Notes on a tour of an Amazon distribution facility (2014) | Hacker News Reader

Notes on a tour of an Amazon distribution facility (2014)

docs.google.com

145 pointsscmoore39 comments

That's a old manual Amazon facility. The newer ones use Kiva robots to bring the shelving units to the pickers, rather than the pickers going to the shelving units.

The manual system isn't that different from the Sears "schedule system" introduced in 1908.[1] Note the line "pickers don't necessarily pick items for a single or even complete order." That's crucial. Orders are split apart and combined pick lists generated. Picked items then flow towards order assembly stations, which Amazon calls "sort". At any one time, some maximum number of orders are in progress, limited by the number of output boxes at order assembly. Amazon does this with computers; Sears did it with clerks and pick slips, with Sears giving each order an assembly bin for a fixed number of minutes.

Separating picking from order assembly and inserting a sort phase reduces the order of the problem. Picking N orders from M items individually means O(N × M) cost, because as inventory becomes larger, the pickers travel more distance. With separate picking and assembly, performance is something like O(log(M) × N) cost, because each picker works in a limited area. That was Sears' big breakthrough.

[1] https://news.google.com/newspapers?nid=1734&dat=19721006&id=...

Kiva robots, for the curious:

https://www.youtube.com/watch?v=6KRjuuEVEZs

By the way, wouldn't segways bringing people to the shelves have to expend less energy and be more efficient than kiva robots bringing shelves to the people?

Hmm.. don't know why asking a question deserves a downvote.

I think you're right that it would be more energy efficient -- since a human weighs less than a rack full of goods.

But energy, compared to the other costs, is a fairly small component of costs.

The bigger question is whether a person is cheaper than a machine.

A few years ago before Amazon bought Kiva, they would sell a complete setup for a large warehouse for $15-20m for 1000 robots. This included a lot of setup, but even if we assume it's just 20m/1000 robots = $20,000 each.

So for $20,000 you can get a robot that will work 24/7 for a few years. Lets pretend the kiva robots are completely junk after 3 years. That's under $7000/yr.

Obviously, that's going to be tough to beat with a human. But for comparison, lets assume a person gets paid federal minimum wage ($7.25/hr), there's no overhead, no management, healthcare, never gets sick, etc. That equals $15000/year. But that's only for 40 hours a week x 52 weeks/year. So really, we'll need 3 people.. so $45000/year. (EDIT: I forgot about weekends, which is another 48 hours. So really, we need 4 people, or $60,000/year.)

Considering that humans are so much more expensive than a kiva robot, the increased energy compared to segways doesn't really effect the outcome. It's clearly cheaper to use robots and simply pay for the increased energy usage.

Energy is cheap; people are expensive. One kilowatt hour costs about ~80 seconds of a human's time, if the kwh and the human are both in Chicago and the human gets federal minimum wage and costs nothing in overhead.

(Any human fitting this definition is also spherical and frictionless, i.e., existing nowhere outside of physics textbooks.)

Bonus points: this remains true if one does not value money but one instead counts costs totally in carbon credits, which is one of those things that the environmental movement passionately believes and yet doesn't spend much time thinking of the implications of.

Not really if you define (as Amazon would) efficiency as order throughput and cost-per-order. Humans are typically more expensive than robots per hour: robots cost upfront and then maintenance + energy, but if those sum up to an hourly rate beyond the cost of humans for a jobs that humans can do at comparable quality... you wouldn't have robots yet. Kiva exists and continues to get deployed...

To fulfill an order with Kiva, a human stands at a ship station (boxes, label printing, tape) near a fast path (door, belt) to outbound. Shelves are brought, the human picks the item from the shelves and assembles the order. That's putting the humans in the loop where robotics is harder/more expensive. Your hard fulfillment rate limit is then the time it takes the human to pick and assemble the order, plus whatever time it takes the robots to bring the shelves over. This scales gracefully (install more robots until the hard latency of order fulfillment becomes robot travel time from the farthest shelf to the human, plus human time) -- robots scale with the number of shelves, humans scale with the number of fulfillment stations.

To fulfill via Segway, you tie up a human for the entire travel time between station and the traveling salesman problem of order picking and back. Or you do as Amazon was doing in this article, and split pick and pack and allow for subdividing orders between pickers, and so on. This works (obviously) and the segway approach would likely improve fulfillment speed for pickers. But it does not save operating expense (headcount).

