Every team has that one person. Every review waits for them, and when they say "I'll be offline next week", three people quietly panic.

For Graph Hacks (WeMakeDevs x FalkorDB) I built something that finds these people in Kubernetes. It's called Understudy. Here's the story, mostly in memes.

Every team has this moment

The lie we tell ourselves

This is a real part of Kubernetes, pkg/kubelet/apis/config. Nine people know the code. Six people can approve changes. Sounds safe.

Except every single path between those two groups goes through one person.

So I stopped counting people

Understudy counts routes instead: someone who knows the code, connected to someone who can approve it, with nobody shared between routes. That count is the "understudy number". Here it's 1.

FalkorDB does the counting with a built-in algorithm called max-flow, so it's an actual measurement, not a vibe.

Meet contributor-25b5f

Everyone shows up as a pseudonym, so I don't know who this is. I just know they should never take a vacation.

So I rehearsed them leaving

Understudy copies the whole graph, takes them off the copy, and checks what breaks. 41 of 79 parts of Kubernetes get weaker. The real data never changes.

It also suggests the cheapest fix. Here, pairing one person who knows the code with one approver is enough to get a backup.

I added an AI agent. It's humble.

You can ask it questions in plain English. It can only answer from the graph, so when something isn't there, it presses the second button. Every answer also shows the queries it ran.

Then I argued with it

Every answer gets fact-checked against the graph before you see it. Here the agent was right, and the check was also right. I fixed it by making the tools say things in plain words, so the agent can quote them.

FalkorDB and I had one small disagreement

I tested every number against networkx. With mixed capacities, max-flow came out low on 8 of 150 test graphs. With every capacity set to 1, they agreed every time. So everything is 1 now. Problem solved by not having the problem.

OK but does it actually work?

I checked it against real people who left Kubernetes. Only 7 of 39 cases had enough data, and none of them hit a 1. So honestly, I can't prove it yet. The whole result, including what didn't work, is in the app anyway.

Try it

It runs on a free server with a tenth of a CPU. Getting it to start in 30 seconds instead of 2 minutes was its own adventure.

Live app: https://understudy-u2qk.onrender.com

So, be honest. Does your team have a contributor-25b5f? Do they know? Tell me in the comments. Bonus points if it's you.