Kartlytics: Applying Big Data Analytics to Mario Kart 64

This post also appears on the Joyeur blog.

If you missed it, Joyent recently launched Manta, a web-facing object store with compute as a first-class operation. Manta makes it easy to crunch on Big Data in the cloud, and we’ve seen it used by both ourselves and others and others to solve real business problems involving Big Data. But it’s not just for user behavior and crash dump analysis: Manta has profoundly changed the way we operate here at Joyent’s SF office.

Mario Kart 64

On a typical Friday afternoon at Joyent’s San Francisco office, it’s been a long week and the engineers are getting restless. I glance over at Bill, who mimes working a video game controller with both hands and nods towards the projector. The war dance has begun, and the trash talking will soon follow. Factions form, and soon everyone’s agreed on only one thing: it’s time to play Kart.

Playing Kart
(photo: Joshua Clulow)

Sound familiar? I know we’re not the only office that brings the same intensity to office video games that we bring to our work. Kart is the perfect game for serious office competition because it’s dead simple to learn but takes a long time to master. It’s a standard arcade-style racing game with cartoonish weapons, encouraging the best of friends to abuse and taunt each other mercilessly.

Kart clip

But it’s not just about the competition. In our years playing Kart, we’ve seen a lot of quirky game behavior, and we’ve always wondered: is it true that the game handicaps successful players by giving them less powerful weapons? Given that, how much does the first lap really matter? And which characters are more likely to win: heavyweights or lightweights?

We’ve also had lots of time to discuss strategy: how much does power sliding matter? How much time do you lose for each banana peel slip, shell hit, bomb impact, or fall off the track?

As serious intellectuals often do, we spent hours discussing these questions, what data we would want to collect to answer them, and even how we might go about collecting it. It sounded like a fun project, so I wrote a program that takes video captures of our Mario Kart 64 sessions and picks out when each race starts, which character is in each box on the screen, the rank of each player as the race progresses, and finally when the race finishes. Then I built a web client that lets us upload videos, record who played which character in each race, and browse the aggregated stats. The result is called Kartlytics, and now contains videos of over 230 races from over the last year and change.

This worked pretty well, but as we collected more videos, the time to process them grew prohibitive. Often I’d want to tweak the program to fix detection of, say, Yoshi Valley, but I couldn’t be sure that the change wouldn’t cause it to do the wrong thing on some other track. To test that, I’d have to rerun the process on all the videos, but that took nearly a full day on my laptop.

Enter Manta

This changed significantly about a month ago when Joyent launched Manta. We like to say that we built Manta to solve all kinds of problems from log analysis to video transcoding, but the reality is now clear: Manta exists to compute analytics on Mario Kart 64 sessions. Immediately after we stood up our production Manta service, I loaded all of the videos we’ve recorded to date. If you’ve set up the toolkit, you can list these videos with:

$ mls -l /dap/public/kartlytics/videos | grep -v json
-rwxr-xr-x 1 dap      71246499 Jun 3 18:35 2012-06-19-00.mov
-rwxr-xr-x 1 dap      66418215 Jun 3 17:25 2012-06-19-02.mov
-rwxr-xr-x 1 dap      67641347 Jun 3 18:37 2012-06-19-03.mov

In recent weeks, I’ve converted kartlytics to use Manta for all of the video processing. Here’s the job to process each video, producing a JSON file for each one describing what happened in all races in the video:

$ export KARTLYTICS_OUTDIR=/dap/public/kartlytics/generated
$ mfind -t o -n '.*.mov$' /dap/public/kartlytics/videos |
    mjob create -w
    -s /dap/public/kartlytics/kartlytics.tgz
    -s /dap/public/kartlytics/bin/video-transcribe
    --init "cd /var/tmp && tar xzf /assets/dap/public/kartlytics/kartlytics.tgz"
    -m "/assets/dap/public/kartlytics/bin/video-transcribe /var/tmp/kartlytics $KARTLYTICS_OUTDIR "'$MANTA_INPUT_FILE'

This job uses a common pattern for Manta jobs, which is to throw an existing program (kartlytics) into a tarball in Manta, specify it as an asset for a job, and run the bundled program inside the job. In this case, the kartlytics.tgz tarball contains a built copy of the kartlytics github repo, which contains the C program and support files for doing the video processing. There’s a separate job for creating this tarball, which is just:

$ export KARTLYTICS_TARBALL=/dap/public/kartlytics/kartlytics.tgz
$ mjob create -w
    -s /dap/public/kartlytics/bin/make-tarball
    -r "/assets/dap/public/kartlytics/bin/make-tarball $KARTLYTICS_TARBALL" < /dev/null

The “make-tarball” script just clones the repo, builds it, and saves the tarball back to Manta.

