It’s really difficult to write about E. M. Forster

When I started getting into A Room with a View a few years back, I kept reading it over and over again. In each reading, I would find something new. Something would be changed, revealed, or transmuted into something new. It was fascinating. But there were still things that I didn’t understand. Some of the places and artiss were obscure. Some of the referenced literature was very obscure.

I started taking notes on the novel. Whenever I came across something that was obscured by distance, time, or education, I would look it up online and produce a little note to myself about it. Eventually, I had enough of these to think, ‘hey, I should share these with other people!’

That’s where things got tricky. After all, I don’t want to share spurious or incorrect notes. I want to share notes that increase the enjoyment of the book for casual readers and fans. But how can I know if one of my notes is incorrect? How can I be sure that what I’m pointing out isn’t very subjective or obvious?

So I thought I should at least read his other novels to get some context. I read his first novel first. Where Angels Fear to Tread is, in most respects, not a great novel. But it does begin to reveal Forster’s unique approach to realism. Next I picked up The Longest Journey. That book is truly boring. It was a slog to get through, and I don’t know how anyone enjoys it without knowing a lot about Forster’s biography.

Then I read The Machine Stops, which is fabulous and unique and ahead of its time. So I thought maybe Forster had a particular gift for short fiction? So I read all of his short fiction that was in collections (which I know now is not all his short fiction). It was excellent, but there wasn’t very much of it. So I went back to the novels. I finished with Howard’s End, A Passage to India, and Maurice, each of which is a masterpiece in its own way.

Then I went back to my original task. I could finally say something about Forster’s most popular book from a position of authority, right?

Well, no. I soon learned that Forster produced even more essays and non-fiction than he produced fiction. I read his guide to Alexandria, and some of his essay collections. I couldn’t get through it all, and some of his collections are difficult to acquire these days.

Then I thought I should read a biography. So I read Wendy Moffat’s excellent book.

At this point, I’m years away from my original task. I have more or less forgotten that I ever wanted to say anything about Forster and his most famous novel. Still, I’ve realized that he also wrote a lot of letters, and the selected letters are available in bound collections. I’m duty-bound to read them, right?

I read as much of the letters as I could, but I think you can see the problem. It’s really hard to say anything with authority about a writer who produced such a massive volume of words as EM Forster. The novels, essays, non-fiction, lectures, and letters amount to such a vast quantity of work that I would put it up against even the most fecund modern novelists. It’s unbelievable.

Then there’s the critics, biographers, and academics. I can see how a grad student could get very dis-heartened. How can you hope to add to the vast discourse on such a popular, beloved author? …

I don’t know if I will ever finish taking notes on A Room with a View (and now Forster’s other novels too). All I know at this point is that the journey has become the goal. Reading and studying Forster’s work and the work about Forster has become a little hobby of mine. Maybe I won’t ever say anything about Forster, but I’ll have a lot of fun not saying it!

SoundCloud, I love you, but you’re terrible

I finally started using SoundCloud for a new jazz/electro project called Fynix. I casually used it in the past under my own name, in order to share WIP tracks, or just odd stuff that didn’t fit on bandcamp. But I never used it seriously until recently. Now I am using it every day, and trying to connect with other artists. I am remixing one track a week, listening to everything on The Upload, and liking/commenting as much as I can.

SoundCloud is the best social network for musicians right now. But it still has a terrible identity crisis. Most of the services seem to be aimed at listeners, or aimed at nobody in particular.

So in this post, I’m going to vent about SoundCloud. It’s a good platform, but with a few changes it could be great.

1. I am an artist. Stop treating me like a listener.

Is it really that difficult for you to recognize that I am a musician, and not a listener? I’ve uploaded 15 tracks. It seems like a pretty simple conditional check to me. So why is my home feed cluttered up with reposts? Why can’t I easily find the new tracks by my friends?

This is the core underlying problem with SoundCloud. It has two distinct types of users, and yet it treats all users the same.

2. Your “Who to Follow” recommendations suck. They REALLY suck.

I’ve basically stopped checking “Who to Follow” even though I want to connect with as many musicians as possible. The recommendations seem arbitrary and just plain stupid.

