Layer 3 · Go deeper · 10 min
Go deeper
Size it, build it, then make the calls an interviewer will push you on.
Back of the envelope
Size it up
Tap through each step of the arithmetic.
How long does saving a 70B model stall 1,024 GPUs?
How often does a 1,024-GPU job get interrupted?
How often should that job checkpoint?
Build-up · 5 steps
Build the high-level design
One functional requirement at a time, ending with a design that works before any scaling.
Step 1 of 5
All GPUs at once
The Job API stores the spec and queues it for its team. The scheduler admits a job only when every GPU fits, then binds all its pods at once, because 127 of 128 pods is the same as zero.
6 interview moments
You’re the candidate
The interviewer pushes on the hard parts. Pick an approach and see how it holds up.
Deep dive 1 of 6
Interviewer
“A 1,024-GPU job needs 128 pods running together, in four whole blocks. How do you admit gangs like it?”
Your move. Pick an approach:
Deep dive 2 of 6
Interviewer
“Every training step ends in an all-reduce, so it runs at the speed of its slowest link. Where do you place a job?”
Your move. Pick an approach:
Deep dive 4 of 6
Interviewer
“fraud-ml reclaims 256 GPUs from a borrower that checkpoints every 70 minutes. How do you take them back?”
Your move. Pick an approach:
Deep dive 5 of 6
Interviewer
“Your 1,024-GPU run is interrupted about every 49.5 hours, mostly by hardware. How do you keep it going?”
Your move. Pick an approach:
Deep dive 6 of 6
Interviewer
“About 300 notebooks run a cell for ten seconds, then wait minutes for their human. How do 40 teams share GPUs for them?”
Your move. Pick an approach:
Build-up · 8 steps
Follow one fine-tune from submit to bill
The control plane decides; the nodes run; checkpoints and the usage ledger remember. Follow one 32-GPU fine-tuning job through it.
Step 1 of 8
Submit
A fraud-ml engineer submits a 32-GPU job from the CLI. The Job API checks identity, team, flavour and image; the spec says what the job needs, never which machines.