Processing all these videos used to take nearly a day on my laptop. It takes just 5 minutes on Manta, and I don’t have to think about spinning up compute instances and then spinning them down when it’s done. This makes a serious difference because I can get feedback on algorithm tweaks in minutes rather than overnight.

There are two other Manta jobs involved: one that takes the race transcripts produced by the first job and aggregates them with the human-entered data about who played which characters in each race, producing a single aggregated JSON file; and one that takes the same race transcripts and produces smaller web-quality videos for each race that you can play directly on the kartlytics.com website. For details, see the the run_all script in the repo.

Try it: Since these videos are public, you can run kartlytics yourself on our collection of videos. You can run these two jobs exactly as written, replacing KARTLYTICS_OUTDIR and KARTLYTICS_TARBALL with paths in your own storage areas (e.g., /$MANTA_USER/public/kartlytics). The above “run_all” script is also parametrized so that anyone can run it on our videos.


The front page of kartlytics.com shows some summary information, including the tracks we play a lot:


We start every session with Luigi Raceway, which is why that track has the most plays. After the first race, the winner chooses the next track, so the distribution diverges after that.

Next on the page we show races from the latest session, then the “wildest races”. This is a proxy for “slugfests”, which are races with just the right circumstances (wide open track, a tight pack, lucky weapons, etc.) for a weapons-based free-for-all. The “wildest” ones listed here are those with the most number of player position changes per minute.

Wildest races

Below that, we have an example of a hotly-debated question: how often does a player go from 1st to 4th in less than 5 seconds? We call that a “Keithing”, and the front page shows all of them. You can sort by person to see how many times it’s happened to each player (but keep in mind that not all players have played the same number of races).

You can click on the date/time of any race to get to the race details page, which summarizes the results, shows a playable video of just that race, and then the all important race transcript, showing not only everything that happened in the race, but with screenshots to prove it:

Event summary

That’s really important, both for debugging and to silence criticism about kartlytics “making stuff up”.

You can also click any player to see that player’s races, how many times they’ve played each track, and which characters (and character classes — lightweight, middleweight, or heavyweight) they tend to play:

Player details


We’ve really just scratched the surface. We’ve only answered a few of the questions I mentioned above. We’ve discussed tracking lots more data, like power slides (to see how much they matter), weapons (to see how the distribution changes with rank), and various mistakes (to prove [or disprove] claims like “I keep running into banana peels today”).

We’d also love to add videos from you! It’s surprisingly easy. To record video, I purchased an iGrabber and a small power amplifier (so we could split the video signal without losing too much quality — you may not need this if you’re playing on a TV instead of a projector). The iGrabber comes with software to record the video, and kartlytics does the rest. As I showed above, you can run the software on your own jobs.

The algorithm behind Kartlytics is really pretty simplistic (see the README in the repo for details), but it can probably be applied to other cult classics that have relatively simple on-screen game status. I know our Seattle office is big on Street Fighter II; surely Street Fighter-lytics can’t be far away?

Fault tolerance in Manta

Since launching Manta last week, we’ve seen a lot of excitement. Thoughtful readers quickly got to burning questions about the system’s fault tolerance: what happens when backend Manta components go down? In this post, I’ll discuss fault tolerance in Manta in more detail. If anything seems left out, please ask: it’s either an oversight or just seemed too arcane to be interesting.

This is an engineering internals post for people interested in learning how the system is built. If you just want to understand the availability and durability of objects stored in Manta, the simpler answer is that the system is highly available (i.e., service survives system and availability zone outages) and durable (i.e., data survives multiple system and component failures).

