Tag Archives: open source

Moving from Storm to Spark Streaming: Taking the Jump

At work, my team has built a true Lambda Data Architecture for our SaaS services (if you don’t know what Lambda is, take a look at a free chapter of Nathan Marz’ book or my prior blog post on why I love Lambda so much). The Lambda architecture is ideal for what we do: sensor analytics:

  • We use the speed layer to perform complex event processing on real-time analytic data
  • We then trigger follow-on windowing analysis based on deferred events and triggers (essentially deferred speed layer calculations across a wider time range)
  • We then follow this with multiple layers of batch layer computations that do everything from analytic wrap-up to discovery of strategic data patterns to update of machine learning models.

Spark: The Promise to Simplify Two Architectures into One

While Lambda is powerful, it is also complicated. It essentially requires you to maintain two virtually separate data architectures: something expensive to build and keep in sync over software releases. This cost is one the reasons some argue against the use of Lambda.

Like many current practitioners of Lambda, we originally used Apache Storm for real-time speed computations and MapReduce for batch layer computations. Contrary to what you might read, you can build Lambda without Storm, MapReduce, or even Hadoop. I have built several using a wide range of technologies (even using human Mechanical Turks for the batch layer). However, up until about 2014, Storm and MapReduce (along with Kafka, HDFS and Cassandra) were the open source, high-scale tools of choice.

Then Spark graduated to a top level Apache project in late 2014.

Since Spark added its real-time streaming capability it has offered the potential of providing both real-time and batch computation from a single software library and platform. Yes, Spark Streaming is micro-batch and more akin to Storm Trident than pure Storm. However, for a broad range of use cases Spark Streaming approximates real-time (soon, I will be writing a post on when I would use Storm vs. Spark Streaming vs. Flink ). However, for all intents and purposes, Spark Streaming was close enough to real-time to process sensor signals that take anywhere from 2 seconds to several minutes to be transmitted from devices over cellular connections to data centers.

So, what is the catch if Spark “offers it all?” Many. While Spark is very powerful, it presents several limitations. First its speed requires lots of memory—which means it is ultimately not as scalable as plain-old MapReduce. Second, it is a bit of a “Swiss Army Knife” of data platforms—which means it essentially is force-fitting a batch model into a real-time one. What would be the tradeoffs of moving this vs. staying with Storm: a real-time specialist proven for scale at places like Twitter and Yahoo!?

Experiment 1: Replacing MapReduce with Spark Batch Jobs

For these reasons, I took my time evaluating Spark before jumping in. First I tried it as batch only tool, replacing some of our MapReduce jobs with Spark (0.9) ones. As expected, the Spark jobs were much faster: about 60x faster. Also as expected, I had to by more expensive servers to get more memory (we use AWS EC2 – not EMR – for several business-specific reasons). Overall I got about 60x performance boost for 4x cost increase. Not too bad of a trade-off.

What I did not expect was what happened to my development costs: they went down significantly. Our developers found it much easier to build and de-bug jobs in Spark vs. MapReduce. How easier: we saw a 75% reduction in bugs and a 67% reduction in development timelines. These savings easily paid for some extra servers (of course, the story would have been different if I had thousands of servers vs. hundreds). When Spark set a new large-scale sorting record a few months later, we felt pretty good that our choice would have a good enough balance of speed vs. scale vs. cost for our needs.

Taking the Big Jump: 700+ Contributors Means A Lot

At this point, I waited. I was curious about the benefits of consolidation. However the developers I trusted at places that had enough resources to use whatever tools still said Storm was more scalable. Having supported thousands of transactions per second for the bulk of my career, I have always been always favored speed and scale (having seen teams die when they could not scale). So I watched

In May 2014, Spark 1.0 came out. Four months later 1.1 was released. Three months later 1.2. I started to look at the release notes and noticed that Spark was becoming one of the most popular open source projects in history. How popular? 3x the contributors of MongoDB. Nearly 6x Cassandra. Almost 11x HBase. (In fact Spark has had 5x the number of contributors of the entire Hadoop ecosystem combined. By the time Spark 1.2 came out, it was clearly evolving faster than Storm, with almost 4x and many total contributors.

Growth in Contributors per Month to major Open Source data platforms (sources: Github and Black Duck software)

When this many people adopt something, good things generally happened. Yes, new features are added more quickly. However, more importantly, developers break things, resulting in less risk and improved performance. When Spark 1.3 came out (the third major release in 90 days), we started a project to move to Spark.

Since then, much of the industry has jumped into the Spark pool (IBM has probably jumped with the most people into the pool). Spark is up to 1.6 as of a few weeks ago (Spark was advancing so quickly we had to migrate to new major versions and regression test twice prior to going to production). Spark MLlib has given us some data science-friendly tools and capabilities. Spark SQL with Apache Parquet  has been mind-blowing–more on that in a few weeks.

