Inspired by a tutorial on TensorFlow that was on HN recently I decided to go and read the TensorFlow paper. This paper has been sitting in my “To Read” folder for quite some time now but for various reasons I never got around to reading it. This is also the first AI/ML paper I’ve read in 2016 so I was excited to dive right in.

At 19 pages long this is one of the longest papers I’ve read. But it is extremely well written, with lots of diagrams, charts, and code samples interspersed throughout the text that make this paper fun to read.

The basic idea of TensorFlow, to have one system that can work across heterogenous computing platforms to solve AI/ML problems, is incredibly powerful. I fell in love with the directed graph API used by TensorFlow to describe computations that will run on it (this may or may not be related to the fact that I also love graph theory). The multi-device (and distributed) execution algorithm explained in the paper is quite intuitive and easy to understand. A major component of multi device / distributed execution of the TensorFlow graph is deciding which device to place a node on. While the paper does explain the algorithm used in section 3.2.1 I wish they had gone into more details and talked about what graph placement algorithms didn’t work, details about the greedy heuristic used, etc.

Sections 5, 6, and 7 were my favorite portions of the paper. Section 5 dives into some of the performance optimizations used in TensorFlow. It would have been awesome if the authors had given more details about the scheduling algorithm used to minimize memory and network bandwidth consumption. I would have also liked to know what other scheduling optimizations were used in TensorFlow as I find scheduling algorithms very interesting.

Section 6 talks about the experience of porting the Inception model over to TensorFlow. While the strategies mentioned in this section are specific to machine learning systems, I feel that some of them can be tweaked a little bit to be generally applicable to all software systems. For instance

“Start small and scale up” (strategy #2)

is directly applicable to any software system. Similarly,

“Make a single machine implementation match before debugging a distributed implementation” (strategy #4)

Can be rephrased as

“Make a single machine implementation work before debugging a distributed implementation”

and be generally applicable to building distributed systems.

Section 7 explains how TensorFlow can be used to speed up stochastic gradient descent (SGD). Again, while the idioms presented in this section are used to speed up SGD, I feel that they are general purpose enough where they can be applied to other algorithms/systems as well. The diagrams in this section are amazing and do a great job of illustrating the differences between the various parallelism and concurrency idioms.

EEG, the internal performance tool mentioned in the paper, sounds very interesting. While it is probably not in the scope of a paper that focuses on TensorFlow I’d love to learn more about EEG. It seems like a very powerful tool and could probably be extended to work with other systems as well.

The paper ends with a survey of related systems. This section proved to be a valuable source for finding new AI/ML and systems papers to read.

I loved this paper.




(inspired by this post)

Here is a (definitely incomplete) list of programming / technology related websites (in no particular order) that I frequently read —

  • High Scalability: Excellent articles on real world software architecture and design.
  • Julia Evan’s Blog: Well written and fun to read posts on a wide range of topics from system programming to machine learning.
  • Preshing on Programming: A great resource for data structures and C++.
  • The Morning Paper: Research papers. Research papers. Research papers.
  • Hacker News: I feel this is the best resource to stay up-to-date on what’s happening in the fields of technology, software engineering, and computer science.
  • Dan Luu’s blog: I hope to one day be able to write technical posts as well as they do.
  • Incredible posts about system engineering concepts.


(working through more Hacker News backlog)

I just finished Think OS: A Brief Introduction to Operating Systems. It had been sitting in my browser for quite some time now and I decided to read through the book this afternoon.

The book stays true to its title and gives a short and sweet introduction to (some) OS and systems programming fundamentals. While it doesn’t dive deep into any topic (for example — TLBs are not mentioned in the chapter about Virtual Memoryit does a good job of laying a foundation for people who have never taken a systems programming / OS course before and are trying to gain an understanding of the fundamentals. That being said there are links to resources (like an explanation on how dlmalloc is implemented) scattered throughout the book that allow the reader to gain a deeper understanding of the topic at hand.

If you’ve never done any systems programming before and are looking to write some low level code this book is a good place to begin to understand how the OS and kernel work. If you’re looking for something more in-depth I’d highly recommend reading Operating Systems: Three Easy Pieces.


The other day I was working through my (quite massive) backlog of saved Hacker News stories and read this gem — Introduction to PostgreSQL physical storage. As someone who has loved databases forever, and has spent over a year building LinkedIn’s next generation distributed graph database, I found this post absolutely fascinating.

As the title suggests this article talks about how data in PostgreSQL tables and databases are actually stored on disk and how free space (to figure out where to store incoming data) is managed. The diagram of the page structure is very helpful in understanding how data is stored in a page. I also really liked the use of PostgreSQL queries used throughout the article to explain the topic at hand by examining a real PostgreSQL instance. The author does a great job at explaining concepts at just the right amount of detail, with several links provided for those interesting in learning more.

(Relevant side bar: the PostgreSQL source code documentation is amazing)


One long standing issue that I’ve had with the LinkedIn GitHub page that I helped design was that it because it relied on the GitHub public API to fetch all the data, if the user accessed the page from a rate-limited IP address the rendered page would be blank as no data would be returned by the API. I had some time on my hands today and decided to fix this bug.