First principles

First, Manta is strongly consistent. If you PUT an object, a subsequent GET for the same path will immediately return the object you just PUT. In terms of CAP, that means Manta sacrifices availability for consistency in the face of a complete network partition. We feel this is the right choice for an object store: while other object stores remain available in a technical sense, they can emit 404s or stale data both when partitioned and in steady-state. The client code to deal with eventual consistency is at least as complex as dealing with unavailability, except that there’s no way to distinguish client error from server inconsistency — both show up as 404 or stale data. We feel transparency about the state of the system is more valuable here. If you get a 404, that’s because the object’s not there. If the system’s partitioned and we cannot be sure of the current state, you’ll get a 503, indicating clearly that the service is not available and you should probably retry your request later. Moreover, if desired, it’s possible to build an eventually consistent caching layer on top of Manta’s semantics for increased availability.

While CAP tells us that the system cannot be both available and consistent in the face of an extreme network event, that doesn’t mean the system fails in the face of minor failures. Manta is currently deployed in three availability zones in a single region (us-east), and it’s designed to survive any single inter-AZ partition or a complete AZ loss. As expected, availability zones in the region share no physical components (including power) except for being physically close to one another and connected by a high-bandwidth, low-latency interconnect.

Like most complex distributed systems, Manta is itself made up of several different internal services. The only public-facing services are the loadbalancers, which proxy directly to the API services. The API services are clients of other internal services, many of which make use of still other internal services, and so on.

Stateless services

Most of these services are easy to reason about because they’re stateless. These include the frontend loadbalancers, the API servers, authentication caches, job supervisors, and several others.

For each stateless service, we deploy multiple instances in each AZ, each instance on a different physical server. Incoming requests are distributed across the healthy instances using internal DNS with aggressive TTLs. The most common failure here is a software crash. SMF restarts the service, which picks up where it left off.

For the internal services, more significant failures (like a local network partition, power loss, or kernel panic) result in the DNS record expiring and the instance being taken out of service.

Stateful services

Statelessness just pushes the problem around: there must be some service somewhere that ultimately stores the state of the system. In Manta, that lives in two tiers:

  • the storage tier, which stores the contents of users’ objects
  • the metadata tier, which maps user object names to the servers where the data is stored

The storage tier

The availability and durability characteristics of an object are determined in part by its “durability level“. From an API perspective, this indicates the number of independent copies of the object that you want Manta to store. You pay for each copy, and the default value is 2. Copies are always stored on separate physical servers, and the first 3 copies are stored in separate availability zones.

Durability of a single copy: Objects in the storage tier are stored on raidz2 storage pools with two hot spares. The machine has to sustain at least three concurrent disk failures before losing any data, and could survive as many as eight. The use of hot spares ensures that the system can begin resilvering data from failed disks onto healthy ones immediately, in order to reduce the likelihood of a second or third concurrent failure. Keith discusses our hardware choices in more depth on his blog.

Object durability: Because of the above, it’s very unlikely for even a single copy of an object to be lost as a result of storage node failure. If the durability level is greater than 1 (recall that it’s 2 by default), all copies would have to be lost for the object’s data to be lost.

Object availability: When making a request for an object, Manta selects one of the copies and tries to fetch the object from the corresponding storage node. If that node is unavailable, Manta tries another node that has a copy of the object, and it continues doing this until it either finds an available copy or runs out of copies to try. As a result, the object is available as long as the frontend can reach any storage node hosting a copy of the object. As described above, any storage node failure (transient or otherwise) or AZ loss would not affect object availability for objects with at least two copies, though such failures may increase latency as Manta tries to find available copies. Similarly, in the event of an AZ partition, the partitioned AZ’s loadbalancers would be removed from DNS, and the other AZs would be able to service requests for all objects with at least two copies.

Since it’s much more likely for a single storage node to be temporarily unavailable than for data to be lost, it may be more useful to think of “durability level” as “availability level”. (This property also impacts the amount of concurrency you can get for an object — see Compute Jobs below.)