However, our journey was not all wine and roses. Some experiences were great. Others are still deeply frustrating today: after six months in live operation, streaming mission-critical data 24×7.  My next blog post will recount in detail what we learned using as real-world data as you can get. Whether you love Spark or Storm, there are some bragging rights to your favorite platform. Stay tuned…

BTW: If you like this stuff, I am hiring 😉

A Practitioner’s Guide: Best NoSQL databases to solve 9 real-time transaction challenges

There are a lot of articles out there praising the features and performance of one NoSQL database over another. However, as a practitioner of the principle “pick the right tool for the job”, I thought I would write an article on picking the right NoSQL database for the transactional challenge you are facing—whether you are a startup, mid-sized enterprise or Fortune-500 company.

Before I get started, here are some caveats. First, I have no affiliation with any the companies who provide, or serve as custodians of, the databases outlined in this post. (I have, however, used them all.) Second, I am a big proponent in open source software. This is NOT based on a philosophical bent, but instead decades of experience scaling platforms for hundreds of millions of transactions per day. Third, I am big believer in ecosystems. When you big a technology with a good ecosystem, you get many benefits: you can hire people who no it, you can find lots of tools and libraries to enhance your use of it, and you benefit from patches to solve problems others have already found for your. Ecosystem size factors pretty heavily in my recommendations, as you may not want to bet your company on an untried technology.

Finally, I should address the elephant in the room. You will probably have someone in your company say something like this:

SQL can do anything NoSQL can do. It’s too risky to use it. We lose ACID compliance. We have a huge learning curve, etc.

Yes, relational technology is great for many uses. However, there are many situations where NoSQL technology can do things bigger, faster, at lower cost, and with less effort. Not just by a little bit, but often factors of a 1000x or more. Simply using RDBMS technology for every challenge is akin to using one hammer to for any type of nail, screw, peg, etc.

With this out of the way, let’s get started.

Challenge 1: Search (and Wildcard Searches)

You want to create a place where people can search for content in your site (and handle the idiosyncrasies of misspelling, grammar, and interchange of words like “one” for “1”). Similarly, you want to allow back office and enterprise application users to perform wildcard searches (such as all orders that contain “Nike”) as quickly as they can search for content on Google.

To solve these problems, use a search engine (which is technically a NoSQL database of inverted indices paired with lots scoring and fast spell checking functions). My favorite is ElasticSearch(ES). It is really fast and gives you some interesting capabilities you can use for specialized search (see Challenge #: Recommendation).  A second choice is Apache SOLR. SOLR is quite a bit slower than ES. However, it is included natively in many NoSQL distributions (HortonWorks, Cloudera, DataStax, etc.) If you already have gone through the cost of implementing these it makes sense to stick with Apache SOLR to get more value out of your investment.

BTW, never give users the option of wildcard or text searches in non-Search databases. It is a performance and scalability nightmare.

Challenge 2: Managing Versioned Pages of Information

You have a content management system for writers, journalists, etc. and want to maintain versions of their articles. You have an information management system for life sciences such as electronic medical records (EMR) or eClinical systems, and need to keep a copy of each version of a medical records page (or a clinical case report form, a.k.a. CRF).

To solve these problems use a document-oriented database. Instead of tracking associations of records (or paragraphs) to data to versions, you can simple store the whole document (often in JSON, preserving markup and annotation data) as a single entry for each version. My favorite is Amazon’s DynamoDB, based on its ease of setup and scalability. A close second for me is MongoDB (especially if you have a requirement for on-premise management). MongoDB would benefit by making it easier to setup multi-node clusters with encrypted data transfer. This still takes too much DevOps work.

Caveat: While Document DBs are great for storing pages as documents, I would not use them to present documents as pages to high-volume end users (e.g., for a content management system). Rather than present the ‘current published’ version of content from a database you should use some sort of cache. The easiest solution to setup is Amazon’s CloudFront CDN. However, if your scale and team are big enough Varnish is a more cost-effective solution.

Challenge 3: Managing Streams of Data, Events, and Actions

You want to manage long streams of information. You want to store all the events in a customer’s life cycle (for customer lifetime value management) and that activity is frequent. You want to store sensor events and GPS position reads for an asset (e.g., movement of freight in a reefer container). This challenge is getting more frequent given the explosion of data for mobile, sensors and social sharing events

The best tool for this is a wide-column database (very different from OLAP columnar analytic databases). Wide column databases bring two advantages: they let you store sparse information very efficiently (imaging storing all the topics a person could Tweet about or all the variables a sensor could capture). Second – with some good engineering work – they scale like mad and let you fetch this information back on 1/1000th of the time as traditional relational databases.

All the major columnar databases are children of Google’s BigTable database (that tells you something). My favorite wide-column database is Cassandra, especially if you want to use it for real-time streaming analytics and complex event processing. A far second choice for me is HBase (IMHO, HBase is essentially Cassandra–but with overly complicated ops management). However, if you already have a large bundled Hadoop installation from a major provider (e.g., HortonWorks, Cloudera, MapR) you probably already have HBase installed. HBase would also be my first choice if you company or groups is primarily a batch analytics shop (i.e., a big MapReduce data warehousing shop).

