CharlieKey takeaways from your favorite podcasts
← 5 November 2026The Pragmatic Engineer

Scaling Slack and GitHub with Maude Lemaire

Maude Lemaire explains how Slack and GitHub scaled under strain, from thundering herds, load testing and scatter queries to why platform and product engineers should rotate and why AI still needs human verification.

Watch on YouTube ↗

Charlie's summary

  1. Migrate risky data layers with a dumb abstraction first, then feature flags, double writes and shadow reads

    At Slack, channel membership lived in three tables joined by a union to fill the sidebar. Maude first copy-pasted about 400 raw SQL call sites into a minimal class, optimizing nothing. She then used many feature flags, double writes, backfills and shadow reads. She still wiped the sidebars of all of Slack for about 25 minutes when she deleted data while the flag was on.

    ▶ 18:27
  2. Load test around the usage patterns that actually break you, such as thundering herds

    Maude says to first pin down where your product will fall over. At Slack, a CEO's “you all get Friday off” message to 100,000 people triggers a burst of custom emoji reactions. Each reaction fans out over the websocket to everyone, so the load explodes. Deciding what to drop, such as emojis before text, is a product call, not just an engineering one. Slack capped distinct custom emoji reactions at 23 per message.

    ▶ 27:09
  3. Event floods punish power users: the heaviest Slack users got the worst experience

    Unlike a thundering herd, an event flood overwhelms one client. Jeff, an IBM admin, was in about 30 workspaces. An enterprise-level message shared across them arrived 30 times with no deduplication, which made his client die constantly. Maude notes these users were also some of Slack's biggest champions.

    ▶ 32:18
  4. Replace headless Chrome with lightweight Go clients that simulate only the critical path

    To test 500,000 websocket connections, Slack's headless Chrome approach cost 37 cents per instance for a 5-hour test, around $150,000, and the pods kept crashing. Maude's team switched to small Go processes that open a websocket and replay API calls, including edge cache requests. She built a recorder tool to capture flows, then pushed PMs to record and add their own use cases to the load tests.

    ▶ 38:29

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.