At my FC, pickers are moving way too quickly and navigating around all kinds of obstacles for a Segway to be practical. I believe one of the primary advantages to the Kiva method is density. Aisles can be stored closer together.

In addition to power being cheap (as others have mentioned), the humans would experience 'downtime' when moving between stops. Downtime for humans costs a lot of money.

I was wondering if the robots have to really carry the whole shelf ?

The shelves could be modified to be like automatic wending machines. The robots could just go to the shelf and the shelf will dispose the correct item to that robot. Then the robot goes to the next shelf for that order and collect all the items required for the order from different shelves and return to the packer.

The packer need to only arrange the items in the final box.

There are systems like that, but they're much more expensive than the Kiva system. Kiva just needs a standard, simple mobile robot base and a lot of sheet metal shelving units. It's reliable, because no one robot is essential; if one fails, another one takes over. All you need is a big warehouse with a flat floor; there's not a lot of custom mechanical engineering of conveyors and tracks.

The Kiva system is low-maintenance and low-skill. All you need on-site is someone to replace batteries and tires and clean the things; anything more than that, just ship the robots back to Kiva HQ in Massachusetts. No need for any on site engineering talent or on call maintenance.

Dispensing racks for picking systems are complex, expensive, and have single points of failure for each item. They're essentially huge vending machines. They do exist, though.[1]

Amazon has an R&D program under way to automate the picking process where the Kiva shelf unit reaches a human. They have a competition for robot picking.[2] The prize is $26,000 for a solution that will save Amazon billions.

[1] http://www.ssi-schaefer.de/en/conveying-and-picking/automati... [2] http://amazonpickingchallenge.org/

My guess is though this is entirely possible it is far too expensive to modify all the shelves to have this behaviour when you have a 1.2M sqft. facility.

Amazon recently announced they would be hiring 1,000 workers in my hometown.[1]

Sorta crazy they're using robots & still need to hire that many people.

http://www.chicagotribune.com/business/ct-amazon-joliet-0811...

Despite popular belief, most Amazon facilities do NOT use kiva robots. The fulfillment center would have to be completely rebuilt on the inside to accommodate them. Heck, even some new facilities still don't use them. I work over at Amazon in one of their largest fulfillment centers in the USA. To answer your question regarding where a tote is sent in the facility is simple. During each shift, there are any number of queues open (usually about 5). When receiving an item from DA, we will scan an item into a tote and the computer determines, based on the size of the items, where in the facility it needs to go (library, or library deep). Aside from a few exceptions, most items are randomly stowed. The stowers all have a rate, and because of that, will find whatever space fits their needs. Pallet mass (items with a lot of the same quantity to fill a pallet) is checked regularly and pulled from the regular shelves and sent to a sorting area (at least that's what they started doing at my FC) where it can be stowed somewhere in "pallet land." Try not to think too much into where items are stowed, it is pretty much a clusterf*ck across the FC.

How does the work for stowers is arranged ? do the computer give them specific sets of items to stow ? or is arranged according to the order the goods arrive ?

And what is this pallet land ? what's the benefit?

Work is arranged by what truck is being processed, we attempt to follow FIFO as much as possible. Queues are determined, to the best of my knowledge, by which area of the FC has the most space (but they probably do it arbitrarily, I'm not in charge of any of that, so I don't really know their method). When a truck is being received, all cartons will be placed inside totes (or "juice carts") and sent up to the queues (there's usually only a couple queues that are "dropping" totes). Stowers are advised to take the totes, which are placed on u-boats, by using the FIFO method. Stowers are not given specific sets of items to stow; they only stow what's given to them.

Pallet land is the part of the FC for products we carry in large quantities (i.e. enough to fill a pallet, quarter pallet, or half pallet). Pallet land is just a nickname we give part of the pallet area. We also have a 3-story rack area full of pallets (at my FC). Some items we have on pallets include some Starbucks, Sonicare, and Darth Vader costumes (omg we have so many).

The thing I took away from this was mistake minimization. As shipping the wrong item likely results in a mess of expensive actions down the road, it is much much better to make sure that the right item is shipped the first time. To this end they are willing to shuffle their entire warehouse just to prevent similar items from being set next to each other.

Shuffled is also a feature. You almost never buy a 32" TV and a 34" TV at the same time, right? So it is highly unlikely that similar items would benefit from being next to each other. It helps for finding if that was a problem, but finding is solved by computers anyway, so the layout can be optimised for pathing instead.

