eCommerce is the business of selling online – a rapidly growing market that was estimated to be worth $1.7 trillion (€1.5 trillion) in 2015. In the UK last year, Black Friday broke the £1 billion (€1.25 billion) threshold in sales. There are roughly 1 billion websites in the world, but only 173 million or so are active, and less than 1% of these are involved in eCommerce or web sales.
The Internet is the new marketplace, but there is a problem: how do businesses selling to online retailers find their sales leads? There is no directory listing all of the online shops out there. By using Google, Bing, etc. businesses can find the top 500+ online retailers but are not able to find critical information about these sites e.g. technology the website is using, merchant size, or contact data.
Companies lose 1000s of hours by manually crawling the internet and entering information about the online retailer into their CRM system. They also use LinkedIn to generate sales leads, but online retailers don’t generally reside here. They live in the world of Instagram, Facebook, Twitter, YouTube, etc. There are no search engines or data mining tools available for these platforms.
This is the opportunity that TAMI has identified and tackled head-on. We have created a B2B search engine; a sales lead generation SaaS platform. We comb through the web looking for online eCommerce sites that would be of interest to our customer base.
We do this using web-crawlers to download content, and then analyze that content to understand whether the site is potentially of interest to our clients.
In doing this, we need to harvest and search through vast amounts of information – terabytes of data must be downloaded and analyzed on a weekly basis.
We then need to augment the sites we have found with additional data, such as social media information, web traffic volumes, contacts for the site, and various types of website ranking metrics. We also need to perform company sizing calculations, data confidence metrics, and analyze websites to assess whether they are related to other sites, shared, or owned by the same corporate entities.
By doing all of this, it means that the online leads we generate are well prequalified, and therefore very little work is required from our customers. They just need to contact the company to identify whether it is interested in the products or service they provide.
A web crawler, much like creepy crawlers in the real world, tends to have a bad name. A badly developed crawler can wreak havoc on a website, with some so bad that they can cause the site to crash. So being mindful of what you do in the real world is very important.
Web crawlers need to be really careful not to overload the site they are interested in analyzing. They’ll get kicked out pretty quickly by any good firewall/traffic monitoring system if they do; but a lot of smaller sites don’t have this facility, and at worst web crawlers could cause the site they are analyzing to crash because of overloading.
Here at SalesOptimize, we actively minimize our footprint on a site; at most, we try to download only a single HTML page from a site around every 8-10 seconds. In reality, because of the volume we analyze, it can actually be minutes/hours between each page we take down. We spread the load across tens of thousands of websites at a time and we like to treat each site we contact as if it were one of our own.
Having software capable of scaling across multiple servers and processing nodes is essential when it comes to accumulating and managing massive volumes of data. There are many ways to manage scale, but the most effective way of scaling that I have come across is to:
We happen to persist the messages in a database before publication, and we always publish to the distributed messaging system from that data source. So for us, chaining further actions is really just about calling a message factory that will persist a set of messages in a database.
We track the message from the database to the queue, and we store various nerdy execution metrics per message, e.g. how long it was hanging around before processing, how long took to execute, and whether any errors were encountered during executions. This helps us to understand the workloads we have and where we need additional services spun up.
When dealing with scale, data pipe speed levels to the on internal networks and to the internet will make a massive difference. Currently, we run all of our services through Azure. Luckily, I also have experience with Amazon environments and both are equally good in my opinion. Regardless of which environment you choose, having massive download/uploads speeds is essential to good scalability and performance. Particularly when dealing with millions of distributed messages flying around the place along with data downloads and uploads taking place at the same time. Reducing your IO wait times on reads and writes to a minimum is essential in the modern environments in which we work.
So my advice would be to make sure you’re happy with the performance of the internal and external environments you choose to run in. If you can manage your IO operations efficiently, it will help you to scale with the best of them.
I’ve found that there is no one database solution to suit everything. Using Relational, NoSQL, Column Store, Graph, and Name value data-stores all make sense in their own use cases. When dealing with scale, each has its place, and understanding its usage is essential. In SalesOptimize, we currently use MongoDB, Microsoft SQL Server, and elastic search. We also use various ETL formats for shifting data between different environments.
We constantly evaluate the types of data we need to store, analyze, and report on. We consider production versus back-office analytical services, what’s best for front of the house (our customer-facing UIs and services), and what works best in the dark rooms where the developers and analysts work.
Because of this fact, there is no magic answer for which data store to use. Understand the options you have available to you: how do you choose one over another? Will it scale? Can it be clustered, shared, and portioned? What is its data retrieval speed like, compared to its analytical speed? Are there reporting and analytical tools available to you to so you don’t need to develop your own? Answering these questions will help you to choose what’s best for the different environments your software needs to work in.
Because you never quite know what you are going to find when analyzing a website, you need to build in a confidence-based approach to your data gathering mechanisms. For example, we look for various shipping companies used by a website. We might search for DHL or UPS, but a match for “UPS” may also be found in the phrase “ups and downs.” So how do we know if UPS is actually used as a delivery mechanism? You need to gather additional confidence in your data and place it in the context of other related data.
In the example above, capitalization and related context may matter; the mention of shipping/delivery or even other carries close to the data you found may (or may not) indicate whether UPS is actually used.
Building up additional confidence metrics with every single data point you are interested in analyzing is essential to providing relevant data to our customers, and probably to yours too.
We run everything in the cloud, but we initially tried to use cloud hosting providers for our data and analytical stores. This was fine at the start as we were building up data, but we found ourselves landing with a monthly bill of almost €10,000 for one service. This was a result of the volume of data we had built up. Needless to say, we had to change tack quite quickly. We really would have liked to hand over the management of large clusters of data to 3rd parties; it saves us having to worry about the daily running and management of these, and frees us up to do what we need to do. Unfortunately, we had to take control of this cost. Otherwise, it would become more and more expensive over time.
This meant that we now needed to get up to speed very quickly; understanding clustering, partitioning, sharding, striping large volumes of data and making those datasets resilient and scalable.
As your business grows, you may be faced with the same challenges. If this is the case, you must be willing to devote time and resources to tackle this. Having a plan upfront (which we didn’t, by the way) will save you time, worry (aka hair), and costs in the long run.
As our company grows, we will ultimately have a team dedicated to looking after these operations. For now, however, we as a developer group are looking after these issues.
When dealing with scale and big data, you need to: