simple capacity problem. A service must handle traffic distributed across many connections, while also preventing concentrated bursts from overwhelming a particular client or service component. Slack’s Koi Pond approach, as described in the interview, illustrates how testing methods can be adapted when a direct simulation is too expensive.
The GitHub account points to a different operational risk: an abstraction that makes application development easier may leave engineers less aware of the queries and infrastructure costs behind a feature. The reported 16-shard constraint connects query design to the company’s ability to scale databases. For engineering teams, the practical lesson is not that a particular framework or abstraction is inherently unsuitable, but that its system-level effects need measurement and attention.
Popular right now
Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer – Almond Cream (KSM150PS)
Breville BES870XL Barista Express Espresso Machine
As an affiliate, we earn on qualifying purchases.
Lemaire’s Work Across Three Companies
Lemaire’s experience spans several performance settings. The source says she spent nearly eight years at Slack leading backend performance infrastructure, then joined GitHub in 2024 to address database scaling. More recently, she worked on performance at Cursor, including time to first token, a measure of how long users wait for the first output from an AI system.
The interview also recalls an earlier performance project at Rent the Runway. In 2015, its home page could take as long as 10 seconds to load, in part because of a high-resolution hero image. Lemaire said she worked on improving the site even when she was told competition was not a reason to prioritize speed. Across these examples, the report frames performance as a product-quality concern as well as an infrastructure issue.
On AI-assisted engineering, Lemaire says she uses large language models but personally verifies their output, particularly for foundational platform work. She argues that ambitious systems depend on stable foundations. The source does not provide a formal policy or measure of how often she uses AI tools; the account is her stated approach.
“The best competition is competition with yourself, knowing that you can do better, knowing that you can drive a better experience.”
— Maude Lemaire, discussing her Rent the Runway performance work
Limits of the Published Account
The source is an interview summary, not a full technical design document. It does not give the publication date, detailed test results from Koi Pond, or information about how the 500,000-user scenario was ultimately handled. The projected $150,000 cost is an estimate for the proposed headless Chrome test, not a reported bill or a general price for load testing.
The material also does not explain when Slack’s emoji cap was introduced, whether the limit varies by workspace, or what other controls are used to manage event floods. On GitHub, it does not give a timeline for the 16-shard limit or describe the later changes, if any, that addressed scatter queries. Those details would be needed to assess the systems’ current configurations.
The Interview’s Technical Follow-Up
The immediate next step for readers is to consult the full Pragmatic Engineer episode and transcript, which the source says are available through its page and podcast platforms. Those materials may provide fuller explanations of the examples and Lemaire’s discussion of working on GitHub’s Rails monolith and using AI tools.
The supplied report does not announce a new product, policy change or engineering milestone at Slack or GitHub. It is not clear whether either company plans to publish further technical details about the systems discussed. Any claims about current shard counts, emoji limits or load-testing practices would require newer information from the companies or Lemaire.
Key Questions
What is the news?
The Pragmatic Engineer published an interview with Maude Lemaire about infrastructure and performance challenges at Slack and GitHub, drawing on her work at both companies.
What was Slack’s Koi Pond tool designed to do?
Koi Pond creates lightweight Slack-like clients in Go to open WebSocket connections for load testing. Lemaire built it as an alternative to a proposed test using hundreds of thousands of headless Chrome instances.
Why did GitHub’s database sharding become difficult?
The report says many queries lacked a sharding key and therefore had to run across multiple shards. These scatter queries became slower as shards were added, limiting GitHub to 16 shards in the account described.
What does Lemaire say about using AI at work?
She says she uses AI tools but personally checks their output, especially on foundational platform work. The interview presents this as her approach, not as a company-wide policy.
Source: newsletter.pragmaticengineer.com