The main problem is that, as a musician, I want to follow other musicians. I want to follow people who will interact with me, and who will promote my work as much as I promote theirs. Yet, the “Who to Follow” list is full of seemingly random people.

Is this person from the same city as me? No. Do they follow lots of people / will they follow back? No. Are they working in a genre similar to mine? No. Do they like and comment on lots of tracks? No.

So why the heck would I want to follow them?

3. Where are my friends latest tracks?

This last one is just infuriating. When I log in, I want to see the latest tracks posted by my friends. So I go to my homescreen, and it is pure luck if I can find something posted by someone I actually talk to on SoundCloud. It’s all reposts. Even if I unfollow all the huge repost accounts, I am stuck looking at reposts by my friends, rather than their new tracks.

Okay, so let’s click the dropdown and go to the list of users I am “following”. Are they sorted by recent activity? No. They are sorted by the order in which I followed them. To find out if they have new tracks, I must click on them individually and check their profiles. Because that is really practical.

Okay, so maybe there’s a playlist of my friends tracks on the Discover page? Nope. It’s all a random collection of garbage.

As far as I can tell, there is no way for me to listen to my friends’ recent tracks. This discourages real interactions.

Ultimately, the problem is data, and intelligence. SoundCloud has none.

You could blame design for these problems. The website shows a lack of direction, as if committees are leading the product in lots of different directions. SoundCloud seems to want to focus on listeners, to compete in the same space as Spotify.

But even if that’s the case, it should be trivial to see that I don’t use the website like a regular listener. I use it like a musician. I want to connect and interact with other musicians.

And this is such a trivial data/analytics problem that I can only think that they aren’t led by data at all. Maybe this is just what I see because I lead our data team, but it seems apparent to me that data is either not used, or used poorly in all these features.

For instance, shouldn’t the “Who to Follow” list be based on who I have followed in the past? I’ve followed lots of people who make jazz/electro music, yet no jazz/electro artists are in my “Who to Follow” list. I follow people who like and comment on my tracks, yet I am told to follow people who follow 12 people and have never posted a comment.

The most disappointing thing is that none of this is hard.

4. Oh yeah, and your browser detection sucks.

When I am browsing your site on my tablet, I do not want to use the app. I do not want your very limited mobile site. I just want the regular site (and yes, I know I can get it with a few extra clicks, but it should be the default).

Tips for Managing Joins in Looker

Looker is a fantastic product. It really makes data and visualizations much more manageable. The main goal of Looker is to allow people who aren’t data analysts to do some basic data analysis. To some extent, it achieves this, but there are limits to how far this can go. Ultimately, Looker is a big graphical user interface for writing SQL and generating charts. Under-the-hood, it’s programmable by data engineers, but it’s limited by the fact that non-technical users are using it.

The major design challenge for Looker is joins. A data engineer writes the joins into what Looker calls “explores”. Explores are rules for how data can be explored, but ultimately just a container for joins. When someone creates a new chart, they start by selecting an explore, and thus selecting the joins that will be used in the chart.

They pick the join from a dropdown under the word “Explore”. This is the main design bottleneck. Such a UI encourages users to have only a limited number of joins that can fit in the vertical resolution of the screen. This means limiting the number of explores, and hence limiting the ways tables are joined. This encourages using pre-existing joins for new charts.

This creates two problems.

  1. A non-technical user will not understand the implication of choosing an explore. They may not see that the explore they chose limits how the data can be analyzed. In fact, a non-savvy user may pick the wrong explore entirely, and create a chart that is entirely wrong.
  2. The joins may evolve over time. A programmer might change a join for a new chart, and this may make old charts incorrect.

The problem is that SQL joins are fundamentally interpretations of the data. Unless a join occurs on id fields AND is a one-to-one relationship, then a join interprets the data in some way.

So how can you limit the negative impact of re-using joins?

1. Encourage simple charts

Encourage your teammates to make charts as simple as possible. If possible, a chart should show a single quantity as it changes over a single dimension. This should eliminate or minimize the use of joins in the chart, thus making it far more future-proof.