If you watch the Kiva video linked above, they also optimize for popularity of items - so as items are picked out of each crate, the remaining items are analyzed. If there's no popular items remaining on the create, they go and park it at the back of the warehouse.

I don't think it's really random.Since it's dictated by the computer, even though it looks random, i'm sure there are some patterns in the order(for example there are items that statistically sell together) ,But maybe they keep it as a trade secret , only to be known by a small group of developers and managers.

To quote Lunarsm below:

> Aside from a few exceptions, most items are randomly stowed. The stowers all have a rate, and because of that, will find whatever space fits their needs.

We only have a few queues to stow into per shift where all the product comes in, so I doubt the computer determines THAT much of it, but I could be wrong.

This is a pretty great description (disclosure: I currently work at Amazon writing software used inside the Fulfilment Centers). I think an important point is there is big variance between facilities throughout the network. For instance, the use of Kiva robots as mentioned by others here. Meaning if you were to go on tours of other facilities you would definitely see new things at each (and of course many similarities).

As I read through the notes, the one thing I really liked was the 4 day work week. Three days off just makes sense? I belonged to a union, and we had the opportunity to work 4 10hr. days, and the members voted it down. I still don't know why they voted it down.

I worked in a wafer fab for a summer and we did four 12 hour days and then three days off. And then three days of work and four off.

It was by far one of my favorite schedules. I rode my bike to work and that was about 45 minutes each way and when you add in the time it takes to get ready for work it was nice to only have to do that 50% of my days. And with four day breaks I was able to really get into projects or go camping for a long time.

I like it. I have half the week where I can goof off and do whatever, and then 4 days of graveyard shifts. It works out quite nicely, especially when paired with the benefits offered. To each their own.

All, thanks so much for the comments. We wrote this up so long ago; it's really interesting to get feedback now.

FWIW, The Infrastructure Observatory is working on more tours like this one, mostly in the NYC area. If you want to get involved (or have an idea for a tour that we should organize), get in touch at http://infraobservatory.com or @NYInfraObserve on Twitter.

-Spencer

Vaguely interesting, as I suppose a similar tour of, say, a Wal-mart or FedEx logistics center might be.

Bonus quote: "Our tour guide used an iPhone, but claimed he’d use a Fire if he wasn’t locked into his carrier."

Does anyone know any good resources/ books on the logistics of setting up and running a fulfillment/distribution facility with the latest tech?

Pretty cool stuff

I wished they would add a diagram, even if it's just hand-drawn. It's would have been much easier to visualize, especially the detailed parts.

Also the naming is sometimes inconsistent across the different authors. e.g. Sometimes folks are called "associates", sometimes "sorters". It got confusing quickly. Would be nice if the naming was clearer

On a side note, there are languages such as French where repeating a word is a proof of bad style (and possibly lack of culture from the author), and others like Swedish (e.g. the Millenium books) where repeating is an appeasing regularity.

British English at least is in the "don't repeat" group.

He doesn't specify exactly how cubbies are selected at the "inbound" stage, beyond saying it's random. I assume it's left up to the associate's own whims and judgements? I imagine the "no identical items in adjacent cubbies" rule improves the potential for picking the same item in parallel while preventing too much concentration, improving availability from different locations. I wonder what other constraints they are given for this stage.

It is indeed quite random. Each stower has a rate, and in order for us to make that rate, we can stow in almost any bin. Some other constraints include using the smallest bin possible for each item and not stowing heavy items up high (the scanner restricts stowing "heavy" items up top).

He also implied that the entire stock of an item isn't in the same place. This makes sense - the computer will tell the picker to get an item from the location that's closest to the rest of the items they're picking at that time.

So, you'd think there would be some logic to how the items were distributed and not just the whim of the associate. (Or maybe that's good enough.)

PDF @ https://docs.google.com/document/export?0=mp&1=ws&2=print&3=...

Looks like initial notes for an IE Analysis. Pretty good for retrospective.

I have no skin in this game and really don't know much about what is going on, but really appreciated that rundown. I've always wanted to know what goes on inside these places.

And something unrelated I've learnt about Google Docs:

"Wow, this file is really popular! Some tools might be unavailable until the crowd clears."

Yep that was interesting. I am part of a Google Apps trial at work, but we don't have the resources to stress test it so I have never seen it under large simultaneous load (seemed to handle it fairly well though).

Reply on news.ycombinator.com