George Cristea

I implemented a directed acyclic graph used in a production system for entitlements where the read/write ratio is higher than 1,000. Additionally, the most frequent operation was to get the ancestors for a node. I extended the data structures from "using rational numbers to key nested sets" to implement a directed acyclic graph. The data structures described by the paper are multi-way trees. The node values in a sub-tree have an encoding in the interval defined by sub-tree root value and its right sibling node.

Work Done:
• Introduced a root node to handle cases not covered by the original paper (e.g., when a node doesn't exist in the system).
• Transformed the data structure to allow representation of a directed acyclic graph and to detect and prevent cycles from being introduced. The code uses several multi-way trees to represent multiple paths from the root node to leaf nodes and each sub-tree is stored as a MongoDB document.
• To ensure data consistency I used Redis locking for document updates. The reason for using Redis has to do with MongoDB concurrency model.
• Using the continued fractions encoding, I was able to ensure that read operations are executed in constant time in terms of the number of MongoDB queries.

I designed and implemented a cloud management system that supports multiple cloud providers, both public and private. The private cloud was based on OpenStack while the public cloud providers were AWS, GCP, and Azure. The system offers the ability to audit cloud accounts, provision new cloud accounts, and is integrated with a corporate finance system.

It was implemented using AWS Simple Workflow Service and had several workflows for each provider. The back-end was implemented using Node.js and PostgreSQL.

A Cassandra NoSQL storage system was trying to use a hard-coded set of ciphers. However in situations when not all ciphers are present/installed on the host the server was failing to create SSL sockets so I created a fix for this.

I implemented a backup manager for a Cassandra cluster that comprised of approximately 20 nodes. The backup manager was using a distributed lock to restrict the number of nodes that were performing backups to only one at a time.

The distributed lock was implemented in Cassandra using a keyspace where the node performing a backup would write its identity. The value stored in the keyspace had a higher expiration period to ensure that the backup progresses in the presence of a node failure.

I implemented an Erlang port to interface with a Berkeley DB High Availability C library. The Berkeley DB HA offers full replication of data from the master to slave nodes.

The Erlang port driver was running the same process as the Erlang VM and it kept track which node is was the master. All the reads were served from the local Berkeley DB storage while the write operations were forwarded to the master using Erlang's native remote procedure calls.

I designed and implemented an engine for calculating the order total, taxes, discounts, and shipping charges. The library's C++ code was generated by a compiler from field dependency declarations. The compiler extracted semantic information from the input (e.g., the price to be displayed should include taxes for EU countries).