2. Give explores long, verbose names

Make explore names as descriptive as possible. Try to communicate the choice that a user is making when they choose an explore. For instance, you might name one explore “Products Today” and another one “Product Events Over Time”. These names might indicate that the first explore looks at the products table, but the second explore shows events relating to products joined with a time dimension.

One of the mistakes I made while first starting out with Looker is naming the explores with single word names. I now see that short names create maintenance nightmares. Before assessing the problems with a given chart, I need to know which explore the maker chose for it, and because the names were selected so poorly, the choice was often incorrect.

I hope these ideas help you find a path to a maintainable data project. To be honest, I have a lot of digging-out to do!

Pride in Software Craftsmanship

As I spend more and more time in Silicon Valley, my views on software management are changing. I read Radical Candor recently, and while I agree with everything in it, I feel like it over-complicates things.

This meditation has been pushed in part by my passion for food. I like going to new restaurants. It brings me joy to try something new, even if it’s not a restaurant that would ever be considered for a Michelin Star. Even crappy looking restaurants can serve great food.

I am often awed by the disconnect between various parts of the restaurant business and the quality of the food. Some restaurants are spotlessly clean, have have beautiful decor, and amazing service… but the food is mediocre. The menu is bland and uninspired, and the food itself is prepared with all the zeal that a minimum wage employee can manage.

Then I’ll go to a dirty looking greek joint down the road, and the service will be awful… but the menu is inspired. It’s not the standard “greek” menu, but it’s got little variations on the dishes. And when the food comes out (finally), maybe it isn’t beautiful on the plate, but the flavors come together to make something greater than the ingredients and the recipe.

What seems to distinguish a good restaurant from a crappy one is pride. At restaurants that I return to, there is someone there, maybe a manager, maybe a cook, maybe the chef who designed the menu, who takes great pride in his work.

There’s a diner by my old house, for instance, where the food is … diner food. There’s no reason to go back to the restaurant… except for the manager. The man who runs the floor, seats the patrons, deals with the kitchen, and does all the little things that make a restaurant tick. He manages to make that particular diner worth going to. And for a guy who has two young kids, that’s terrific.

I am starting to think that the same basic principle applies to software engineers. I’ve met brilliant engineers with all sorts of characteristics. Some of them have a lot of education and read all the latest guides. Others have little education, and don’t read at all. The main thing that makes them good engineers is that they take pride in their work. They care about the quality of their work, regardless of how many people are going to use it, or how much time they put into it. They write quality code because their work matters.

So when it comes to managing software projects, I’m starting to think that all of these systems boil down to two basic steps.

  1. Put your engineers in a position to take pride in their work.
  2. Get out of the way.

Obviously, the first step is non-trivial. It’s why there are so many books on the topic. But at the end of the day, pride is what matters.

Book Review: Life of a Song

I recently had the chance to read Life of a Song: The fascinating stories behind 50 of the worlds best-loved songs. It’s a concise collection of fifty Life of a Song articles from the Financial Times. As I rarely have a reason to visit the FT website, and I only occasionally catch the Life of a Song podcast, the book was a great opportunity to catch up on what I’d missed. Regular readers may find nothing new in the book, but for pop fans and die-hard listeners, the short collection is definitely worth a read.

Life of a Song: The fascinating stories behind 50 of the worlds best-loved songs

The book consists of fifty articles from the regular Life of a Song column collected into book form. Each article takes on a different, well-loved tune from twentieth century popular music. Songs covered include ‘My Way’, ‘Midnight Train to Georgia’, ‘1999’, ‘La Vie en Rose’, and ‘This Land is Your Land’. There are only a few songs in the list that I didn’t know off the top of my head, including ‘Song to the Siren’, and ‘Rocket 88’. The articles usually include some remarks about the songwriter, often quoting them about their creation. Then they cover the journey from composition to hit recording, and usually mention other interpretations that followed the hit.

Each article appears to be less than 1000 words. As you might expect, that’s a lot to cover in that much room. So each article is pretty topical, relating a single anecdote about it, and only touching on the rest. For instance, in the article about ‘Like a Rolling Stone’, the author relates the recording process that shaped the final sound.

