Slack - the case study

Boards

The nine boards on one canvas, and the story below runs through them in the order Slack hit each problem.

Before you start. Slack is the chat app companies use instead of email, and it was built on one idea: a company signs up, gets its own private space, and everything about that company lives together in one place. This lesson is the story of what happened when Slack started selling to companies far too big for that idea.

The idea in 30 seconds

Slack assumed one company equals one space, and every hard problem since has been prying that assumption apart while live.
The boot payload grew with headcount, so compression could not help. The answer was a cache of objects at the edge.
A thundering herd is a synchronised arrival, and it is the shape behind the old message servers and both famous outages.
Hashing channels across a ring deleted the per-customer blast radius. A channel became an address, not a room.
A shard key is a frozen product assumption, and changing team_id took three years even with the right tool.
A queue that needs free memory in order to drain cannot recover from being full. Put a log in front of it.
Enterprise features pull load-bearing columns out of your schema, which is how privacy stopped living in the channel id.

Read the story · Skip to the numbers