MongoNYC 2012: E-Commerce at UnderArmour, Inc.

E-Commerce at UnderArmour __________ Jesse Dailey [email protected] github/jldailey

The Caveat This presentation discusses a closed-source application owned by
a publicly-traded company, so code snippets are pseudo-code and all dates, statistics, and numbers are only my best personal estimates.

What is this talk about? And not about?

Goals: for 2011 1. Act quicker, and sell more. ◦
Get developers out of the way. 2. Empower authors to control Publishing. ◦ Preview ◦ Scheduling ◦ Rollback 3. Test any thing. ◦ Deeply integrated AB Testing.

The Plan • Divide the existing application into three sections:
◦ the "runtime", ◦ "author time", ◦ and the "service" layer. • What we are showing today ◦ The new runtime, a CMS designed for MongoDB. ◦ The new author time UI, done in Flex.

MongoDB

Batches Publishing: • Preview • Scheduled Activation • Rollback Every
piece of content is associated with an integer batch number, and batches are scheduled to begin and end at certain times.

Batches: how-to Documents in Mongo that are batch-aware have a
"batch" attribute, a single integer. To find the document with the highest active version: var doc = collection("pages").find({ ... real criteria first ..., $in: { batch: timeslice } }) .sort({ batch: -1 }) .limit(1); There is a more elegant way to use the timeslice that we will save for later, if someone asks about it in Q&A.

Content A tree of objects, stored in BSON documents. Every
node in the tree has a "type" attribute. var page = { type: "page", body: "Hello World" }; Each "type" maps to some handler code. reducer["page"] = function(node) { return [ "<html><head>",node.head, "</head><body>",node.body, "</body></html>" ]; } Two-phase rendering.

reduce: the first phase. A central recursive function, • evaluates
the handler for each node • output accumulates output as a (larger, but simpler) tree. This is why the previous example was: return [ "string", object, "string" ]; instead of return "string" + object + "string"; At the end of reduction, it is still "turtles all the way down".

finalize: the second phase. Results in a single string. A
second recursive function, • joins simple elements as strings. • invokes finalize on nodes that survived reduction. Next, an example of a node that survives to be finalized, • { type: "region" }

Regions: out-of-order rendering. A region lets you: • label a
position in the output. • each label gets a content buffer. • during reduction, other nodes contribute to these buffers. • during finalization, the buffer is joined. { type: page, head: ["<style>",{ type:region, name:"css" },"</style>"] body: [ "<div class='demo'>Hello World</div>", { type: region, action: "add", name: "css", content: ".demo { color: red; }" } ] }

"script" and "style": out-of-order function script(node) { return { type:"region",
action:"add", name:"js", content: node.text } }

Dependencies: scripts {type: "script", src: "my-jquery-plugin.js", depends: "jquery" } {type:
"script", src: "jquery-1.6.min.js", provides: "jquery" } • reduce builds a dependency tree. • finalize renders tree to $script (into the "js" region).

Dependencies: application logic { type: "anything", require: "tax-calculated" } Here,
"tax-calculated" is the name of a Processor. A Processor defines two methods: • pre() • post() Each method is called once in any request where it is required.

AB Testing • The code "homepage-test" gets sent to the
mPath Learning API from Conductrics. { type: "abtest", code: "homepage-test", items: { A: { ... } B: { ... } } }

Programming { type: "data", key: "FOO" } { type: "each",
of: [ ... ] as: { ... } } { type: "with", person: { ... } as: [ "Hello ", { type: "data", key: "PERSON.NAME"} ] } { type: "if-data" } { type: "case-data" }, etc...

Caching { type: "cached", key: { type: "list-join", items: [
{type:"data", key:"PARENTKEY"}, "my-key-component" ] }, content: { ... } }

What makes this powerful? Why bother?

5-minute UI video

Event Queue w/ Blind Locking Goal: Cluster-wide publish / subscribe
event queue. Strategy: • Capped Collection ◦ maintains insertion order cheaply. ◦ can't delete, just mark. • Readers of the Queue must use "Blind Locking"

Blind Locking 1. Each document in the capped collection gets:
◦ { lock: 'unlocked' } 2. Each Reader is assigned/generated an 8-digit token: ◦ 'reader01' 3. To pop an event off the Queue safely under load: ◦ Update the first 'unlocked' lock to 'reader01'. ◦ Find the first document with 'reader01'.

Reading from an Event Queue. 1: db.event.queue.update( 2: { lock:
'unlocked' }, 3: { lock: 'reader01' }); 4: 5: doc = db.event.queue.findOne( 6: { lock: 'reader01' }); 7: 8: if( doc != null ) { 9: doc.status = invokeEventHandlers(doc); 10: doc.lock = 'finished'; 11: db.event.queue.save(doc); 12: }

Acknowledgements (aka Wall of Email) UnderArmour, Inc - Dev Team
Brian Massey <[email protected]> Brian "Magic" Johnston <[email protected]> Jeff "Switzerland" Switzer <[email protected]> Ken Valencik, Ron Wilson, and others. Twin Technologies, LLC Ben Elmore, CEO <[email protected]> Abraham Lloyd, Solutions Exec. <[email protected]> Conductrics, LLC Nate Weiss, CTO <[email protected]> Jesse Dailey <[email protected]>

Thank You Three Talks that inspire (and aren't on TED):
1. Bret Victor, "Inventing on Principle": ◦ youtu.be/PUv66718DII 2. Jaron Lanier, "Learning by Experience & Play": ◦ youtu.be/F9eFZpdSeRU 3. Alan Kay, "Normal considered Harmful": ◦ youtu.be/FvmTSpJU-Xc

MongoNYC 2012: E-Commerce at UnderArmour, Inc.

MongoNYC 2012: E-Commerce at UnderArmour, Inc.

mongodb

More Decks by mongodb

Featured

Transcript

E-Commerce at UnderArmour __________ Jesse Dailey [email protected] github/jldailey

The Caveat This presentation discusses a closed-source application owned by

What is this talk about? And not about?

Goals: for 2011 1. Act quicker, and sell more. ◦

The Plan • Divide the existing application into three sections:

MongoDB

Batches Publishing: • Preview • Scheduled Activation • Rollback Every

Batches: how-to Documents in Mongo that are batch-aware have a

Content A tree of objects, stored in BSON documents. Every

reduce: the first phase. A central recursive function, • evaluates

finalize: the second phase. Results in a single string. A

Regions: out-of-order rendering. A region lets you: • label a

"script" and "style": out-of-order function script(node) { return { type:"region",

Dependencies: scripts {type: "script", src: "my-jquery-plugin.js", depends: "jquery" } {type:

Dependencies: application logic { type: "anything", require: "tax-calculated" } Here,

AB Testing • The code "homepage-test" gets sent to the

Programming { type: "data", key: "FOO" } { type: "each",

Caching { type: "cached", key: { type: "list-join", items: [

What makes this powerful? Why bother?

5-minute UI video

Event Queue w/ Blind Locking Goal: Cluster-wide publish / subscribe

Blind Locking 1. Each document in the capped collection gets:

Reading from an Event Queue. 1: db.event.queue.update( 2: { lock:

Acknowledgements (aka Wall of Email) UnderArmour, Inc - Dev Team

Thank You Three Talks that inspire (and aren't on TED):