Feb 20, 2018

Operationalising Data Science

0 comments

 

Operationalising Data Science is something of a difficult question for us here at Elastacloud, given the wealth of options available (AzureML, Spark, python web service etc). Recently as part of Renewables we’ve been updating our model to the latest and greatest and wanted to move away from using AzureML, which is notoriously bad at scaling. We are investigating a number of long term solutions to this problem, which will most likely involve DataBricks (spark) and ML Pipelines. In the short term we wanted to prove out a solution using Azure Machine Learning Model Management. This provides capabilities such as;

 

  • Model versioning

  • Tracking models in production

  • Deploying models to production through AzureML Compute Environment with Azure Container Service and Kubernetes

  • Creating Docker containers with the models and testing them locally

  • Automated model retraining

  • Capturing model telemetry for actionable insights.

 

Once the appropriate artifacts have been created it’s a small set of simple steps to deploy that given model as a web service. One big advantage is the ability to scale the number of nodes in the container service, giving us significantly more control over Azure ML. We can always deploy the derived docker container into other environments should we choose.

 

However, I digress, the purpose of this post is primarily around how we dealt with integrating this into our existing application. Originally, the AzureML model was invoked directly from an Azure Stream Analytics (ASA) instance (via a UDF). This was not possible using Azure Model Management. Instead, we restricted our ASA to emit the data for which we wanted to make predictions to an Event Hub. We wrote an Azure Function with an Event Hub Trigger (yes, we could call this directly from ASA, but that tightly couples us and inhibits our ability to hive off this stream for other purposes). This Azure Function emits to another Event Hub from which an ASA job reads and write the predictions to a database.

 

 

One problem that did keep cropping up during testing was a “Too many requests for service irrmodel (overloaded)” error. This as you can imagine was somewhat unexpected given we have a service that should be responsive to whatever we want to throw at it. It took us a while to track down and understand, but, long story short, the Azure Function was pulling messages off the event hub at too high a rate. Given by default it will attempt to process 64 messages at a time and given ASA batches messages it writes to event hub (~600 messages packed into a single event hub message) we were in effect trying to process ~38400 messages at a time. We reduced this by overriding the defaults in host.json and coupled with batching requests in the Azure Function the error was avoided. The problem now is to figure out the optimal values for message throughput and parallel invocation of the deployed model.

 

 

New Posts
  • Something that comes up quite frequently when people start using Spark is "How can I filter my DataFrame using the contents of another DataFrame?". People with SQL experience will immediately look to trying to replicate the following. SELECT * FROM table_a a WHERE EXISTS (SELECT * FROM table_b b WHERE b.Id = a.Id) So how do you do this in Spark? Well, some people will try to use the Column.isin method which uses varargs, this is okay for a small set of values but if you have a couple of large DataFrames then it's less than optimal as each row needs to be evaluated against the list. So what's the other choice? We can use joins to do the same thing. There are 2 we can use, a SEMI JOIN which is equivalent to our above example of running EXISTS; the other is ANTI JOIN which is equivalent to a NOT EXISTS. Using the above example and keeping the table names as DataFrame names we could re-write this in Scala as: table_a.join(table_b, Seq("Id"), "left_semi") These 2 joins are unique in that they only return the output of the left DataFrame, without any content from the right DataFrame. So what does this look like in practice. Well using Azure Databricks we can quickly create some sample data to try them out. First lets create a couple of DataFrames. First lets runs a simple query to find heroes which have an arch-enemy. This uses the SEMI JOIN to keep records in the left DataFrame where there is a matching record in the right DataFrame. Now, lets have look for heroes who've been a little more active and have removed their arch-enemies (for now). This time we've used an ANTI JOIN to keep only those records in the left DataFrame where there are no matching records in the right DataFrame. You'll notice that in the examples the join condition uses the slightly longer form, that's because in this example the columns we're joining on have different names, and also because there is a column in both DataFrames which have the same name.
  • Recently I needed to deploy an Azure Data Lake Store - Gen 2 instance and thought I'd take the opportunity to use some custom ARM template functions . These aren't something you often see in the example templates but can be really useful if there's a complex expression which you find yourself writing repeatedly within a template. If, for instance, you routinely create resource names based on a prefix, unique name and a suffix then this could save you a few keystrokes. In essence you are simply parameterizing the expression as follows: In this way you can use this simpler expression where you would have previously used the more complex version. [namespace.function(parameter1, parameter2)] If you want to see what this looks like in a full template then checkout this simple ARM template I put together for creating a Data Lake Store - Gen 2 instance over on GitHub.
  • Documentation is not something people often spend time reading, or if they do then its to quickly find the one thing their after and then get out as quickly as possible, very similar to how I do my Christmas shopping. Sometimes it's worth spending time reading the documentation though as there can be some useful bits of information hidden in summary descriptions, links etc... One such item is the Azure Data Lake Store client. If you find yourself reading or writing a lot of files and your doing it in multiple tasks (or threads, but you should be using Tasks if possible), then reading the docs can really help you out. For instance this snippet taken from the description at the top of the documentation page . If an application wants to perform multi-threaded operations using this SDK it is highly recomended to set ServicePointManager.DefaultConnectionLimit to the number of threads application wants the sdk to use before creating any instance of AdlsClient. By default ServicePointManager.DefaultConnectionLimit is set to 2. Okay, so how bad can things be if you don't read this? Well, to answer that I created an ADLS instance and uploaded a number of small parquet files. Then wrote an application to read each file (using the excellent Parquet .NET ) and return the number of records in the file, each file is processed in it's own Task and each uses the same AdlsClient instance. The simple process being followed here is to get a list of files, call " ProcessPath " on each and then when all the files have been process output the results. The output of this initial version is as follows: It's not too bad, but with multiple tasks I would have expected it to be better. Looking at the documentation snippet above it suggests we need to change the ServicePointManager.DefaultConnectionLimit value, but what to? Well doing some digging around came across a suggestion from Microsoft Support which, for ASP.NET, is to limit the number of requests that can execute at the same time to 12 per CPU (or 12 per core). So let's give that a go and see what happens. The code change for this is pretty simple and we can use System.Environment to get the number of processors available. So does it make much of a difference? Well, yes, quite a lot of difference actually. I ran the code in both variations a few more times to check it wasn't intermittent networking issues, other processes on my laptop interfering etc... but no, it really does make that much of a difference. So next time you're working with multiple tasks sharing resources, maybe spend a bit of time reading the documentation to see if there's anything which can make a difference to your application.