Object-oriented / low-level design problems, each in its own folder with the code and a short README (problem, the classes, the pattern used, how to run). Java and Python, mixed.
Each design has a small demo main you can run:
cd parking-lot && java Main.java # Java 17 or newer, no separate compile step
cd rate-limiter && python3 main.pyTo run every demo at once, from the repo root:
pip install -r requirements.txt # numpy, used by 11 of the Python demos
python3 run_all.py # one PASS or FAIL row per design
python3 run_all.py kv-cache-pool dynamic-batcher # only the designs you namerun_all.py exits with status 1 if any demo fails. It compiles and runs the Java demos when javac
is installed, and skips them when it is not. CI runs it on every push to main and on pull requests,
with Python 3.12 and Java 17.
48 designs in five groups: ML and AI infrastructure · ML building blocks · Systems and concurrency · Data structures and storage · Classic object-oriented interview problems
| Design | Language | Summary |
|---|---|---|
| kv-cache-pool | Python | Paged KV cache memory, as in vLLM's PagedAttention: sequences grow page by page from a shared pool, and LRU eviction skips sequences that are still generating. |
| dynamic-batcher | Python | Flushes a batch when it reaches the max size or the oldest request hits the max wait. A fake clock makes every batch boundary exact and testable. |
| inference-server | Python | Client threads send requests; one worker groups them into batches and runs one forward pass per batch, within a wait budget. |
| gpu-memory-allocator | Python | Device memory allocator with two strategies, best-fit free list and buddy system, compared on fragmentation as blocks are freed. |
| gpu-quota-manager | Python | Multi-tenant GPU scheduler with weighted fair shares. It takes back only GPUs above a tenant's fair share and reports Jain's fairness index. |
| llm-rate-limiter | Python | Limits requests per minute and tokens per minute at the same time, with two token buckets per client. |
| model-router | Python | Canary and weighted traffic splits across model versions, with fallback to the next backend when one fails. |
| model-registry | Python | MLflow-style model versions and stages (None, Staging, Production, Archived). Illegal moves are rejected and only one version is in Production. |
| experiment-tracker | Python | MLflow-style runs with params and metric series. A context manager marks each run finished or failed, and runs can be compared by a metric. |
| feature-store-client | Python | Online lookup of the latest feature values with staleness flags, and point-in-time-correct historical lookup for training data. |
| streaming-feature-aggregator | Python | Count, sum and average per entity over a sliding time window, updated as events arrive and age out. |
| ab-test-assignment | Python | Hash-based experiment bucketing: the same user always gets the same variant, splits follow the weights, and no assignment table is stored. |
| embedding-cache | Python | LRU cache that computes an embedding on a miss and tracks hit rate on a Zipfian query stream. |
| vector-store | Python | Top-k vector search with an exact brute-force index and an IVF (k-means) index, with IVF measured by recall@k. |
| prompt-template-engine | Python | Prompt templates with named variables and partial binding, rendered as plain text, a few-shot prompt or chat messages. |
| Design | Language | Summary |
|---|---|---|
| tokenizer | Python | Byte-level BPE tokenizer. It learns merges from a corpus, and encode then decode gives back any text exactly, even text it never saw. |
| mini-autograd | Python | Reverse-mode autodiff over scalars (+, *, **, relu, tanh), checked against finite differences. |
| tensor-ndarray | Python | A flat buffer with shape and strides: reshape, transpose, broadcasting and 2-D matmul, checked against NumPy. |
| compute-graph-executor | Python | Runs a target node's ancestors in topological order with a shared cache, so a shared node is computed only once. |
| kernel-dispatcher | Python | Picks a kernel by (op, dtype, backend), falls back to a generic kernel, and raises an error when no kernel exists. |
| data-loader | Python | PyTorch-style Dataset and DataLoader with batching, shuffling every epoch, transforms and drop_last. |
| ml-pipeline | Python | sklearn-style pipeline (imputer, scaler, one-hot encoder, k-NN) that learns its statistics from training data only. |
| Design | Language | Summary |
|---|---|---|
| thread-pool | Python | Fixed worker threads on a shared queue. submit returns a future that gives back the result or raises the task's error. |
| object-pool | Python | Thread-safe pool of reusable objects with a size cap. Objects are reset on release, and a context manager makes sure each one is returned. |
| priority-scheduler | Python | Priority job queue with aging, so low-priority jobs cannot starve. Uses lazy deletion on a heap. |
| circuit-breaker | Python | Closed, open and half-open states that stop calls to a failing service and test when it has recovered. |
| rate-limiter | Python | Token bucket, fixed window and sliding window log limiters behind one interface. |
| consistent-hash-ring | Python | Hash ring with virtual nodes. Adding a node moves only about 1/N of the keys. |
| metrics-aggregator | Python | Streaming count, mean and standard deviation (Welford), plus p50 and p99 from an exact or a reservoir-sampling sketch. |
| logging-framework | Java | A chain of log handlers (console, file, alert), each writing only messages at or above its own level. |
| notification-service | Python | Broadcasts a message to subscribed users over the channels each one picked: email, SMS or push. |
| Design | Language | Summary |
|---|---|---|
| lru-cache | Java | O(1) get and put with a HashMap and a hand-written doubly linked list. |
| cache-with-eviction-policies | Python | Fixed-size cache with swappable LRU, LFU, FIFO and random eviction. |
| bloom-filter | Python | Bloom filter sized from a capacity and a target false-positive rate, using double hashing. It never gives a false negative. |
| lsm-memtable | Python | LSM-tree write and read path: memtable, flushed sorted runs, tombstones and compaction. Reads always return the newest value. |
| Design | Language | Summary |
|---|---|---|
| parking-lot | Java | Multi-floor lot that fits motorcycles, cars and buses into sized spots, issues tickets and charges hourly fees. |
| elevator-system | Java | Several elevators, a nearest-elevator dispatcher and one class per elevator state, simulated floor by floor. |
| vending-machine | Python | Coins, product selection, change and sold-out slots, with one class per machine state. |
| atm | Python | Card and PIN session with limited attempts, balance checks, and cash split into notes. |
| library-management | Java | Books and their physical copies, checkout and return, and a notice to members when a book comes back. |
| tic-tac-toe | Python | Two players on a 3x3 board with win and draw detection. Scripted moves make the demo repeatable. |
| snake-and-ladder | Python | Multiplayer game on a 100-square board with snakes, ladders and a seeded die. |
| splitwise | Java | Shared expenses split equally, by exact amounts or by percentage, with the net amount each person owes the others. |
| movie-ticket-booking | Python | Shows, seat maps, seat locking and booking, payment, and confirmation notices. |
| car-rental | Java | Fleet search, reservations, and daily or weekly pricing, with a factory for vehicle types. |
| chess | Python | 8x8 board, six piece types with their own move rules, and turn order. Moves are pseudo-legal (no check detection). |
| hotel-booking | Java | Room search over date ranges, booking, and a checkout invoice with weekend pricing. |
| meeting-scheduler | Java | Books shared meeting rooms without double-booking and tells the organizer and attendees either way. |
Patterns show up where they fit — strategy, factory, state, observer, chain of responsibility — not forced.
New designs are welcome. Each one is a folder with a runnable demo and a short README. See CONTRIBUTING.md for the layout and a list of open ideas.