Metadata tier

The metadata tier records the physical storage nodes where each object is stored. The object namespace is partitioned into several completely independent shards, each of which is designed to survive the usual failure modes (individual component failure, AZ loss, and single-AZ partition).

Each shard is backed by a postgres database using postgres-based replication from the master to both a synchronous slave and an asynchronous standby. Each database instance within the shard (master, sync slave, and async slave) is located in a different AZ, and we use Zookeeper for election of the master.

The shard requires only one peer to be available for read availability, and requires both master and synchronous slave for write availability. Individual failures (or partitions) of the master or synchronous slave can result in transient outages as the system elects a new leader.

The mechanics of this component are complex and interesting (as in, we learned a lot of lessons in building this). Look for a subsequent blog post from the team about the metadata tier.

Compute Jobs

Manta’s compute engine is built on top of the same metadata and storage tiers. Like the other supporting services, the services are effectively stateless and the real state lives in the metadata tier. It’s subject to the availability characteristics of the metadata tier, but it retries internal operations as needed to survive the transient outages described above.

If a given storage node becomes unavailable when there are tasks running on it, those tasks will be retried on a node storing another copy of the object. (That’s one source of the “retries” counter you see in “mjob get”.) Manta makes sure that the results of only one of these instances of the task are actually used.

The durability level of an object affects not only its availability for compute (for the same reasons as it affects availability over HTTP as described above), but also the amount of concurrency you can get on an object. That’s because Manta schedules compute tasks to operate on a random copy of an object. All things being equal, if you have two copies of an object instead of one, you can have twice as many tasks operating on the object concurrently (on twice as many physical systems).

Final notes

You’ll notice that I’ve mainly been talking about transient failures, either of software, hardware, or the network. The only non-transient failure in the system is loss of a ZFS storage pool; any other failure mode is recoverable by replacing the affected components. Objects with durability of at least two would be recoverable in the event of pool loss from the other copies, while objects with durability of one that were stored on that pool would be lost. (But remember: storage pool loss as a result of any normal hardware failure, even multiple failures, is unlikely.)

I also didn’t mention anything about replication in the storage tier. That’s because there is none. When you PUT a new object, we dynamically select the storage nodes that will store that object and then funnel the incoming data stream to both nodes synchronously (or more, if durability level is greater than 2). If we lose a ZFS storage pool, we would have to replicate objects to other pools, but that’s not something that’s triggered automatically in response to failure since it’s not appropriate for most failures.

Whether in the physical world or the cloud, infrastructure services have to be highly available. We’re very up front about how Manta works, the design tradeoffs we made, and how it’s designed to survive software failure, hardware component failure, physical server failure, AZ loss, and network partitions. With a three AZ model, if all three AZs became partitioned, the system chooses strong consistency over availability, which we believe provides a significantly more useful programming model than the alternatives.

For more detail on how we think about building distributed systems, see Mark Cavage’s ACM Queue article “There’s Just No Getting Around It: You’re Building a Distributed System.”

Inside Manta: Distributing the Unix shell

Today, Joyent has launched Manta: our internet-facing object store with compute as a first class operation. This is the culmination of over a year’s effort on the part of the whole engineering team, and I’m personally really excited to be able to share this with the world. There’s plenty of documentation on how to use Manta, so in this post I want to talk about the guts of my favorite part: the compute engine.

The super-quick crash course on Manta: it’s an object store, which means you can use HTTP PUT/GET/DELETE to store arbitrary byte streams called objects. This is similar to other HTTP-based object stores, with a few notable additions: Unix-like directory semantics, strong read-after-write consistency, and (most significantly) a Unixy compute engine.

Computation in Manta

There’s a terrific Getting Started tutorial already, so I’m going to jump straight to a non-trivial job and explain how it runs under the hood.

At the most basic level, Manta’s compute engine runs arbitrary shell commands on objects in the object store. Here’s my example job:

$ mfind -t o /dap/stor/snpp | mjob create -qom 'grep poochy'

This job enumerates all the objects under /dap/stor/snpp (using the mfind client tool, analogous to Unix find(1)), then creates a job that runs “grep poochy” on each one, waits for the job to finish, and prints the outputs.

