Interview questions

Interview questions (with the counter-questions you'll get)

Q1. How would you find all restaurants within 3 km of a user?

They're testing: the core geospatial pattern.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Index restaurants by a spatial cell: geohash, a quadtree, S2 or H3, or use PostGIS or Redis GEO. Find the user's cell at a size roughly matching 3 km, fetch restaurants in that cell and its 8 neighbours, compute haversine distance, filter to 3 km and sort. Restaurants rarely move, so the index can live in the database and be cached.
3
4**Counter-questions:**
5- *"Why not just an index on latitude and longitude?"* → A B-tree sorts by one column, so a 2D range becomes a band scan.
6
7**Trap:** forgetting the neighbouring cells.

Q2. Explain how geohash works.

They're testing: whether you understand the encoding, not just the name.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Repeatedly halve the longitude and latitude ranges, recording a bit for which half the point is in, alternating between them; every 5 bits become a base-32 character. Longer strings mean smaller cells, and a shared prefix means the same larger cell, so prefix queries find nearby points.
3
4**Counter-questions:**
5- *"What's the edge problem?"* → Points close together across a cell border can have very different geohashes, so you always search neighbours.
6
7**Trap:** claiming that nearby points always share a prefix.

Q3. Geohash vs quadtree: when do you use each?

They're testing: trade-offs.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Geohash is a fixed grid encoded as a string, easy to store and index in any database and to shard by prefix, but cell sizes don't adapt to density. A quadtree adapts, splitting only crowded areas, which keeps each cell's count bounded, but it's usually an in-memory structure and harder to distribute.
3
4**Counter-questions:**
5- *"How do you keep a quadtree updated with moving riders?"* → Remove and reinsert when a point leaves its leaf; most moves stay in the same leaf. Or rebuild periodically.
6
7**Trap:** not mentioning density.

Q4. Why might Uber use hexagons (H3)?

They're testing: awareness beyond geohash.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** In a hexagon grid, all six neighbours are the same distance from the centre, unlike squares, where diagonal neighbours are farther. That makes rings, smoothing and movement between cells more even, which suits demand and pricing per area. H3 also has 16 resolutions, each about a seventh of the area of the coarser one.
3
4**Counter-questions:**
5- *"Any downside?"* → Hexagons don't subdivide perfectly into smaller hexagons, so parent and child cells don't line up exactly.
6
7**Trap:** saying hexagons are "just prettier".

Q5. How do you handle millions of location updates from moving drivers?

They're testing: write-heavy design.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Keep only the latest position per driver in an in-memory geo index (Redis GEO or in-memory cells), sharded by region. Update in place when the cell doesn't change. Expire drivers who stop pinging. Send the full stream to a log or queue for history and analytics, written in batches.
3
4**Counter-questions:**
5- *"What if the in-memory store crashes?"* → Positions are rebuilt within one ping interval, since every driver resends their location every few seconds. Replicate for faster recovery.
6
7**Trap:** writing every ping to the main relational database.

Q6. How do you find the K nearest drivers, not just those within a radius?

They're testing: a common variation.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Start with a small cell and its neighbours; if fewer than K candidates are found, widen to the next ring or a coarser cell, and repeat. Then sort by exact distance. Remember the K-th nearest candidate found in a ring might not be the true K-th nearest until you've checked far enough to cover that distance.
3
4**Counter-questions:**
5- *"Is straight-line distance good enough?"* → For a first filter, yes. For assignment, road distance or ETA matters more, so compute ETAs for the few nearest candidates.
6
7**Trap:** stopping as soon as K points are found without checking the distance bound.

Q7. How would you shard a geospatial service?

They're testing: combining Chapter 15 with geography.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** By region, usually city, because queries are local. Split hot cities further by cell prefix. Near boundaries, query both shards or let shards overlap. Monitor hot spots, which move with time of day.
3
4**Counter-questions:**
5- *"Why not shard by driver ID?"* → Then a nearby query has to ask every shard (scatter-gather), because nearby drivers are spread across all of them.
6
7**Trap:** hash-sharding by ID for location queries.

Q8. How do you compute the distance between two coordinates?

They're testing: a basic correctness detail.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** The haversine formula, using the Earth's radius of about 6,371 km, gives the great-circle distance. Euclidean distance on raw degrees is wrong because a degree of longitude shrinks with latitude, about 108 km at 13°N but about 56 km at 60°N.
3
4**Counter-questions:**
5- *"Is haversine exact?"* → It assumes a sphere; the Earth is slightly flattened. For delivery distances, the error doesn't matter.
6
7**Trap:** using degrees as if they were kilometres.

Q9. Two orders ask for the same nearest rider at the same moment. What happens?

They're testing: linking geospatial search to concurrency.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** The search only proposes candidates. Assignment must be atomic: a conditional update that marks the rider as assigned only if they're still available (Chapter 17), so one order wins and the other moves to the next candidate.
3
4**Counter-questions:**
5- *"Where does the 'available' flag live?"* → Close to the location data, or in a store that supports atomic updates, with the assignment recorded durably.
6
7**Trap:** assuming the geo index handles assignment.

Q10. Which database would you use for geospatial queries?

They're testing: practical knowledge.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** For moving points with high update rates, Redis GEO or a custom in-memory index. For stored places with rich queries, PostgreSQL with PostGIS. For search combined with location ("pizza near me"), Elasticsearch with geo queries. MongoDB's 2dsphere indexes are also an option.
3
4**Counter-questions:**
5- *"Why not build your own?"* → Usually you shouldn't; but for very high update rates and custom matching, companies do build in-memory indexes.
6
7**Trap:** naming a tool without saying why.
8
9---
10
11## Practice coverage
12
13Every section in this chapter is tested at least once.
14
15| Section | Tested by |
16|---|---|
17| 18.1 Why a lat/lng index fails | Level 1A.1, Level 2.2, Q1, Q8 |
18| 18.2 Spatial indexes | Level 1B, Q3 |
19| 18.2.1 Geohash | Level 1A.2, Level 1A.3, Level 3.1, Q2 |
20| 18.2.2 Quadtree | Level 1A.4, Level 3.5, Q3 |
21| 18.2.3 S2 and H3 | Level 1B.2, Q4 |
22| 18.3 Radius queries | Level 2.3, Level 3.2, Level 3.4, Level 4, Q1, Q6 |
23| 18.4 Moving drivers | Level 1A.5, Level 2.1, Level 2.4, Level 5.3, Q5 |
24| 18.5 Sharding by region | Level 5.4, Q7 |
25
26---

18. Geospatial Indexing