The simplest fix for this bug is to cache the GiHub API response in a file, and when you get rate-limited by the GitHub API fall back to reading from a cached API response. Since the raw API response contained lots of information that was not required to generate the website, I decided to add an intermediate filtering step to only extract the relevant information from the raw GitHub API response. The JSON data generated by this filtering step is the final cache used by the webpage.

To test the code I’d written and to make sure everything works as expected I needed to rate limit myself. This was easily done using the (amazing) Python Requests library.

You can find my fix for this bug here.

Update — I realized that my original patch failed to use the GitHub API response when the user was not rate limited. My last commit should fix this.



Inspired by Spotify’s year in music feature (I wrote a post on it as well), I decided to analyze music related data that I had at my disposal. The data that I chose was the list of all the artists that I’ve seen live (78 at the time of doing this analysis).

There were two things that I wanted to surface from this data:

  1. Which genres of music have I seen the most live?
  2. Which artists should I see next, based on the artists I’ve already seen?

To answer both these questions I decided to use the Echo Nest API. And Python. All the code I wrote to analyze the data can be found here. I wrote this code when I should have been sleeping so the quality is not the best. Oh well.

About halfway through writing the code I decided that generating a word cloud for #1 would be cooler than simply listing the top genres. After failing miserably to get word_cloud working on my machine I decided to use an online word cloud generator instead. Here’s the resulting word cloud:

Screen Shot 2016-02-23 at 6.58.51 PM

The technique I used to answer #2 was to get the list of similar artists for each artist I’ve seen live, remove artists that I’ve already seen, and keep track of how many times each unseen artist is listed as a similar artist. Here are the top recommendations generated by my algorithm (format: <artist, number of times listed as similar artist>):

  1. Swedish House Mafia, 5
  2. The Raconteurs, 4
  3. Cut Copy, 3
  4. Beach Fossils, 3
  5. Kaiser Chiefs, 3
  6. Iron Maiden, 3
  7. Dio, 3
  8. Ellie Goulding, 2
  9. Black Sabbath, 2 (seeing them in September)
  10. Animals as Leaders, 2

My recommendation algorithm is extremely simple but produced surprisingly good results.

The Echo Nest API is incredible.

P.S. I tried using pyechonest but there didn’t seem to be a way to retrieve artist genre information which is why I decided to use their API directly. 


I loved Fred’s post on the Zen of Erlang. I decided to check out his blog on the bus ride back from work today and read a few of his other posts. Two posts stood out to me.

Lessons Learned while Working on Large-Scale Server Software is, in my mind, required reading for any software engineer working on backend and infrastructure systems. Knowledge of lot of the concepts mentioned in this post (like the CAP Theorem, or the Fallacies of distributed computingis essential in developing robust software systems. Fred’s style of writing is lots of fun to read, and I really his views on computer networks in this post —

There’s nothing more dangerous than someone going for a stroll over the network without knowing what it entails.

The network owes you nothing, and it doesn’t care about your feelings. It doesn’t deserve your trust.

The network is a necessary evil, not a place to expand to for fun.

The second post that stood out to me on how Queues Don’t Fix Overload. He explains in simple terms why queues (when used incorrectly) seem to solve your scaling problems in the short run while introducing a whole new class of problems of their own. As mentioned in the post, identifying bottlenecks in your system and scaling and fixing those is the correct way to deal with system overload.


Have you ever been in a situation in which something has “gone wrong” (intentionally vague) between two git commits, say c1 and c2, and you’re trying to figure out which commit caused the issue? In other words, your code works fine at c1, but not at c2. Thus, a commit in the range (c1, c2] resulted in your code being in a “bad” (for some definition of “bad”) state.

One approach is to look at all the commits between (c1, c2] and see if any commit stands out as something that might have caused the issue. But there are times when looking at the changes is not enough, or it’s not clear why any of the changes would have broken anything, and you need to do some other work (run integration tests, run performance suites, run UI tests, etc.) in order to pinpoint the breaking commit.

“Why, this seems like a perfect opportunity to use binary search to figure out which commit caused a problem! All I need to do is a binary search in the range (c1, c2]. For a particular commit in this range (starting in the middle) I simply need to git checkout the code at that point, do whatever work I need to (explained above), and then make a decision on whether I need to search in the ‘upper half’ or ‘lower half'”

Enter git bisect. It allows you to focus on what went wrong, without having to manage the git + binary search state. In our scenario we’d simply mark c1 as a good commit, and c2 as a bad one, and then let git bisect work its magic in enabling us to discover what went wrong between (c1, c2].

I love git.


After 24+ hours of traveling I’m back in San Francisco! The long journey gave me a lot of time to think. And read. And sleep (I think my superpower is falling asleep on airplanes).

Here are some of the things I read:


One of my vacation goals was to work through my Pocket list. While I’m not close to being done (I’ve only been looking at the Computer Science related links. I also have a number of music related links on my list) here are the highlights so far: