Part one of this pair explored the original algorithm the WiredTiger write-ahead log used to consolidate writes in order to minimize IO. It used atomic compare-and-swap operations in two phases to accomplish this without time-consuming locking. This algorithm worked extremely well as long as there were no more than a few threads running per core. But its reliance on busy-waiting to avoid locking caused a logjam when the number of threads increased beyond that limit -- a serious problem given that many MongoDB workloads would have a large number of threads per core. This issue was blocking MongoDB’s goal of making WiredTiger the default storage engine in v3.2.
This story has a happy ending thanks to my colleague, Senior Technical Service Engineer Bruce Lucas. Bruce had initially uncovered the logjam and reported it to me; together, we overcame it without compromising any other workloads. Because Bruce’s mindset was not colored by the legacy of the original approach, he was able to provide the critical insight that paved the way for the solution, allowing WiredTiger to become the default storage engine in v3.2.
Code can't be optimized; it can only be optimized for a set of conditions. When conditions change, optimizations can become bottlenecks, and when that happens, a thorough audit of assumptions might well hold the key to the solution.
The WiredTiger write-ahead log exemplifies this principle. It’s a critical codepath within a high-performance storage engine, and I have optimized it heavily to avoid I/O and locking. But some of the conditions I had initially targeted became invalid when WiredTiger became a storage engine in MongoDB. When a colleague of mine investigated a case of negative scaling found during testing, he uncovered a serious bottleneck in the write-ahead log… call it a "logjam". That investigation ultimately led us to rethink our assumptions and optimize for new conditions. We validated the new approach with a quick prototype, and then covered all the intricacies and edge cases to produce a fully realized solution.
In part one of this two-part series, I’ll dive deep into the innards of the WiredTiger write-ahead log. I’ll show how it orchestrates many threads writing to a single buffer without locking, and I’ll explain how two conflicts between that design and the new conditions produced the logjam. Part two will focus on how we eliminated the bottleneck. I’ll analyze its root causes, describe the key insight that enabled our solution, and detail the new algorithm and how it reflects our current conditions.
On January 10, I released a badly broken version of the MongoDB C Driver, libmongoc 1.5.2. For most users, that version could not connect to a server at all! Luckily, in under 24 hours a developer reported the bug, I reverted the mistake and released a fix. Although it was resolved before it did any damage, this is among the most dramatic mistakes I've made since I switched from the PyMongo team to libmongoc almost two years ago. My error stemmed from three mistaken assumptions I've had ever since I changed projects. What were they?
Here's how the story began. In December, a libmongoc user named Alexey pointed out a longstanding limitation: it would only resolve hostnames to IPv4 addresses. Even if IPv6 address records existed for a hostname, the driver would not look them up -- when it called getaddrinfo on the hostname to do the DNS resolution, it passed AF_INET as the address family, precluding anything but IPv4. So if you passed the URI mongodb://example.com, libmongoc resolved "example.com" to an IPv4 address like 220.127.116.11 and tried to connect to it. If the connection timed out, the driver gave up.
What do you do with a third-party tool that proves your application lacks a feature? Add that tool to your continuous integration system (after adding the feature, of course)! In our case we have added linearizable reads to MongoDB 3.4 and use Jepsen to test it.
What is Linearizability?
Linearizability is a property of distributed systems first introduced by Herlihy & Wing in their July 1990 article "Linearizability: a correctness condition for concurrent objects" (ACM Transactions on Programming Languages and Systems Journal). Peter Bailis probably provides the most accessible explanation of linearizability: "writes should appear to be instantaneous. Imprecisely, once a write completes, all later reads (where “later” is defined by wall-clock start time) should return the value of that write or the value of a later write. Once a read returns a particular value, all later reads should return that value or the value of a later write."
Recently we published a piece by A. Jesse Jiryu Davis about his undertaking to prove that getaddrinfo was thread-safe on OS X, thus enabling Python to do away with an unnecessary and troublesome mutex around hostname resolution. The convolutions of tracking down that evidence, and the shroud of secrecy involved in all correspondence with Apple, inspired us to render the piece in a whimsical, high-fantasy style. It was called "The Saga of Concurrent DNS in Python, and the Defeat of the Wicked Mutex Troll"
It seems the unconventional style has inspired a couple of dramatic readings, which we’re just thrilled about. We’d love to share them with you.
That challenge was taken up by Mason Egger, who created his YouTube channel BSD Synergy to serve those not yet ready for the firehose of insider expertise that is BSD Now. He was so tickled with what he heard on BSD Now that he read the entire thing for BSD Synergy episode 20, against a backdrop of images that set the scene beautifully.
Allan, Mason, thank you both for bringing our work to life!
We cannot confirm any rumors that we are in talks with Peter Jackson to do a motion picture adaptation of this story.
