Interview Experience | Backend / Networking | π₯π₯π₯ | Amazon
2026 (Oct-Nov) Software Development Engineer, Software Defined Network Controller, Amazon Leo @ Amazon β Redmond, WA β Virtual β 2 rounds | π Neutral | Hard
Round 1 β Coding
Input is a list of links (u, v, start, end, delay). A link can only send during [start, end), arrives delay later. A packet leaves terminal src at t0; earliest time it reaches a gateway?
Asked if a node can hold the packet; he said assume yes for now. Dijkstra, departure is max(t, start), skip if past end. His case passed.
Part 2: multiple gateways, any one is fine. Return when a gateway pops. One line.
Part 3 is where it got hard: satellites can't buffer, must forward immediately; only ground nodes can wait. I just removed waiting and kept earliest arrival per node. He gave a case: leave the ground early and you hit satellite A before A's next link opens; leave two seconds later and you make it. Earliest isn't best.
Stuck a while. Hint: "when you leave the ground, what do you actually control?" Ground can wait, so you pick a whole range of departure times, which becomes a range of arrival times at the satellite. State becomes (node, [a, b)): intersect with [start, end), shift by delay, prune if empty; collapse to the earliest time again at a ground node or gateway. Endpoints only come from link start/end, so finite states. Forgot to shift the right endpoint; a test came out one second early before I found it.
Part 4: each link carries at most k packets per window, deliver as many of a batch as possible. Said max flow on a time-expanded graph. Didn't work out how big the graph gets, out of time.
Round 2 β System Design
Controller computes new route tables and has to push them to thousands of satellites and gateways. Satellites are only reachable over ground stations.
First drew controller β queue β per-satellite agent pulling. He asked: a path crosses 5 satellites, 3 on the new table and 2 on the old, what happens to packets. Loops or blackholes.
Switched to versioned tables with an activation time T: push ahead, store, swap together at T. Controller confirms T only after all affected nodes ack; otherwise cancel the version and keep the old one. He pushed on partial activation. I said per path: once every node on a path acked, that path can switch. He said then the controller tracks ack state per path, which is huge. Discussed for a bit, didn't converge.
How long satellites keep the old table: at least one cycle past T, for rollback.
Last: satellite clocks drift, so T isn't exactly simultaneous; what about the few hundred ms of mixed tables. Said make-before-break, old and new paths both valid for a while. Didn't finish how to pick the guard time.
For finding roles and applying, LandOffer.ai aggregates job postings and auto-fills applications for free.