On take four of the remake, serendipity strikes. Session guitarist Al Kooper, 21, a friend of the band, walks in holding his guitar, hoping to join in. He is deemed surplus to requirements, but Dylan decides he wants an organ in addition to piano, and Kooper volunteers to fill in. He improvises his part, as he would later recall, ‘like a little kid fumbling in the dark for a light switch’. And suddenly the song turns into the tumbling, cascading version that will become the finished article.

There’s two pieces of information that you need to know about this book in order to enjoy it.

  1. It is a collection of short articles by many contributors.
  2. Those writers are almost entirely arts journalists, rather than trained musicians.

This book was written by a lot of authors. I counted fourteen contributors, each of whom appears to be an English journalist. This can lead to the book feeling somewhat disjointed. Each author is comfortable talking about their own domain of the music industry. Some interpret the lyrics, others relate interviews with creators, others pick up on business maneuvers behind the scenes.

In the introduction, David Chael and Jan Dalley write that the book “is not about singers, or stars, or chart success – although of course they come into the story. It is about the music itself”. If you are a musician, this may leave you expecting musical analysis, lyrical breakdowns, or at least comparisons to similar songs. The book “is about music” in as much as it tells stories about musicians, but it is strictly an outsiders perspective. There’s no illusion that the writers were part of the culture of the song, or involved themselves with the people in the story. A reader shouldn’t expect that in a collection such as this.

My favorite article is the one about ‘Midnight Train to Georgia’. That song has so much soul, that it surprised me to learn that the original title, given to the tune by its white songwriter, was ‘Midnight Plane to Houston’.

The soul singer Cissy Houston… decided to record its first cover version… But the title irked. It wasn’t the collision of Houstons – singer and subject – that bothered her, but one of authenticity. If she was going to sing this song, she had to feel it. And, she later said, ‘My people are originally from Georgia and they didn’t take planes to Houston or anywhere else. They took trains.’

Ultimately, Life of a Song is a great book to read on the way to and from work, or to sit in your book bin next to your favorite chair. It’s a book that can be read in lots of small chunks, and each chunk reveals a little bit more about a song than the recording.

Now if you don’t mind, I need to catch a plane to Houston.

Other Music of 2017

On I did my year-end roundup of the best EDM of 2017. Like any of these lists, it’s a very opinionated list. In this post I want to mention a few other albums I really liked that didn’t make the cut.

There are a few albums that I really enjoyed this year, but weren’t really striking in any way. They were enjoyable to listen to, but they didn’t stand out. Biggest amongst them is probably the new album from Odesza. It’s a fine album, but it nothing in it is unexpected. For me, one of the biggest surprises of the year was the amazing live show Odesza put on to support a relatively mediocre album. Another album that was good, but not great, was Float bySlow Magic. I really enjoyed listening to The Invincible EP by Big Wild, but in the end it just sounds kind of generic.

There are a few other albums that I discovered in 2017, and listened to a lot, but were actually released in previous years. I somehow missed Braincase by Electric Mantis when it was originally released. It’s a dope album of instrumental trap. Also great is the Kindred Spirits EP by Jai Wolf. I managed to see Jai Wolf twice in concert this year because he was coincidentally playing at events I wanted to see. His pure, old-fashioned turntable mastery absolutely dominated festivals crowded by much bigger names, and his EP is worth a few dozen listens.

Last up are the albums that BARELY missed the cut. There were hundreds of great albums released in 2017. Even focusing on EDM exclusively leaves more great albums than I could possibly list. I listened to Full Circle by Oliver probably twenty times. It’s great for getting psyched up for a hard job, or just for getting your hips moving. The new album from Giraffage is also great. It’s an album that really lives up to his previous releases. It feels mellow, sexy, and fun all at the same time. When I saw him live in San Jose, the show was filled with college kids, and that made me feel old. But I think it’s an album that has wide appeal to all listeners to EDM.

And that’s all the other albums from 2017 that I want to talk about.

Okay, obviously that’s not true. I loved so many more albums. LCD Soundsystem made a triumphant comeback. Bonobo released a very solid album earlier in the year. Bonnie and Clyde were the darlings of the internet for a week or two. I thought Sofi Tukker was over-hyped, but then they were terrific live. BVD Kult released a pretty paint-by-numbers pop/EDM track that I absolutely adored. Vitalic released a disappointing album. Jerry Folk released an album that I have very mixed feelings about. Some people from YouTube released some really good music. And someone named Andrew Applepie made a bunch of music that is actually really great in a quirky kind of way.

There was so much great music in 2017. Looking forward to 2018, there is a lot to be pessimistic about. Vladimir Putin seems intent on starting a war. Trump and the Republicans seem intent on destroying democracy in America. But still, it’s going to be a great year for music, and for human culture generally.

DIY Music Branding

Branding, marketing, advertising … they’re necessary evils. Whenever I start a new project I take time to think out the image I want to project. I wish that music could just be music. I wish it could just be sound heard and not seen. But that’s naive. We live in a post-MTV world where music listeners connect their music with a lifestyle, an image, and a brand.

When I started a new beat-driven music project, I was thinking of making trancey electronica under the name of Fynix. So I created a sleek, futuristic logo based on similar designs on groups such as Odesza, and Armin van Buuren.

I’m pretty proud of it.

But after making more music, I have drifted away from trance into a more soul-influenced, keyboard-based style. So the old branding makes no sense.

So how can I connect my image with soul music and soul-inspired electronica? I started by looking at similar acts. I like this art from Faking It by Calvin Harris.

Faking It by Calvin Harris

I also like the branding for Charles Bradley. The image for Spotify Sessions is particularly nice.

Charles Bradley Spotify Sessions

I also want to connect with older soul artists like Sam Cooke and James Brown. Even though my music may not share many qualities with theirs, I want my imagery to put me in the same bucket.

Sam Cooke Wonderful World

After looking at a bunch of images of Sam Cooke, James Brown, Charles Bradley, and some electronica acts, I came up with a few guidelines.

  • Warm colors. James Brown and Sam Cooke use a lot of oranges, browns and reds on their album covers. Using the same color palette will help.
  • Cosmopolitan. Charles Bradley and other funk artists use a lot of images of the city. Maybe this is an artifact of the association with Detroit. Whatever the cause, I’d like to project a cosmopolitan image.
  • Natural Fabrics. Leather jackets are big in the Charles Bradley and James Brown branding. I don’t want a picture of myself in my branding, but if I could connect with something physical that would be good.
  • Sans Serif Fonts. Flat-colored text using basic fonts.

The funk and soul branding also prominently features portraits. Mostly artists looking at or near the camera, with strong lighting. Often the artist is wearing a suit, or a leather jacket. I might do something like this in the future, but I would need a photographer.

For now, I executed my branding guidelines pretty directly. I am from Pittsburgh, so I found a warm-colored picture of Pittsburgh on Wikimedia Commons. Then I wrote “FYNIX” in the middle using a Sans Serif font. Then I applied the newsprint effect to associate my brand with newspapers, which have a real-world physicality that unifies the ideas of the city and natural fabrics.

Here’s the result.

Fynix Branding

I like this because it is legible even when scaled very small, and it looks like the Calvin Harris branding, while subtly calling to mind soul acts of the past.

Sometimes It’s Okay to NOT Write Unit Tests

I recently lost about two-and-a-half days to unit/integration tests. At Mighty Networks, we are pretty proud of our test coverage, and we make writing tests part of the development process. Developers are required to write tests for every feature they implement, but in the past few days I’ve seen that this policy needs to be applied flexibly.

A few months ago we wrote a pretty expansive integration with the iTunes Store. Since we allow people to sell subscriptions through our app, we required a pretty complex integration. Apple’s developer APIs are notoriously crappy, so this required a full team effort. One developer wrote a series of tests for our Apple integration.

In theory, the tests are very thorough. But getting real data for testing is virtually impossible. So the developer faked up a json file, then wrote a preprocessor to generate fake data in a format that looked like Apple’s format. Then he wrote tests.

You may see where this is going already.

The tests he wrote essentially tested his preprocessor. Rather than testing the actual methods used in the integration with Apple, the tests looked at the values generated by the preprocessor. Essentially, by writing a clever object to fake Apple data, he removed the actual integration from the tests.