A suggestion we received both in this discussion and on lobste.rs was to use canvas to render the data points. We had originally avoided canvas because of time constraints, lack of team familiarity with canvas, and the complications it introduced with regards to mouse interactions. However, Noah proposed a combination of SVG and canvas that strikes a balance between canvas' performance and SVG's convenience, complete with a demo. It piqued my interest, and so I decided to explore it in some more detail here.
Tell us about the time you made DNS resolution concurrent in Python on Mac and BSD.
No, no, you do not want to hear that story, my friends. It is nothing but old lore and #ifdefs.
But you made Python more scalable. The saga of Steve Jobs was sung to you by a mysterious wizard with a fanciful nickname! Tell us!
Gather round, then. I will tell you how I unearthed a lost secret, unbound Python from old shackles, and banished an ancient and horrible Mutex Troll.
Let us begin at the beginning...
A long time ago, in the 1980s, a coven of Berkeley sorcerers crafted an operating system. They named it after themselves: the Berkeley Software Distribution, or BSD. For generations they nurtured it, growing it and adding features. One night, they conjured a powerful function that could resolve hostnames to IPv4 or IPv6 addresses. It was called getaddrinfo. The function was mighty, but in years to come it would grow dangerous, for the sorcerers had not made getaddrinfo thread-safe.
As ages passed, BSD spawned many offspring. There were FreeBSD, OpenBSD, NetBSD, and in time, Mac OS X. Each made its copy of getaddrinfo thread safe, at different times and different ways. Some operating systems retained scribes who recorded these events in the annals. Some did not.
Because getaddrinfo is ringed round with mystery, the artisans who make cross-platform network libraries have mistrusted it. Is it thread safe or not? Often, they hired a Mutex Troll to stand guard and prevent more than one thread from using getaddrinfo concurrently. The most widespread such library is Python's own socket module, distributed with Python's standard library. On Mac and other BSDs, the Python interpreter hires a Mutex Troll, who demands that each Python thread hold a special lock while calling getaddrinfo.
Fuzz testing is a method for subjecting a codebase to a tide of hostile input to supplement the test cases engineers create on their own. In part one of this pair, we looked at the hybrid nature of our fuzzer -- how it combines “smart” and “dumb” fuzzing to produce input random enough to provoke bugs, but structured enough to pass input validation and test meaningful codepaths. To wrap up, I’ll discuss how we isolate signal from the noise a fuzzer intrinsically produces, and the tooling that augments the root cause analyses we do when the fuzzer finds an issue.
In part one of two, we examine how our fuzzer hybridizes the two main types of fuzzing to achieve greater coverage than either method alone could accomplish. Part two will focus on the pragmatics of running the fuzzer in a production setting and distilling a root cause from the complex output fuzz tests often produce.
What's a fuzzer?
Fuzzing, or fuzz testing, is a technique of generating randomized, unexpected, and invalid inputs to a program to trigger untested code paths. Fuzzing was originally developed in the 1980s and has since proven to be effective at ensuring the stability of a wide range of systems, from filesystems to distributed clusters to browsers. As people attempt to make fuzzing more effective, two philosophies have emerged: smart, and dumb fuzzing. And as the state of the art evolves, the techniques that are used to implement fuzzers are being partitioned into categories, chief among them being “generational” and “mutational.” In many popular fuzzing tools, smart fuzzing corresponds to generational techniques, and dumb fuzzing to mutational techniques, but as we will see, this is not an intrinsic relationship. Indeed, in our case, the situation is precisely reversed.
Until last year, Jeremy Mellema was a history teacher. Now, he's teaching computer programming. When I visited his class in the Bronx this month, he had 30 students with 30 MacBooks, completing exercises in Python. They had just finished a lesson on data types, and now they were tackling variables. In Jeremy's class, the first variable assignment is:
tupac = "Greatest of All Time!!"
Computer Science for All
A year ago, New York City mayor Bill de Blasio announced Computer Science for All, an $80 million public-private partnership. The goal is to teach computer science to every student at every public school. But first, the schools need curricula and 5000 teachers need training.
Here at MongoDB, our VP of Education Shannon Bradshaw oversees MongoDB University, which trains IT professionals to use MongoDB. When he heard about CS4All, he wanted us to contribute. He proposed that we set aside budget for two paid fellowships, and recruit public school teachers to spend the summer with us. We would develop them as teachers, and help build curricula they could take back into schools this fall. MongoDB staff would share our expertise, our office space, our equipment, and the MongoDB software itself.
Shannon pitched his proposal to the company like this: "As many of us know, it’s still unusual for students to encounter computer science, let alone databases, in their classrooms before entering college. I believe this absence directly contributes to the gender and racial disparity we see today across our industry." The CS4All project improves access to these subjects for many more students in our city, and MongoDB could be part of it from the beginning.