Fourteen Tools for a Productive Distributed Team

A geographically distributed software development team (“distributed team”, for short) is simply one where developers don’t work in close physical proximity (within a few hundred feet). In such a time you interact mostly via electronic means.

To some readers a distributed team will sound like an obviously ridiculous idea, while to others it will sound quite normal. A great many companies are distributed nowadays, as are essentially all open source projects, including the largest and most successful such projects. (Does open source seem like a toy to you? Not “real”? Consider this: in your real project you need large amounts of money to get everyone to work together to build something. Open source projects somehow accomplish this without the carrot and stick of money. Which is more impressive?)

Distributed teams appear to be most common in small firms, but certainly aren’t rare elsewhere. At some large firms, there are teams that work at the same campus or even in the same building, yet are inconveniently spread out enough that they are borderline distributed.

At Oasis Digital we operate mostly as a geographically distributed team, or rather as a set of teams (for various projects). We are scattered around our city, our country, and in a few cases, globally. We’ve learned a lot about how to succeed this way, summarized as 14 tools for a productive distributed team:

(1) Online meeting  / telephone. Effective online meeting tools are available inexpensively or sometimes free; telephone calls, including mobile, long distance, and conference calls are free or inexpensive. Either way, don’t hesitate to spend time together and live discussion, with at least voice and ideally video. Especially the early stages of a project, this is by far the closest match to in-person interaction, of all the tools in this list. At Oasis Digital, we typically use a regular weekly discussions as a baseline, with more discussions of all kinds as needed.  (Frequently at the beginning of a project, less frequently over tim.)

(2) Instant Messaging. Beware that your time is valuable though; don’t type out a long conversation, when it gets complex talk live instead. The key value in IM is as a substitute for the awareness of who is available and working, that would be obvious in the same office.

(3) Email is especially helpful to create a thread, a trail of key decisions. I suggest summarizing key decisions, especially when they span between different teams, in a (hopefully concise) email – even if the decision was reached by meeting, phone, IM, in-person, etc. As above, beware that email can consume unlimited amount of time, and don’t hesitate to  switch to live discussion if it gets complex.

(4) Airplane / Car / Train. The most powerful tool to help a distributed team work well is occasional in-person interaction.

(5) Documents. Communicate a complex idea, especially one that will need to be explained again and again, clearly in a document. (An aside on “template” documents: I’ve found them to be of limited value, though we use them when needed to meet customer requirements.)

(6) Screen Sharing. Using Remote Desktop, VNC, and other tools, it’s possible to do “pair programming” from half a world away. In our projects we do a only a little of this, but a few hours of pair programming on a tricky area can work wonders. Screen sharing is also extremely useful for demonstrating new features, reproducing bugs, working with customers, and much more.

(7) Screencasts are video/audio recordings of the computer screen and a person talking; they are very useful for explaining a feature or module to another developer. Recording then viewing a screencast is not as effective as sitting alongside someone explaining in real-time, but it has the tremendous advantage of re-playability. A series of (high quality) screencasts explaining important parts of a system will get new team members up to speed quickly.

(8) Audio recordings. Record an audio explanation of a feature or bug, while walking around or driving, on an inexpensive hand-held digital voice recorder. Then send it to another developer as is, or have it transcribed. Again, it’s not quite as effective or high-fidelity (of either the audio or the ideas) as an in-person explanation, but can be replayed, sent to multiple recipients, etc. A caveat here is that people read much faster than they listen, so if you have something long-winded (or long-lived, important) to say, have it transcribed then edit it. You may find, as I have, that it’s far easier to compose a long detailed explanation this way, than via direct typing.

(9) Screen Shots. Don’t just tell; show. Show another developer on your team what you mean by taking a screenshot them marking it up.  Unlike a screen recording, it’s easy to go through a set of screenshots and update just one of them in the middle if things change.

(10) Mockups. A distributed team leaves more room for misunderstanding of desired results. Counteract this by building a mockup (paper, Excel, etc.) of what you want.

(11) Issue / Bug Tracking. If you don’t have an issue tracker for your team, stop reading this now and go get one: Mantis, Trac, Jira, there are hundreds to choose from. (Go ahead. I’ll wait.) A common approach for collocated teams is to use a tracking system loosely, updating it occasionally. In this usage model, tracker items act mostly as high level token for more detailed information exchanged in conversations. This approach can be made to work on a distributed team, but in my experience it is unwise. Instead, get every issue in to your tracker and keep it up to date aggressively. Don’t let the sun go down with your issue tracker mismatching reality. Assume that others on your team will check the issue tracker then make important decisions relying on the data therein. If they see something that was fixed 3 days ago, but still marked as broken, bad decisions can easily result. Anyone on the team should be able to see the essential status of every current issue in the tracker, not by a time consuming process of asking someone about each item.

