Smart Video Filter project

After 2 years and 3 months of efforts I finally released the Smart Video Filter project. It started as a long hope of mine from several years ago: as a YouTube viewer, I should be able to filter only universally liked videos with high ratio of likes to dislikes. The YouTube search results and recommendations often contain objectively bad videos, e.g. which are low quality/take 10% of the screen or contain static pictures or are offensive/cruel or are aimed at advertising some product with an unrelated video. All those videos would get a lot of dislikes and relatively few likes, such that their ratio = likes/dislikes is low. The videos I would really like to watch are the ones with very high ratio = likes/dislikes. These videos are objectively amazing and are out of this world. Hence the idea of the filtered search, where for any combination of search terms the user could filter only the videos with high enough "ratio". The idea can be trivially extended to ratios of likes to views, comments to views, dislikes to views, etc.

A small concern in the above approach is that videos of similar quality with fewer views tend to naturally have high ratios of likes to dislikes (or likes to views). The videos with fewer views tend to be watched by subscribers/aficionados/lovers. As the video gains popularity, it gets exposed to a wider audience, which typically likes it less. Then the ratio goes down. The primary goal is to find the best of all videos. Then we can for each number of views check how many videos have a higher ratio of, e.g. likes to views and only provide the results in top x% of the ratio. A uniform sample of videos across all range of views would be readily obtained by selecting videos in top x% (of some ratio) without specifying the target number of views. That is precisely the idea behind  Smart Video Filter.

Though the idea appears simple, the implementation took many evenings over a long period of time. The project is quite rich in required expertise: UX design, architecture, back-end engineering, front-end engineering, and devops skills. The technologies include NoSQL databases MongoDB and Elasticsearch, front-end technologies Angular 5 and Angular Material from Google, CentOS 7 administration and cluster administration. The project went through several stages. The initial UI was written in AngularJS and then rewritten in Angular 5. The initial backend datastore included PostgreSQL, which was unable to support the required I/O loads and was switched to MongoDB. Initially the project used Elasticsearch 2.1 and then gradually migrated up to Elasticsearch 6.2. Retrieval and refresh of videos and channels metadata need to rely on non-trivial algorithms to be efficient and not use up a daily quota before noon. All that relies on heavily multi-threaded and fault-tolerant Java backend. The underlying cluster system implements high availability, when after a complete loss of one machine all systems still work. YouTube terms of service and developer policies are quite strict and took a while to comply with. Recently, I got an audit review from YouTube compliance team, and well, they aren't shutting me down yet.

I'm greatly enjoying the final implementation myself, while searching for the best videos and filtering over duration in a fine-grained way. My long-term hope came true! Even now I got heavily distracted using the service instead of writing the post. 46:1 ratio video of Donald Trump "singing" Shake it Off by Taylor Swift is an amazing composition! The project has obvious ways to improve to substitute the YouTube original search even more: provide recommendations and play videos on the site without redirecting to YouTube. However, Smart Video Filter is a "small market share" application. If the "market maker" YouTube itself was to implement it, then lots of videos would not be regularly shown in search results/recommendations, which would have discouraged the content creators. Hope you enjoy this niche service as I enjoy it myself!

HA (high-availability) setup for Smart Video Filter

With high expectations on website and service availability in 2018, it is especially important to ensure the redundant DR (disaster recovery) copies of the service are running at all times and are ready to take on the full PROD load within seconds. Hosting companies like Amazon have long solved this problem for standard services, e.g. for Elasticsearch cluster. Since cluster always runs with 1 or more replicas, the replica node is ready to take over for a short period of time, till a new primary is spun up and synched after the primary failure. A level of abstraction such as Kubernetes also allows creation of high-availability service.

With all available options, what should we use in the real world? It depends on the budget and the available hardware. My recently released service, Smart Video Filter, is a low budget solution working on 2 physical machines running CentOS 7. Two enterprise-grade SSDs with large TBW (terabytes written) resource is substantially cheaper than AWS in terms of storage cost and provisioned IOPS cost. It is recommended to run HA setups with 3 machines, but 2 machines (PROD and DR) provide enough reliability and redundancy in most cases. Four different services on those machines needed to seemlessly switch between PROD and DR: ElasticSearch, MongoDB, Mining Service, and Search Service.

ElasticSearch setup over 2 machine includes creating a cluster with 1 primary and 1 replica. ElasticSearch reads and writes happily proceed even if one of those nodes is down. No special setup is necessary.

MongoDB setup on 2 nodes is trickier. MongoDB has a protection against a split brain condition. The cluster does not allow writes, if a primary is not chosen. A primary can only be chosen by the majority of nodes and there is no majority with 1 out of 2 nodes down. Addition of an Arbiter instance is recommended in such cases. However, a simple arbiter setup isn't going to work, if the arbiter is deployed on one of data nodes. If the entire node goes down, then it takes the arbiter down with it. What I ended up implementing is a workaround of the split-brain protection, when a MongoDB config is overwritten by the mining service. The mining service provides an independent confirmation that one of data nodes is dead, and adds an Arbiter on a different machine to the cluster, while removing an arbiter running on the same machine as the failed data node. Node health detection by the mining service is described below.