The tests looked correct. They seemed to show that our Apple code worked. But really they were mostly testing the test code itself.

So when I modified a related system, and added a few tests, I suddenly saw a massive cascade of failures all over the place. The failures were of different types too. Sometimes there was a null value, or an unexpected ID, or an error seemingly from Apple.

It took me a day to figure out that the Apple integration itself wasn’t failing, only the preprocessor wasn’t set up to actually work with the rest of the system. Then I took another day-and-a-half to pull out the worst part of the system and replace the non-tests.

I don’t blame the developer who wrote the tests. It’s a common mistake, and we all did it at least once.

In part I blame Rails, because it encourages black/white thinking about software development.

The developer followed the rule that he needs to write tests for every new feature. When he integrated with Apple, he diligently wrote his tests.

The problem arose when he realized that he couldn’t run production code to get real data. He didn’t know how to write a test for the algorithm, so he wrote code that generates Apple-like data, then tested that.

The developer failed to see that writing tests is a guideline, rather than a rule. In this case, it is very difficult to test every part of the integration. It’s acceptable to write tests of the core process, without testing specific return values and specific pieces of data. The tests gave the impression of working code, and full test coverage. But they hid a few problems with the integration by testing for specific values, rather than algorithmic correctness.

So what can we do?

Senior developers need to encourage junior developers to talk about problems that arise when following “the rules.” Senior developers need to encourage an environment where it’s okay to admit that portion of the code just can’t be tested. Or at least to see that a portion of the code can’t be tested in the same way as most of the code. Senior developers need to encourage critical thinking and analysis in situations where the strict interpretation of the rules may not lead to the best results for the development process.

In which I complain about the lcd soundsystem show…

This feels like a blog post from 2004. I want to complain about some super popular thing as if anyone cares about my opinion. Whatever. I’m going to write it anyway.

I went to see LCD Soundsystem at The Bill Graham last night. The auditorium was packed with a concert audience that actually made me feel young for once. The show was generally excellent, at least in the music sense. It was a great performance. Maybe you could complain that some of the performances were virtually identical to London Sessions. Or maybe you could complain that they only played five tracks off the new album. But that’s picking nits. They ended with All My Friends, so I really can’t complain too much about the music.

LCD Soundsystem at The Bill Graham

And here’s the part where I get up on my soapbox about some nonsense.


Point the fucking lights at the band. No. NO!. Stop your shit. Nobody cares about your art, we just want to see the fucking band. Seriously.

The band was back-lit for 2/3rds of the show. Bright spots were pouring over James Murphy’s shoulders into the audience’s eyes. He looked fabulous in silhouette. At least I think he looked fabulous. It was tough to see him at all. Since the lights were pointed at the fucking audience.


I can’t believe I bought this shirt.

Terrible LCD Soundsystem Shirt

All you had to do was show the picture of James Murphy, and underneath it, write “LCD Soundsystem”. Instead you gave us this monstrosity.

“So Evan, why didn’t you buy the other shirt?”

I did buy it, and it’s a fucking tie-dye.

Terrible Tie Dye LCD Soundsystem T-Shirt

Seriously? SERIOUSLY?!?!? Has anyone who cares about clothing ever actually worn a tie-dyed t-shirt?

Anyway, I know those are pretty minor points. But I was really excited to finally see one of my favorite bands, and these two little things really grated on me.

PS. Yes, there were even more tragic shirt options. In plain white.

Learning to Talk About Inaccuracy for a New Data Engineer

About a month ago, the engineering team at Mighty Networks was impacted by China’s now-defunct one child policy. A parent of our data engineer was having health problems. Because there were no other children to help out, he was forced to relocate his family back to China.

It took us around six months to hire him. With a bunch of data projects now in the pipeline, we couldnt’t go through the hiring process again. Fortunately, I was excited to step into the breach. I’ve never formally trained as a data engineer, but I built a data warehouse from scratch for another startup, and I’ve always had a passion for numbers.