(12) Source Control. You’re already using this, of course. For the future, consider a distributed source control tool (bzr, git, svk, Mercurial, etc.) for your distributed team – if you’ve only used centralized tools (CVS, SVN, ClearCase, etc.) you’ll be surprised how helpful it is to have the project version history available locally, among other benefits. (See Linus explain distributed source control.)

(13) Code Review. Some collocated teams enjoy social camaraderie, but squander the benefits of proximity by working in technical isolation. Your distributed team can outperform them easily: avoid technical isolation. Read each other’s code, comment on it, learn from each other.

(14) Automated Builds. Surely there is hardly anyone left by 2007 who isn’t using an automated, continuous build process on a server somewhere. The value of this is amplified in a distributed team, because the stream of successful builds (we fire off a build after each commit) helps keep everyone in sync.

A related but separate question is where distributed team members work: in their homes, in offices outside their homes, in public spaces, etc. I believe the best answer is a blend of these; I mentioned hiding in a cafe to get some work done before, and I’ll write more about the Where question in a later post.

Another related question is whether any of this is even a good idea; stayed tuned to read more on this as well.

My name is Kyle, and I’m an Infoholic

I recently read Tim Ferriss’s book The Four Hour Work Week, colloquially called 4HWW. The book is short, dense with ideas, and easily worth the $12 price. I recommend the book in spite of:

  • Questions about the veracity of Ferriss’s claimed accomplishments
  • Criticisms that some of his techniques are not as broadly applicable as he makes them sound
  • The fact that the author apparently fell for a bogus chain-letter email and reprinted it on page 284. Ooops – how embarrassing!
  • I’d guess he’s spending more like 80 hours per week promoting his book over the last few months, with many media appearances, interviews, etc.

Among his main points (outsource more, delegate more, sell products rather than services, travel, etc.), the key idea that stood out for me is the “low-information diet”: read less, watch less, surf the web less. This is nothing new of course (I even touched on it myself in an earlier post), but Ferriss makes a compelling case.

Unfortunately, upon self-examination the truth hurts:

  • I read too many books, even though I’ve gotten rid of many recently
  • I read too many magazines.
  • I read too many web sites.
  • I subscribe to too many RSS/Atom feeds.
  • I check email too often.

In my defense, I also somehow write a lot of software, solve many customer problems, and much of the information I consume is at least tangentially related to those sources of value. I read quickly, and I don’t watch television, so this excessive consumption is not as time consuming as it could be.

Still, I need something closer to Ferriss’s low-information diet. I don’t have the guts to go cold turkey, and part of the service we offer to our customers is fast response to problems, so I won’t go as far as he suggests. I will spend less time consuming input and more time producing output.

Update in 2009: This remains an ongoing struggle, but I do quite often manage entire days of producing most of the day and consuming only in short breaks.

Help! My Hierarchy is Slow – Faster Hierarchies with Nested Sets

A great many applications, including many that I’ve worked on, have a hierarchy of things: of parts, of people, of organizations, etc. The way most of us represent such hierarchy is with the first thing that generally comes to mind: make each Widget have a parent Widget, with a table like so:

create table widget (widget_id int, parent_widget_id int, other_fields_here);

This representation is called an “adjacency list”, and is simple and easy. You can readily build a tool to manipulate a hierarchy stored this way. Many off the shelf visual components, for both client-side and web applications, know how to manipulate hierarchies represented this way. Some reporting tools know how to report on hierarchies represented this way.

However, for answering common questions like “who all is under person X in the hierarchy”, the adjacency list approach is unwieldy and slow.

There are various other approaches to representing a hierarchy, most of them discussed in detail in Joe Celko’s articles and books, prominently in the book Trees and Hierarchies in SQL for Smarties. If you work with SQL and hierarchies, buy this book now.

One approach Celko is especially fond of is the “nested set” representation. You can read about it online here and here.

Of course, changing an entire application to use nested sets might be a very big deal in a mature application. That’s OK; in most cases we can get much of the benefit by building a nested-sets “cache” of the share of the hierarchy, with a table like so:

create table widget_hier_cache (widget_id int, left int, right int);

Each time the hierarchy has changed, or before each time we need to run complex queries, delete the rows in this cache table and repopulate them based on the current canonical adjacency data. Celko offers SQL code in his book to do that, which could be translated to work in the stored procedure language of the DBMS at hand. But what about DBMSs that don’t offer stored procedures, such as lightweight local databases (SQLite), MySQL, etc.? The translation must be done in application code instead.

I wrote such code in Delphi a while back, in the process of getting a full understanding of this problem; I’ve cleaned it up and now offer code for download here (DelphiAdjacencyNestedSets.zip), under an open source license (MIT license – use it all you want). I tested this today with Delphi 2007 Win32, but it should work fine at least back to Delphi 7. As far as I can tell with some searching, this is the only Delphi code for translating adjacency to nested sets available on the internet. This code doesn’t know about databases – it is a module to which you feed adjacency data, and from which you get back nested set data. It includes DUnit test cases.

I’ve also put the code on github, for easy browsing and forking.

(Update in August 2007: a new version (DelphiAdjacencyNestedSets2.zip) optionally tolerates “orphan” nodes and forests. Update in January 2008: a newer version (DelphiAdjacencyNestedSets3.zip) propagates an integer value down the hierarchy; it is named Tag as a nod to the .Tag property on VCL components. Both of these newer versions were tested with Delphi 2006/2007 also.)

The essence of the translation is a depth-first traversal of the hierarchy, and of course this can be easily implemented in other languages; the Delphi code is easy to understand, so don’t be afraid to take a look even if you need some Java or C# or PHP etc. I also stumbled across this PHP nested sets implementation, which offers a set of functions to maintain (insert, update, etc.) a hierarchy stored as nested sets, rather than only translate from adjacency to nested sets.

Another useful way to represent a hierarchy for fast querying is with a transitive closure table. I’ll write this up in a future post; it turns out to be especially useful (and necessary) to make arbitrary hierarchies work in the Mondrian OLAP server.

YouTube Scalability Talk

Cuong Do of YouTube / Google recently gave a Google Tech Talk on scalability.

I found it interesting in light of my own comments on YouTube’s 45 TB a while back.

Here are my notes from his talk, a mix of what he said and my commentary:

In the summer of 2006, they grew from 30 million pages per day to 100 million pages per day, in a 4 month period. (Wow! In most organizations, it takes nearly 4 months to pick out, order, install, and set up a few servers.)

YouTube uses Apache for FastCGI serving. (I wonder if things would have been easier for them had they chosen nginx, which is apparently wonderful for FastCGI and less problematic than Lighttpd)

YouTube is coded mostly in Python. Why? “Development speed critical”.

They use psyco, Python -> C compiler, and also C extensions, for performance critical work.

They use Lighttpd for serving the video itself, for a big improvement over Apache.

Each video hosted by a “mini cluster”, which is a set of machine with the same content. This is a simple way to provide headroom (slack), so that a machine can be taken down for maintenance (or can fail) without affecting users. It also provides a form of backup.

The most popular videos are on a CDN (Content Distribution Network) – they use external CDNs and well as Google’s CDN. Requests to their own machines are therefore tail-heavy (in the “Long Tail” sense), because the head codes to the CDN instead.

Because of the tail-heavy load, random disks seeks are especially important (perhaps more important than caching?).

YouTube uses simple, cheap, commodity Hardware. The more expensive the hardware, the more expensive everything else gets (support, etc.). Maintenance is mostly done with rsync, SSH, other simple, common tools.
The fun is not over: Cuong showed a recent email titled “3 days of video storage left”. There is constant work to keep up with the growth.

Thumbnails turn out to be surprisingly hard to serve efficiently. Because there, on average, 4 thumbnails per video and many thumbnails per pages, the overall number of thumbnails per second is enormous. They use a separate group of machines to serve thumbnails, with extensive caching and OS tuning specific to this load.

YouTube was bit by a “too many files in one dir” limit: at one point they could accept no more uploads (!!) because of this. The first fix was the usual one: split the files across many directories, and switch to another file system better suited for many small files.

Cuong joked about “The Windows approach of scaling: restart everything”

Lighttpd turned out to be poor for serving the thumbnails, because its main loop is a bottleneck to load files from disk; they addressed this by modifying Lighttpd to add worker threads to read from disk. This was good but still not good enough, with one thumbnail per file, because the enormous number of files was terribly slow to work with (imagine tarring up many million files).

Their new solution for thumbnails is to use Google’s BigTable, which provides high performance for a large number of rows, fault tolerance, caching, etc. This is a nice (and rare?) example of actual synergy in an acquisition.

YouTube uses MySQL to store metadata. Early on they hit a Linux kernel issue which prioritized the page cache higher than app data, it swapped out the app data, totally overwhelming the system. They recovered from this by removing the swap partition (while live!). This worked.

YouTube uses Memcached.

To scale out the database, they first used MySQL replication. Like everyone else that goes down this path, they eventually reach a point where replicating the writes to all the DBs, uses up all the capacity of the slaves. They also hit a issue with threading and replication, which they worked around with a very clever “cache primer thread” working a second or so ahead of the replication thread, prefetching the data it would need.

As the replicate-one-DB approach faltered, they resorted to various desperate measures, such as splitting the video watching in to a separate set of replicas, intentionally allowing the non-video-serving parts of YouTube to perform badly so as to focus on serving videos.

Their initial MySQL DB server configuration had 10 disks in a RAID10. This does not work very well, because the DB/OS can’t take advantage of the multiple disks in parallel. They moved to a set of RAID1s, appended together. In my experience, this is better, but still not great. An approach that usually works even better is to intentionally split different data on to different RAIDs: for example, a RAID for the OS / application, a RAID for the DB logs, one or more RAIDs for the DB table (uses “tablespaces” to get your #1 busiest table on separate spindles from your #2 busiest table), one or more RAID for index, etc. Big-iron Oracle installation sometimes take this approach to extremes; the same thing can be done with free DBs on free OSs also.

In spite of all these effort, they reached a point where replication of one large DB was no longer able to keep up. Like everyone else, they figured out that the solution database partitioning in to “shards”. This spread reads and writes in to many different databases (on different servers) that are not all running each other’s writes. The result is a large performance boost, better cache locality, etc. YouTube reduced their total DB hardware by 30% in the process.

It is important to divide users across shards by a controllable lookup mechanism, not only by a hash of the username/ID/whatever, so that you can rebalance shards incrementally.

An interesting DMCA issue: YouTube complies with takedown requests; but sometimes the videos are cached way out on the “edge” of the network (their caches, and other people’s caches), so its hard to get a video to disappear globally right away. This sometimes angers content owners.

Early on, YouTube leased their hardware.

Business of Software 2007 Conference

I just registered for the Business of Software 2007 conference this fall (Oct 29 and 30) in San Jose. The organizers have put together an impressive list of speakers, and the schedule is mostly one-track: everyone in the same session. I prefer conference like that, to those with many “tracks”.

This is the first year for this conference, so I have no idea how popular it might or might not be. At “worst” case, it’s a small crowd, which usually results in great conversations.

Update: Brian Button pointed me to this interview with Tim Lister, which includes hints about the salacious topics he’ll talk about at the conference (Tim, co-author of Peopleware, is a headline speaker at the BOS conference).

Pipe RGB data to ffmpeg

A while back I asked on the ffmpeg mailing list how to pipe RGB data in to ffmpeg. I described it as follows:

in my code I am building video frames, 720x480x24bit. I have in mind generating a large number of these, as long as a full DVD worth at 30fps, then using ffmpeg (followed by dvdauthor) to encode them in to MPEG2 for DVD usage.

There were a few replies, but no definitive answer. With considerable experimentation, I got it to work. It turns out that (as far as I can tell) ffmpeg does not have the ability to accept piped in RGB frames. It will however accept piped in data in its “yuv4mpegpipe” format. With some searching and reading I found that this is roughly akin to the format of raw DV video; each frame consists of a header something like this:

YUV4MPEG2 W%d H%d F%d:%d Ip A0:0 C420mpeg2 XYSCSS=420MPEG2

… then an LF character, then data for the the Y, U, and V “planes”. The Y data is full resolution, while the U and Y are half-resolution (this is called “420” in the video world). These planes are uncompressed, one byte per pixel. All of my past work with computer video (going back to Commodore 64s and Apple IIs) has arranged all of the bits for each pixel within a few bytes of each other; this format (with all the Y data for the whole frame, then all the U data, then all the V data) is starkly different.

The essential problem remaining was how to convert RGB to YUV. Happily there are plenty of online references for this. Unhappily there are few fast implementations, and a naive implementation will be very slow. I solved this problem by finding and hiring an expert in low-level data processing with MMX, SSE2, etc. instructions. (I am not in a position to publish that code here.)

In retrospect, though, there are routines included in Intel’s “Integrated Performance Primitives” library which perform this transformation in a highly optimized way. IPP is a bargain: for only a few hundred dollars you get a wealth of high optimized ready-to-use library routines for signal processing.

The ffmpeg piping solution consists, therefore, of:

  1. A module which generated frames in RGB format, to contain whatever contents your application requires.
  2. A module to very quickly convert these to YUV in yuv4mpegpipe format (write your own, or use routines in IPP, for the RGB->YUV420 part).
  3. Pipe this data stream to ffmpeg with stdin; ffmpeg is invoked something like this: ffmpeg -y -f yuv4mpegpipe -i – -i audio.mp3 -target ntsc-dvd -aspect 4:3 foo.mpg

By using a multicore CPU and threads, this whole process can be made to happen in real time or better (i.e., one second of “wall clock” processing time, for one second of finished MPEG2 video). The resulting MPEG2 file can be used with a DVD authoring application to produce a ready-to-burn DVD ISO image.

Update: the data format above is published here as part of the mjpegtools man pages.