Latest / The CTO Podcast with Fexingo: Technical Leadership, Architecture, and Engineering Org / How Slack Rebuilt Its Platform for Ten Thousand Integrations
Transcript
- Lucas: In 2021, Slack had about 2,400 apps in its marketplace. By 2025, that number had more than tripled to over 8,000. And each one integrated with Slack's API in a different way — from simple bots to deep workflow automations. Luna: That's a massive jump. So how did their architecture keep up without constant outages? Lucas: Good question. Originally, Slack had a monolithic API — essentially one big REST endpoint that handled everything. That worked when you had a few hundred integrations, but as the marketplace grew, they hit serious reliability and consistency issues. Luna: So they needed a complete re-architecture? Lucas: Exactly. They started by moving to a unified GraphQL gateway. That let them centralise authorization, rate limiting, and data fetching. GraphQL gave developers flexibility — they could request exactly the data they needed — and Slack could enforce security at a single point. Luna: But isn't GraphQL harder to cache? I've heard that complaint. Lucas: It is. Slack had to build a sophisticated caching layer on top, using persisted queries and CDN caching for common requests. But the trade-off was worth it because it dramatically improved developer experience — the number of support tickets related to API calls dropped thirty percent in the first year. Luna: Thirty percent — that's a concrete win. What about the event-driven side? I know Slack apps rely heavily on event subscriptions. Lucas: Right, that's where Kafka came in. They built an event delivery system that handles billions of events per month. Each app subscribes to specific event types, and Slack's platform guarantees at least once delivery. But they also had to handle scale — during peak hours, some apps receive millions of events a minute. Luna: So how do they prevent one noisy app from overwhelming the system? Lucas: They implemented per-app rate limiting, but also per-customer rate limiting. If one app is making excessive calls for a particular workspace, that workspace's quota gets hit first — isolating the blast radius. They also introduced a sandboxed runtime for apps that do compute on Slack's side — things like slash commands or interactive components. Luna: A sandboxed runtime — so they run untrusted code? Lucas: Exactly. Apps deployed to Slack's platform run in isolated containers with resource limits. If an app has a memory leak or an infinite loop, it gets killed automatically without affecting anyone else. That was a big architectural shift from the early days when apps ran on their own servers and just called the API. Luna: That must have been a massive migration — convincing thousands of developers to change how they build. Lucas: It was. They did it incrementally over about two years. First, they introduced new SDKs that abstracted away the old API calls. Then they provided migration guides and a testing environment that mirrored the new stack. They also kept backward compatibility for existing apps — the old endpoints still worked, but they encouraged developers to move over. Luna: And did they have any major incidents during the migration? Lucas: A few. One that I remember is when the GraphQL gateway had a misconfiguration that caused a partial outage for about forty minutes. But their incident response was solid — they had feature flags to fall back to the old API, so they could revert quickly. The key was that they never broke the developer experience for long. Luna: That's a good lesson — invest in gradual migration and maintain escape hatches. Lucas: Yeah, and that kind of community support actually mirrors Slack's approach to their own platform — they invested heavily in documentation and examples, which made the migration smoother. It's a reminder that a lot of the best infrastructure — whether it's Slack's platform or this show — depends on the people who find it valuable enough to support. Luna: True. And if you're someone who gets a lot out of these deep dives and wants to help keep the show ad‑free and independent, you can find us at buy me a coffee dot com slash fexingo. Every contribution goes straight into making episodes like this possible. Lucas: Exactly. So with that in mind, let's talk about the next challenge Slack faced after the core platform rebuild — how they handled versioning and breaking changes. Luna: Versioning is always tricky with a platform that has thousands of integrations. Lucas: It is. Slack adopted a 'versioned SDK' approach — they release new SDK versions that bundle API changes, but the underlying endpoints remain versioned via headers. This way, apps can opt into new features on their own timeline. They also maintain a long-term support version for critical integrations. Luna: So you don't force upgrades on anyone. Lucas: Right. The only time they deprecate an endpoint is if there's a security issue or it's been obsolete for over a year. They give six months notice and provide automated migration tools in the SDK. Luna: That sounds like a lot of engineering effort just to keep the platform running. Lucas: It is, but it's the cost of maintaining a platform that hosts thousands of apps. They measure success by the number of monthly active apps — which has grown over thirty percent year over year since the migration. So the investment paid off. Luna: So where are they now? Are there new challenges on the horizon? Lucas: The big one is AI. Slack is seeing a wave of apps that use large language models to summarise messages, generate responses, or automate workflows. These apps have very different traffic patterns — they often make long-running, streaming API calls. So Slack is now working on supporting WebSocket-based streaming in their platform to handle that. Luna: Interesting — so the platform keeps evolving. Lucas: Exactly. The architecture they built is modular enough to support whatever comes next. That's the real takeaway: a well-designed platform gives you the flexibility to embrace new technologies without starting from scratch. Luna: And it's a good reminder for anyone building a platform to think about extensibility from day one.