Still, I’ve definitely struggled a little bit in the new role. One of the things I’ve struggled with most is how to communicate numbers to the business team. It’s fine to make pretty visualizations, but how do I communicate the subtlety in the data? How do I communicate the fact that the numbers are as accurate as we can get, but there are still some sources of error ever-present in the system?

I came up with the following guidelines to help me talk to the business team, and I thought they might be useful to other programmers who are in a similar position.

Sources of Error

There are two categorical sources of error in any data analysis system.

  1. Data warehouse replication problems
  2. Bugs and algorithmic errors

Data Replication Issues

Inaccuracy of the first type is unavoidable, and is a universal problem with data warehouses. Data warehouses are typically pulling in huge amounts of data from many sources, then transforming it and analyzing it. In our case, we have jobs that should pull data hourly, but these jobs can fail due to infrastructural errors, such as an inability to requisition server resources from Amazon. So we have jobs that run daily as a fallback mechanism, and we have jobs to pull all the data for each table that can be run manually.

Typically, the data should be no more than an hour off of real time.

When ingestion jobs fail, the data can be recovered by future jobs. Typically, data replication errors do not result in any long-term data loss.

Bugs and Algorithmic Errors

It’s important to remember that the data analysis system is ultimately just software. As with any software project, bugs are inevitable. Bugs can arise in several ways and in several different places.

  1. Instrumentation. The instrumentation can be wrong in many ways. New features may not have been instrumented at all. Instrumentation may be out of date with the assumptions in the latest release. Instrumentation could be conditionally incorrect, leading to omitted data or semi-correct data.
  2. Ingestion. The ingestion occurs in multiple steps. The data has to be correctly propagated from the database, to the replicated database, to the data pipeline, to the data warehouse. Errors in ingestion often occur when only part of this process has been updated. In our case, fields must be added to RedShift, to Kinesis Firehose (for events), to Data Pipeline (for db records), then they must be exposed in Looker.
  3. Transformation and Analysis. The presentation of advanced statistics rests on several layers of analysis and aggregation. A small typo, or mistake in one place can lead to a cascade of errors when that mistake effects a huge amount of data.

How to Talk About Inaccuracy

The best way to talk about inaccuracy is to talk about what steps you have taken to validate the data.

  • What did you do to validate the instrumentation? How did you communicate the requirements and purpose of the new events to the developers? Did you review their pull requests and ensure that the events were actually instrumented?
  • What did you do to validate the ingestion? Did you see events coming in on a staging environment? Did you participate in testing the new feature then verify that your tests percolated through to staging analytics? Did you read the monitoring logs?
  • What did you do to validate the analysis? Did you compare the resulting data to the data in another system? Did you talk through the results with a colleague? Did you double-check the calculations that underlie your charts? Even when they were created by other/former developers? Did you create an intermediate chart and verify the correctness at that level of analysis? When you look at the data from another angle/table, do the results make sense with your new results?

Don’t dwell on the sources of error. Talk about what you have done to minimize the sources of error. In the end, this is software. Software evolves. The first release is always buggy, and we are always working to refine, fix bugs, and improve.

Make a plan to validate each data release like the rest of the team validates the consumer-facing product. Use unit tests, regression tests, and spot=checks with production to validate your process.

Top Line Numbers vs. Other Numbers

In general, you can never be sure that a number is absolutely 100% correct due to the assumptions in the process, and the fact that you must rely on the work of many other developers. Most charts should be used in aggregate to paint a picture of what is happening. No single number should be thought of as absolute. If possible, you should try to present confidence intervals in charts or use other tools that represent the idea of fuzziness.

But as in everything, there are exceptions.

With particularly important numbers, if the amount of data that goes into them is relatively small, then we can manually validate the process by comparing the results with the actual production database. The point is that for the most important, top-line calculations, you should be extremely confident in your process. You should have reviewed each step along the way and ensured to the best of your ability that the number is as close as possible to the real number.


When you’re trying to communicate the accuracy of your data to the business team…

  • Focus on what you have done to validate the numbers.
  • Keep in mind that the data analysis process is software that evolves toward correctness as all software does.
  • Validate data analysis like you validate any other software.
  • Where it’s possible and important, do manual validation against production so you can have high confidence in your top line numbers.