Flipkart

PART 1: ALGORITHMIC PROBLEM - DISTRIBUTED TASK SCHEDULER BASE PROBLEM You are designing the core execution engine for a distributed build system. You are given a list of tasks, where each task has a specific execution time, and a list of dependencies (e.g., Task B cannot start until Task A finishes). Task: Design an algorithm to calculate the minimum time required to complete all tasks. tasks: An array of integers where tasks[i] represents the time it takes to execute the i-th task. dependencies: An array of pairs [u, v] indicating that task u must be completed before task v can begin. Assuming you have an infinite number of worker nodes (meaning any number of independent tasks can run perfectly in parallel), what data structure and traversal method will you use to find the minimum total completion time? FOLLOW-UP 1 Your infrastructure budget has been cut. You no longer have infinite workers; you now have exactly K worker nodes. How do you modify your scheduling algorithm to minimize the total execution time while respecting both the dependencies and the K concurrency limit? FOLLOW-UP 2 The workers are now distributed across different data centers, and network partitions are common. Tasks can fail mid-execution and must be retried. If a task fails, all of its downstream dependencies must remain blocked. How would you design the backend architecture and state management to handle distributed task execution, failure detection, and dynamic retries without causing race conditions? PART 2: AI DISCUSSION QUESTIONS What is Tokenization in the context of LLMs? Tokenization is the process of breaking down text into smaller pieces called tokens (which can be words, subwords, or individual characters) before feeding them into a model. Since models only understand numbers, these tokens are mapped to numerical IDs. A token is typically about 3/4 of a word in English. What is Prompt Injection? Prompt injection is a security vulnerability where a user deliberately provides input designed to override or manipulate the original instructions (the system prompt) given to the AI. This can trick the model into ignoring its safety guardrails, leaking sensitive information, or executing unintended commands. How does a Mixture of Experts (MoE) architecture work? Instead of a single massive neural network processing every query, an MoE model consists of several smaller, specialized sub-networks ("experts"). When a prompt is received, a routing mechanism determines which specific experts are best equipped to handle that particular input, activating only a fraction of the total model. This allows for vastly scaled model capacity without a proportional increase in compute cost per token. What is a System Prompt? A system prompt is the foundational set of instructions given to an AI model by its developers before it ever interacts with a user. It defines the model's persona, its operational boundaries, formatting rules, and safety constraints. It acts as the invisible framework guiding how the model should interpret and respond to user inputs.