I can run this one-liner from my laptop to search thousands of objects for the word “poochy”. Instead of downloading each file from the object store, running “grep” on it, and saving the result back, Manta just runs “grep poochy” inside the object store. The data never gets copied.

Notice that our Manta invocation of “grep” didn’t specify a filename at all. This works because Manta redirects stdin from an object, and grep reads input from stdin if you don’t give it a filename. (There’s no funny business about tacking the filename on to the end of the shell script, as though you’d run ‘grep poochy FILENAME’, though you can do that if you want using environment variables.) This model extends naturally to cover “reduce” operations, where you may want to aggregate over enormous data streams that don’t fit on a single system’s disks.

One command, many tasks

What does it actually mean to run grep on 100 objects? Do you get one output or 100? What if some of these commands succeed, but others fail?

In keeping with the Unix tradition, Manta aims for simple abstractions that can be combined to support more sophisticated use cases. In the example above, Manta does the obvious thing: if the directory has 100 objects, it runs 100 separate invocations of “grep”, each producing its own output object, and each with its own success or failure status. Unlike with a single shell command, a one-phase map job can have any number of inputs, outputs, and errors. You can build more sophisticated pipelines that combine output from multiple phases, but that’s beyond the scope of this post.1

How does it work?

Manta’s compute engine hinges on three SmartOS (illumos) technologies:

  • Zones: OS-based virtualization, which allows us to run thousands of these user tasks concurrently in lightweight, strongly isolated environments. Each user’s program runs as root in its own zone, and can do whatever it wants there, but processes in the zone have no visibility into other zones or the rest of the system.
  • ZFS: ZFS’s copy-on-write semantics and built-in snapshots allow us to completely reset zones between users. Your program can scribble all over the filesystem, and when it’s done we roll it back to its pristine state for the next user. (We also leverage ZFS clones for the filesystem image: we have one image with tens of gigabytes of software installed, and each zone’s filesystem is a clone of that single image, for almost zero disk space overhead per zone.)
  • hyprlofs: a filesystem we developed specifically for Manta, hyprlofs allows us to mount read-only copies of files from one filesystem into another. The difference between hyprlofs and traditional lofs is that hyprlofs supports commands to map and unmap files on-demand, and those files can be backed by arbitrary other filesystems. More on this below.

In a nutshell: each copy of a Manta object is stored as a flat file in ZFS. On the same physical servers where these files are stored, there are a bunch of compute zones for running jobs.

As you submit the names of input objects, Manta locates the storage servers containing a copy of each object and dispatches tasks to one server for each object. That server uses hyprlofs to map a read-only copy of the object into one of the compute zones. Then it runs the user’s program in that zone and uses zfs rollback to reset the zone for the next tenant. (There’s obviously a lot more to making this system scale and survive component failures, but that’s the basic idea.)

What’s next?

In this post I’ve explained the basics of how Manta’s compute engine works under the hood, but this is a very simple example. Manta supports more sophisticated distributed computation, including reducers (including multiple reducers) and multi-phase jobs (e.g., map-map-map).

Because Manta uses the Unix shell as the primitive abstraction for computation, it’s very often trivial to turn common shell invocations that you usually run sequentially on a few files at a time into Manta jobs that run in parallel on thousands of objects. For tasks beyond the complexity a shell script, you can always execute a program in some other language — that’s, after all, what the shell does best. We’ve used it for applications ranging from converting image files to generating aggregated reports on activity logs. (In fact, we use the jobs facility internally to implement metering, garbage collection of unreferenced objects, and our own reports.) My colleague Bill has already used it to analyze transit times on SF Muni. Be on the lookout for a rewrite of kartlytics based on Manta.

We’re really excited about Manta, and we’re looking forward to seeing what people do with it!

1 Manta’s “map” is like the traditional functioning programming primitive that performs a transformation on each of its inputs. This is similar but not the same as the Map in MapReduce environments, which specifically operates on key-value pairs. You can do MapReduce with Manta by having your program parse key-value pairs from objects and emit key-value pairs as output, but you can also do other transformations that aren’t particularly well-suited to key-value representation (like video transcoding, for example).