Big Data

Kappa Architecture on Bluemix

Share this post:

During the last three years I’ve been responsible for helping the US Masters, Roland Garros and primarily Wimbledon understand how people are engaging with the events on social media. Their requirements have led to a messy combination of stream processing, batch processing and static indexing. I’ve started to look at how a Kappa architecture might help us produce a cleaner solution.

Streaming and Batch Processing
All three sports events are transient in the level of interest they attract. Their very nature entails a high volume of data over a short period of time (1-2 weeks), before receding to a low background level. It is important to each event that the analysis of social media be done in realtime – giving their media teams the opportunity to react to what is being said. They also need to be able to access historical analysis in order to put what is happening now in context.

The data analysis requirements therefore are:
Large volume of messages (~30m)
Realtime analysis
Historical analysis

While data processing for streamed (realtime) and batch (historical) data are well understood, solved problems. The crux of the challenge here is in combining both.

Lambda Architecture Option
Lambda architectures have gone someway to address this problem, they still have stream and batch processes, but allow the same code to be used in both, reducing development and maintenance cost. Apache Spark can be used in this way (it was the solution we used for the 2016 sports events). This is a good option, but had some limitations in our case.

The user interface in our application did not have a clear distinction between historic and realtime analysis. Why would a user expect there to be one? The initial view was of realtime analysis, but a user could scroll back in time, seamlessly switching between stream and batch analytics. With careful coordination this could have been made to work, but that coordination would have been difficult to implement and left a lot of space for things to go wrong, something I wasn’t comfortable with. The sports events projects already have a lot of moving parts, so I’m always anxious to avoid adding complexity.

Stream and Store as a Fudge
Our solution for 2016 was something of a fudge, albeit a pragmatic one. We processed all the data in realtime but stored the results by time snapshots in an Elastic Search index. All data displayed in the UI was served from the static snapshots, meaning the ‘realtime’ data was not truly realtime, but just the latest snapshot. These were made every 30 seconds, which was as close to realtime as the end users required.

Though this solution did its job, it was not one I was ever happy with. The realtime data was not true realtime and for some applications 30 seconds might matter. It also made it difficult for us to change the analysis. If we improved an algorithm we had to reprocess all the data. The system was setup to process streamed data, so the only way to reprocess data was was to replay the historical data as if it was new (at the same time as the genuinely new data was arriving), and output the results to a second Elastic Search index – the sort of coordination mess I was looking to avoid in the first place. This re-analysis chaos meant I was loathe to reprocess data. The process took several hours and users would see no results while it was running. We avoided doing this at all costs, meaning we didn’t improve the analytics as much as we would have liked to.

Kappa Architecture
Given this background I was searching for a cleaner solution when I came across the idea of a Kappa architecture. This architecture makes use of an immutable, append only log. All data is stored on the log and all processing is run against it. Processing is always done as streaming, but using a log as the data source means you can set the position in the log where processing should start from. Only care about realtime? Start at the end of the log. Need to analyse all the data? Move to the front of the log and process to the end. The idea with Kappa is that we have the tools and compute resources to process data so quickly that we can always process the full data set, rather than storing a current value and processing transactions on it like we would do in a traditional database.

We were already using IBM Bluemix for the sports events analysis and it has its own immutable, append only log, Message Hub, which is a managed instance of Apache Kafka. Looking towards 2017 I set out to prototype a Kappa architecture on Bluemix.

I wanted to build something that would abstract the details of the Kappa Architecture away from developers. I wanted them to be able to run Elastic Search-like queries and not worry if the data was being processed in batch, stream or a combination of the two. The system should give them a result and automatically update that result if and when the underlying data changed.

Kappa Bluemix Prototype
The prototype I developed allows users to send queries via a HTTP Post message. They are returned a URL to a websocket which will push the results to them. There is no ‘final result’ in Kappa Bluemix as there is always an expectation that there will be more data arriving. The prototype assumes the data is in JSON format, but this is a limitation of the query engine, not the underlying architecture.

That there is no distinction between batch and streaming data queries has simplified things for the developer, but there is a performance hit. Having to process the whole log, even for a simple query presents a huge overhead compared to a traditional indexed database. With the prototype I’ve made some effort to mitigate this. Kappa Bluemix shares queries. If two or more users submit identical queries, the second query uses the same output (and subsequent updates) as the first. This form of streaming cache minimises the overhead for common queries, though does nothing for new queries.

The initial performance penalty of a new query is visible to the user in the time take to be return an answer, but once processing has caught up to the end of the log it’s just as efficient as regular stream processing.

Future Direction
There is a lot to do to make this prototype in to something I could use on a real project.

There’s currently no way to run a query in a distributed mode, but there is no reason why map-reduce styled programming could not be used. Without this it will never be fast enough to be practical for interactive applications.

I need to investigate if this architecture has any negative side effects on Message Hub. I’m not aware of any, but the way I have lots of connects and disconnects, multiple consumers with dynamically generated group IDs and continual resetting to the front of the queue is different to how Message Hub would normally be used. I’m wary of any unintended consequences this may have on the infrastructure.

I would also like to implement the prototype in OpenWhisk – it’s currently Java based. OpenWhisk is an appealing platform given that the system is likely to be idle for much of the time and then rapidly spike to be extremely active when processing data from the start of a queue.

One side effect of the architecture that I’ve come to appreciate is that you are submitting persistent queries. If I post a query to count the number of records with the name, ‘fred’, I’ll be returned the answer now, but assuming I leave the websocket open that answer will be updated if the data changes. This can simplify UI code significantly, a developer doesn’t have to code to get an answer and then check for updates, it’s all one query. This shares lot of the simplicity inherent in reactive programming. This kind of persistent query would be prohibitively expensive using a traditional database, but with Kappa you get it free. I’ve come to think that above the cleanliness and simplicity of the processing, this may be the real benefit in this architecture.

Try It Out
If you’d like to have a play with what I’ve been working on, please download the example from the Emerging Technology github pages and please get in touch, I’d welcome any input, ideas, suggestions or criticism.

More stories

IBM Player Spotlight Built with IBM Watson

The Emerging Technology team within IBM Research UK, have been working with FOX Sports to enhance their viewer’s experience of watching ‘The Beautiful Game’. The IBM Player Spotlight Built with IBM Watson is an AI-powered tool built for the FOX Sports’ studio team to access complex statistical analyses using an easy-to-use conversational interface. With the […]

Continue reading

Call for Code Finalists

Call for Code This year, IBM, in partnership with the American Red Cross, the UN, the Linux Foundation and many more, launched the Call for Code. Developers have revolutionised the way people live and interact with virtually everyone and everything. Where most people see challenges, developers see possibilities. The Call for Code is a multi-year, […]

Continue reading

Emerging Technology at THINK 2018

Emerging Technology attended the THINK DevZone at IBM’s flagship THINK conference in Las Vegas, showcasing recent IoT projects!   Health and Safety IoT The first demonstration on offer was Bob – a lone construction worker working on an asset such as a powerline. Whilst Bob is working alone, his health and safety needs to be […]

Continue reading