Search Service makes use of the health API. One instance of the service is deployed on PROD and DR node each. Each instance deploys a RESTful endpoint with a predictable response, which consists simply of the string "alive". Each instance also deploys a client to read this status from itself and from the other node. When both nodes are alive, the PROD node takes over. When DR node detects that it is alive, but the PROD node is not alive, then it takes over. Each node is self-aware: it detects its role by comparing its static IP (within a LAN network) to the defined IPs of the PROD and DR nodes. When the node takes over, it uses port triggering on the router to direct future external requests to itself. It was shown within testing that port triggering can switch the routed node within seconds.

Mining Service employs same health API + another external API, which reports, whether the job is running. When PROD or DR nodes are ready to take over, they let the current job finish on another node, before scheduling jobs on itself. Jobs should not run on PROD and DR node simultaneously. The health detection also helps to switch a MongoDB Arbiter between nodes to ensure MongoDB can elect a primary.

After the full setup is implemented, the system is capable of correctly functioning with only disruptions of several seconds, when one of the machines goes down entirely. This was readily demonstrated within testing, when all services remained highly available throughout a rolling restart of 2 machines!

Specialization Review – Leading People and Teams (University of Michigan)

Here is my review of "Leading People and Teams" specialization taken on Coursera from Aug 2017 till Jan 2018. Courses in this specialization are ranked very high between 4.5 and 5.0. I passed with the average grade of 99.0%. The specialization consists of 3 courses focusing on leadership and team work, 1 course emphasizing human relationships (HR) side of management, and a capstone project.

The first course, "Inspiring and Motivating Individuals", is quite inspirational indeed. Surprising research evidence suggests that most employees around the world are not engaged/motivated at work, and lots of them are even actively disengaged. The course outlines the origin of meaning of the work, the importance of company vision and engagement, the drivers of people motivations, and the ways to align the employees with the company's goals. 

The second course, "Managing Talent", is aimed primarily at managers conducting the onboarding, managing performance and evaluations, coaching the team members, and maintaining continuity of talent. Research shows that managers play crucial role in personnel turnover. A variety of organizational behavior effects and biases are discussed, such as Dunning-Kruger effect, availability error, racial bias, and gender bias. Knowledge from this course might, just like CSM and PMP certifications, backfire in startups or companies without rigid structure, where many of the standard techniques are not followed.

The 3-rd course, "Influencing people", is probably the most practical of the specialization. If outlines the bases of power and the bases of strong relationships with people and goes in great depth with examples. The course offers practical advice on how to positively interact with colleagues, how to build relationships, and how to gain influence, while protecting oneself from unwanted influence. Expert knowledge, information power, and referent power are presented as influencing means without formal authority. The material assumes a workplace in US, which provides great insight into the informal expectations for immigrant workers. E.g. the expected level of socializing at the workplace is different around the world, and is somewhat higher than average in the US.

The 4-th course, "Leading teams" takes it to the higher level of team dynamics. It provides practical advice for improving team work, coordination, output, and overall happiness. The course discusses topics as team structure, team size, subteams and splits based on demographics/similarity. Coordination problems and common design making flaws are emphasized and the prevention methods are presented. Psychological safety is presented as a cornerstone for team performance. Team charters and team norms are discussed. Performance-oriented vs. learning-oriented mindsets are shown to provide different outcomes.

The final capstone project, "Leading People and Teams Capstone" is automatically graded as a pass. It offers 3 options on improving leadership skills: (1) solve a real-world leadership business case, (2) take on a leadership challenge at work, or (3) interview a business leader to gain insight of their practices. The option (2) is probably best aligned the main goal of the course to improve the learner's leadership skills.

Overall, I had a great experience taking the specialization. It emphasizes that leadership skills is not something a person born with. They should and readily are acquired as a result of systematic work. A lot of material is focused on leading without formal authority, which is especially helpful to team members of self-organizing Scrum teams in the software industry. The courses are filled with real-life stories and interviews with people from the industry, which help solidify the concepts. Many pieces of homework are peer graded. Assignments of the others provide insight into ideas, styles, and techniques of people at various stages of career ladder. Those techniques summarize real-life experiences of people managing their subordinates, resolving conflicts, influencing the team, which might not otherwise be accessible to the learners.

Specialization is taught by instructors from the University of Michigan, Ross School of business: Scott DeRue, Full Professor, business school Dean; Maxym Sych, Associate Professor; Cheri Alexander, Chief Innovation Officer. All three are charismatic, knowledgeable, and are great presenters. The material is delivered very coherently and to the point. The lecture slides are very detailed and are great for returning the the material in the future.