This is the fifth post on my series on the inner workings of the Amazon Fulfillment Center. Here I will look at the software that runs behind all the processes and makes this performance possible. Other companies would probably plaster the label “Industry 4.0” all over this, but at Amazon they just do it.
Please note that most of the images and all of the videos are courtesy of Amazon.
The Heart of Fulfillment: Amazon Fulfillment Technologies (AFT)
Above I talked a lot about the physical aspect of the Amazon Fulfillment Center. While this is nicely done, the real power of Amazon fulfillment is invisible in the software, called Amazon Fulfillment Technologies (AFT). They claim it is the largest fulfillment execution engine in the world, and is their connection between the physical and virtual world. Due to the huge number of widely different products, this is a highly challenging and complex task. In a single day, Amazon handles around 37 million orders.
Most other companies would plaster the label “Industry 4.0” all over this, but Amazon does not really mention Industry 4.0. They just do it. I like that. Nevertheless it includes a large number of advanced machine learning technologies for forecasting, text comprehension, image recognition, translation, speech recognition, and a lot more.
Example: Bin Images
One example are the images from the pods used for machine learning. The tape holding the items in the pod is transparent to make identification easier. After picking, photos are taken of the pod shelves (called bins). One of these you can see here on the right. Machine learning is used to answer questions like
- How many objects are in the bin?
- Is there a certain product in the bin?
- How many of a certain product are in the bin?
A data set with half a million bin images is available via Github for machine learning. I looked at a few images, and for quite a few of them I would have trouble myself to answer correctly.
While initially based on Oracle, the Amazon fulfillment database eventually became just too big to manage. In May 2017 they switched to their own Amazon Aurora system, although they turned off their last Oracle database only in October 2019. Aurora is a cloud-based MySQL- and PostgreSQL-compatible database (both are open source database management systems). Amazon claims that it is five times faster than a normal MySQL database and three times faster than a PostgreSQL database. Amazon also claims it costs them as little 1/10th of the previous Oracle database, both in licensing (60% reduction) and administration (70% reduction). The maximum size per database seems to be 64TB, although Amazon migrated around 75 petabyte of data from Oracle to Aurora (including also other databases like Alexa, Prime Video, Amazon Music, and more).
A computer decides what to pick, when to pick, activates the Kivas if available, and organizes the shipment. And this computer system is thoroughly refined. Based on a customer’s order, it makes an initial plan but refines this plan at every step based on minute details such as the likely route of the truck, the available space on a truck, and many more aspects that I can only imagine. Hence the fastest and/or cheapest path to the customer is frequently updated based on new information.
This system includes actual artificial intelligence. While AI is talked about a lot, it is still rarely seen in industrial applications of logistics and production management. One of the features is also machine vision to analyze the pictures of the Kiva pods or other products.
In FRA3, the average time between an order and the corresponding parcel leaving the docks is an astonishing 2 hours and 45 minutes – and FRA3 is one of the slower fulfillment centers due to the handling of clothes and the manual picking. Their goal is to improve this to 2:30 hours. The current record of Amazon between the order of the product by the customer and the ringing of the doorbell of said customer is … hold on to your seat … 3 minutes!
I need more time if I want to get something from my basement! One of my past jobs considered it as Just in Time to get a part from the warehouse across the street within 3 days. And Amazon delivered a parcel in 3 minutes! This is quite a statement on the power of their physical side of the fulfillment, but even more of the software side. Granted, the customer lived right across from the fulfillment center, and also granted this is the best time out of who-knows-how-many gazillion parcels delivered, but I find this highly impressive.
In some locations, Amazon even delivers regularly within 2 hours. This Amazon Prime Now delivers a wide range of goods within 2 hours for Amazon Prime customers in selected locations. Many larger cities in the USA are covered, as well as some cities in the rest of the world. Only a few years ago, a delivery time from ordering to ringing the doorbell within a few hours was unbelievable for end customers, unless you pretty much paid a taxi to bring the stuff to you. Now Amazon does it for free!
In a previous article on Industry 4.0 – What Works, What Doesn’t, I talked about where Industry 4.0 makes sense. One of the areas was logistics. Industry 4.0 can be very expensive. Logistics is a segment where there is an economy of scale. If you have programmed one Kiva, then you have programmed them all. If you have the software back-end for one fulfillment center, then (with minor adjustments) you have it for all. Hence it makes sense for Amazon to pour significant resources into their Amazon Fulfillment Technologies software, since the benefit can be shared across their hundreds of fulfillment centers. Overall, an amazing piece of work that I believe will continue to be improved on.
While still a lot is unknown to people outside of Amazon, I consider this Amazon Fulfillment Technologies (AFT) exceptional. Getting so much data together, keeping track of it, and updating it in real time is quite a challenge. But Amazon has the resources to push this software (they are hiring a lot), and the implementation has company-wide benefits. Overall pretty cool. Now, go out, and organize your industry!
- The Inner Workings of Amazon Fulfillment Centers – Part 1
- The Inner Workings of Amazon Fulfillment Centers – Part 2
- The Inner Workings of Amazon Fulfillment Centers – Part 3
- The Inner Workings of Amazon Fulfillment Centers – Part 4
- The Inner Workings of Amazon Fulfillment Centers – Part 5
- The Inner Workings of Amazon Fulfillment Centers – Part 6
2 thoughts on “The Inner Workings of Amazon Fulfillment Centers – Part 5”
Another excellent article.
I liked how you mentioned that one of the greatest drivers to Amazon fulfillment is not something that physically moves product but the software that runs the entire system. It was very interesting to learn about the software Amazon Fulfillment Technologies. AFT must be very efficient for Amazon as they use it as their fulfillment execution engine to run through a wide variety of products and receive millions of orders a day. It was also very interesting to read about Amazon Aurora, Amazons own fulfillment database. Over this past summer, I worked at a company that used Oracle to manage order fulfillment but hearing Amazon was too big to be able to manage their orders is crazy to hear. It was cool to learn about how Aurora uses cloud-based MySQL and PostgreSQL and is faster than each of these databases comparatively.