TL;DR
Get business pricing on networking and server 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. The discussion covers Slack’s approach to heavy WebSocket traffic and load testing, GitHub’s database and sharding constraints, and Lemaire’s practice of checking AI-generated work on foundational engineering tasks.
The Pragmatic Engineer has published an interview with Maude Lemaire, a principal engineer, about scaling Slack and GitHub and the infrastructure challenges she encountered at both companies. The discussion offers concrete examples of handling large volumes of user activity, testing capacity, and managing database load in widely used software products.
Lemaire spent nearly eight years at Slack, where she led the backend performance infrastructure team, before joining GitHub in 2024 to work on database scaling. The interview also discusses her later performance work at Cursor, including efforts related to time to first token. The source identifies her as the author of Refactoring at Scale: Regaining Control of Your Codebase.
One Slack example concerns a customer preparing to bring 500,000 users onto the service. Lemaire initially considered using headless Chrome instances to simulate Slack sessions, but the report says that approach would have cost about $150,000 for a five-hour test. Instead, she built Koi Pond, a tool that used lightweight Slack-like clients written in Go to establish WebSocket connections. The account describes the tool as an alternative to running a full browser for every simulated user.
The interview also describes two forms of traffic pressure at Slack. Custom emoji reactions can produce a large number of fan-outs across a workspace, while event floods can overwhelm an individual client receiving activity from many busy channels. The source says Slack limits custom emojis to 23 to help avoid the fan-out problem. These examples illustrate distinct pressures: spreading activity across many users and concentrating a surge of events on one connected client.
How Traffic Shapes Slack’s Design
The examples show why scaling a collaboration service involves more than adding servers. Features that look small to an individual user can multiply into substantial network work when activity is distributed across a large workspace. In the emoji example, the report illustrates how 10 reactions across 100,000 users could amount to one million fan-outs; at 100 reactions, it gives a figure of 10 million. The source uses this to explain the engineering rationale for a product limit, not to quantify measured production traffic.
The load-testing account also highlights a practical trade-off: tests need to reproduce the connections a service must support, but a full browser can be an expensive way to create those connections. Koi Pond’s lightweight clients let the team simulate a large number of WebSocket connections with less machinery, according to the interview. For readers building or operating large systems, the story connects capacity planning to both architecture and the cost of testing.
Products Worth Considering
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
EcoFlow Delta 3 Max Portable Power Station – 2048 Wh LiFePO4 Battery, 2400 W Output (3400 W X‑Boost)
Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade
As an affiliate, we earn on qualifying purchases.
From Slack Infrastructure to GitHub Databases
At GitHub, Lemaire encountered a different scaling problem: the interaction between a developer-friendly framework and the database work it generated. The source says GitHub’s Rails application used Active Record, and that loading the homepage could trigger around 100 SQL queries across three database clusters. It describes many of those queries as scatter queries, which lack a sharding key and must be sent across shard instances.
According to the interview, this pattern constrained GitHub’s ability to add database shards: the company could not go beyond 16 shards without worsening scatter-query problems. The example is presented as a lesson in how abstractions can obscure infrastructure costs from application developers. The source does not provide dates for when these specific database constraints applied, so they should be read as an account of the system Lemaire worked on, not necessarily a description of GitHub’s current architecture.
The conversation also reaches beyond infrastructure. Lemaire recalls working on Rent the Runway’s performance in 2015, when its homepage could take up to 10 seconds to load fully, and says she believed performance mattered even without direct competitive pressure. She describes a similar emphasis on judgment in her use of AI: for foundational engineering work, she personally verifies LLM output rather than treating it as authoritative.
“The best competition is competition with yourself, knowing that you can do better, knowing that you can drive a better experience.”
— Maude Lemaire, as quoted in The Pragmatic Engineer interview
What the Interview Leaves Open
The interview is a retrospective account, and the source does not establish whether the Slack emoji limit, Koi Pond, or GitHub’s cited 16-shard constraint remain in place today. It also does not provide current capacity figures, comparative performance measurements, or independent technical documentation for the examples. The figures should be understood in the specific scenarios described, rather than as a complete picture of either company’s present-day systems.
The source says Lemaire joined GitHub in 2024 and later worked at Cursor, but does not give a detailed employment timeline beyond those points. It also does not quantify the effect of her performance work at Cursor or provide a broader assessment of how AI tools affect engineering outcomes. The interview presents her practices and experience; it is not a formal evaluation of those tools or systems.
Where to Hear the Full Discussion
The source directs readers to the full Pragmatic Engineer episode, available on YouTube, Apple, and Spotify, with a transcript and episode timestamps on the report page. Those materials are the next place to check for fuller explanations of Lemaire’s examples and any qualifications not captured in the summary.
No product launch, new infrastructure change, or follow-up milestone is announced in the source material. The immediate development is the publication of the interview; any assessment of whether the systems or limits discussed have since changed would require newer information from Slack, GitHub, or Lemaire.
Key Questions
What is the news about Maude Lemaire?
The Pragmatic Engineer published an interview with Lemaire about scaling Slack and GitHub, drawing on her experience at both companies and later performance work at Cursor.
How did Lemaire approach Slack’s 500,000-user load test?
The report says she built Koi Pond, using lightweight clients written in Go to create Slack-like WebSocket connections. This replaced her initial plan to use headless Chrome instances, which the source estimates would have cost about $150,000 for a five-hour run.
What database scaling issue did she describe at GitHub?
The interview says some Rails and Active Record queries lacked a sharding key and had to run across shards. The source links these scatter queries to a limit of 16 shards in the system Lemaire discussed; it does not say that this is GitHub’s current limit.
What does Lemaire say about using AI for engineering?
The source says Lemaire uses AI in her work but personally verifies its output for foundational tasks. It does not provide a quantitative comparison of AI-assisted and unaided engineering work.
Source: newsletter.pragmaticengineer.com