Challenge 4: Recommendation Engines

You want to recommend the best thing given other things your customer has viewed, liked or purchased. You want to recommend the best business to connect with based on given customers relationships with similar businesses. You want to find potential new customers based on their similarity to existing customers.

In reality, the best tools for recommendation are machine learning based. However, graph databases can make things easier machine learning algorithms or more basic recommendation features (such “users who liked this also like these”) to get the right information. Graph databases are still a bit of nascent market. The biggest leader is Neo4J. However, if you already have a Cassandra, HBase or (I am not sure why) Oracle Berkley DB installation you could simply install Titan to use this data. This would be my number one choice as Cassandra and HBase are fantastic ways to store very sparse info on customer preferences (e.g., all the things the viewed, liked or bought) thanks to their ability to support hundreds of thousands of columns per row—and read quickly along a column down a single column across many, many rows.

If you are just getting your feet wet in the recommendation engine space, you can also use ElasticSearch’s boost feature to highlight content for recommendation based on item scores. (These scores can be simple post-transaction calculations or fully robust machine learning driven scores). I once used ES to setup a recommendation engine in less than two weeks that actually beat a Google Search result. ES (and Apache SOLR) also have the “more like this” recommendation feature available out-of-the-box.

Challenge 5: Relationship- and Attribute-based Exploration

You want to create a site that users can browse to find “things like this” or “things related to that.” One example is searching for restaurants based on confluence of style, ingredients or similarity to restaurants I like (this could be equally applied to wine exploration on Lot18 or art exploration on Art.sy). I very popular use of this is browsing sports statistics, such as using Pro-Football-Reference.com for Fantasy Football team assembly.

For these challenges I would go immediately to a graph database. Out of the box, Neo4J would be my first choice. If I already had a big Hadoop or Cassandra installation, I would go to Titan.

Challenge 6: Knowledge Base Exploration

You want to create knowledge base to find the right “how to” content to address a problem, e.g., fix my account, figure out how to perform a feature in an app, etc. You could do this from customer facing site, call center, support desk, etc.

I spent many years working with companies who spent tens of millions of dollars using exotic database solutions to solve these problems. However, none of them worked as well as Search Engines. Search Engines are fast, scalable and handle all the vagaries of user behavior (such as spelling and grammar errors, use of “one” vs. “1”, etc.). In addition, thanks to Google, people are now “trained” to use Search to find answers to their questions.

ElasticSearch’s boost feature makes an amazing knowledge base manager. Search finds the relevant content and attributes such as views, votes, helpfulness ratings, etc. drive the boost score to raise answers up or down. The best example for this is StackOverflow (the developers crib notes for just about any problem).

Challenge 7: Sorted Catalog Browsing

You want to let users search a catalog of items, then switch to sort by things like price and rating or switch to browser by category and sub-category. You want to allow users to look at broad categories and drill into sub-categories to find items to buy. Basically how we all start looking for something at Amazon.com.

The starter technology for this is Search (again ElasticSearch is my recommendation of choice). You can use ES’s aggregation features to search and browser by category. However, as you get really big (i.e., millions of items in your catalog) or you wish to allow users to pivot from search results into structured browsing or explicit assignment of SKU items to categories you would add a columnar store database such as Cassandra or HBase. To see this action, search on ‘shirts’ at Amazon. Notice the phrase ‘Choose a Department to Sort’? This forces you to move from a search result to a columnar-based query based on nested keys. (To be completely clear, Amazon uses its own home-built tech for this, not Cassandra or even the stuff they expose on AWS for the rest of us to use. However, the principles are similar).

Challenge 8: Capturing Logs for Analysis

You want to capture logs of information for later forensic analysis. This could be for debugging, performance analysis, security and penetration analysis, or simply for audit compliance.

My number one choice for this is Riak. It is fast and simple to use. (The Mozilla Foundation uses this to capture all crash report data). A second choice for this is MongoDB as it is already bundled with many log analysis applications. (However, Mongo is more costly to scale than Riak). Another choice is simply streaming this data to a file store such as Apache HDFS to later interrogate with HBase, Hive, PrestoDB, or ElasticSearch.

Challenge 9: State Machine and Session Management

Ok, this one a bit technical. However sometimes you need to keep track of info for rapid lookup for later. You may want to keep track of the pages a person recently viewed on your site. You may want to keep a pre-approval status. You might want to hold a ticket for concert in temporary lock before purchase. You may want to keep the state of a customer or item for complex rule processing (e.g., keep track that I am exercising for fitness target alerts).

The best types of databases for these challenges are key-value Stores. My favorite is Redis. It is ease to setup and super fast. However, Redis is optimized for in-memory option. It is less optimal if you need guaranteed persistence of the information you want to store for rapid lookup. If you need persistence (such as storing pre-calculated algorithm or pre-qualification models), and have no pre-existing infrastructure, I recommend MemcacheDB. However, if you already are using Cassandra, HBase or DynamoDB you can simply use that as rapid key-value store (with persistence). You may even find this more cost-effective than setting up a Redis cluster even if you do not need persistence


There are a lot of NoSQL technologies out there. The trick, like all things in life, is the pick the right tool for the job.