This is a Java 8 implementation of a distributed file system with caching. Each proxy server can handle multiple clients concurrently, while multiple local proxies connect to a centralized server. The file-caching proxy employs a check-on-use technique to guarantee consistency at open-close session granularity, similar to the andrew file system (AFS). It also supports whole-file caching with LRU eviction.

This is a Java 8 implementation of the distributed consensus protocol Two-Phase Commit. In Phase I, the coordinator proposes a transaction for each participant in hope of unanimous consent. Upon receiving all votes from participants, it enters Phase II to distribute the final decision (either commit or abort) to each participant. Both the coordinator and participants use Write-Ahead logging for persistent state and crash recovery. A timeout mechanism is implemented to handle message lost.

BusTub is a relational database management system written in C++17. It is part of the database project course taught by Professor Andy Pavlo at CMU. I implemented the following four components of the system.

• A Buffer Pool Manager is in charge of the mapping from disk pages and memory pages (Think of it as an OS built specifically for databases). It can fetch disk pages into memory and handle eviction following the LRU-K policy.
• A concurrent extendible hash table-based indexer and iterator for faster retrieval and traversal of table records.
• A query execution engine with rule-based optimiser. To support real SQL queries, it follows the volcano model and implements core executors such as sequential scan, index scan, insert, delete, projection, filter, nested loop join, hash join, aggregation, window functions, sort, and limit. The rule-based optimiser employs common techniques including predicate push-down, projection push-down, and top-N to come up with faster execution plans.
• Multi-Version Optimistic Concurrency Control (MVOCC). To manage concurrent transactions, MVOCC uses timestamp to define the visibility of each transaction and ensures that readers and writers do not block each other. It supports both SNAPSHOT isolation level and SERIALIZABLE isolation level. A garabage collector is invoked periodically in a separate thread to prune the version chains and remove entries no longer visible by any transactions.

This is a client-server implementation of remote file procedure calls written in C. The client side offers interposed file system calls including open, close, read, write, lseek, stat, unlink, getdirentries etc. The server could handle multiple clients operating on different files concurrently. These RPCs are implemented on top of TCP connection. Both the client/server uses customized stub to serialize/deserialize the function parameters, results, and errno.

QryEval is a full-fledged search engine based on the open source Apache Lucene. It leverages the python inheritance properties for extensibility and presents a uniform interface to users. It features the following key components:

• Query Parser: morphological preprocessing and tokenization
• Ranker: basic boolean search operators, such as exact match, BM25 and Indri
• Pesudo Relevance Feedback: self-supervising query expansion
• Diversification: explicit query intents to faciliate xQuAD and PM2
• Reranker: supports BERT-based dense retrieval and Learning-to-Rank based model