Tango is a stand-alone, RESTful service that Autolab uses as the back-end for autograding. Tango receives grading jobs from Autolab’s front-end, adds them to a job queue, assigns them to available containers for grading and shepherds the job through the process. In its early days, Tango was mostly used for jobs that ran in under 5 seconds. However, over the recent semesters, Autolab has grown to host classes like Distributed Systems, Machine Learning and Storage Systems–classes with significantly higher compute requirements. As we looked into how to handle larger loads, we were running into the problem of how to manage queued jobs and distribute them to different instances in our back-end. To this end, we decided to take the initial step for turning Tango into a distributed system by implementing a persistent memory using Redis and using the producer consumer model.

Tango's initial Architecture Initial Tango architecture with in-memory job queue

The existing monolithic model with a single process and an in-memory queue was posing multiple problems:

  • Running a web server that handled large file uploads concurrently along with a job manager that dispatched jobs from the job queue to new threads made the process prone to crashes.
  • Crashes caused tall the jobs stored in the in-memory queue to disappear.
  • After a crash, all the containers running a job would suddenly be left in a so-called limbo state because they are no longer managed by any Tango process.

Therefore we decided to improve our architecture by switching to multi-process model with a persistent queue.

Persistent Memory Model using Redis

The in-memory job queue was the barrier to making the system robust. We want to store jobs on an independent system and make it persistent.

Tango with Persistent Memory Model Tango with Persistent Memory Model

In this persistent memory model, even if Tango restarts it will keep the same job queue. We could even spin up multiple instances of Tango and they can all work concurrently, sharing the same queue (with a couple problems which I will mention in the next section).

Choice of Redis

We started looking into possible solutions: traditional relational databases like MySQL, document-oriented databases such as MongoDB, messaging queues such as RabbitMQ and key-value stores such as Redis. Looking at the amount of data and the level of nestedness, we decided that we don’t need a database with full-range of functionality and robustness. We needed something fast and simple to use. Our old job queue was just a simple Python dictionary; therefore, a fast key-value store like Redis seemed like the right fit. Setting it up and playing around with it was a breeze, therefore we decided to go with this option.

Implementation

While making the switch, we wanted to keep the ability to run Tango without a dependency on the external service in order to keep backwards compatibility and simplify setting up a local development environment.

In order to achieve this we created (pseudo)abstract classes that include the common methods and decide on what is used under the hood based on the configuration.

Since Python doesn’t really have abstract classes, we have define methods that would initiate the appropriate class and return it.

# This is an abstract class that decides on 
# if we should initiate a TangoRemoteDictionary or TangoNativeDictionary
# Since there are no abstract classes in Python, we use a simple method
def TangoDictionary(object_name):
    if Config.USE_REDIS:
        return TangoRemoteDictionary(object_name)
    else:
        return TangoNativeDictionary()

See the full code here.

In the future, we can have other implementations, such as MongoDB, and just add it as an option here. For example:

def TangoDictionary(object_name):
    if Config.USE_REDIS:
        return TangoRemoteDictionary(object_name)
    elif Config.USE_MONGO:
        return TangoMongoDictionary(object_name)
    else:
        return TangoNativeDictionary()

Serialization and Marshalling

Redis, being a key-value store, only accepts strings as values. Even though the objects we are trying to store are quite simple and flat, we do not want to create a new key-value pair for every attribute of the class. Instead we decided to serialize what we call the ‘remote objects’ using the pickle module included in the standard python library.

A sample method to put the object into the store looks like

def set(self, id, obj):
    pickled_obj = pickle.dumps(obj)
    self.r.hset(self.hash_name, str(id), pickled_obj)
    return str(id)   

To get it back you would need:

def get(self, id):
    unpickled_obj = self.r.hget(self.hash_name, str(id))
    obj = pickle.loads(unpickled_obj)
    return obj

Pickling seemed like a good solution at first, but when we needed to nest remote objects, e.g. put TangoMachine*s into the remote *TangoQueue, we ran into the problem of self.r(type of redis.StrictRedis) not being serializable. To get around this, we defined custom getstate and setstate methods that would not include self.r attribute in the serialized string, but set it back when the object is deserialized.

def __setstate__(self, dict):
    self.__db= getRedisConnection()
    self.__dict__.update(dict)

Details of the implementation can be found on TangoObjects module.

Producer Consumer Model

So far we have covered how Tango receives grading jobs and stores them. The second most important part of the process is to read from the queue and distribute the jobs to the available containers(cluster of Virtual Machine or Docker). In the initial architecture, the consumer was part of the HTTP server process. A single Tango process would both produce and consume the jobs.

Having a shared queue gave us the ability to run multiple producers as explained in the previous section. However, we only needed one consumer to distribute the jobs.

Therefore, we decided to separate the consumer from the HTTP server (producer) as a standalone process.

Tango with Prod/Com Model

Tango with Prod/Com processes and Persistent Memory Model

The advantage of this architecture is that we can launch an arbitrary number of HTTP servers as we receive more load. There would be only one consumer process which is responsible for reading from the queue and assigning the jobs. This process gives us a lot of flexibility because it can be run on any node, as well as be stopped and migrated to a different machine at any time.


We are currently testing the new architecture and will be posting more about the outcomes as it gets used by more users and we come across interesting cases.

Please feel free to ask questions in the comments below!