Spring Camp 2016 - List query performance improvement using Couchbase
2 likes1,075 views
The document discusses an asymmetric replication architecture designed to enhance item listing performance using Couchbase and RabbitMQ, aimed at addressing the performance issues of a legacy API at TicketMonster. It outlines the challenges of migrating from a monolithic structure to an asynchronous model with improved data consumption flows while managing over 40 database tables. The implementation results in improved response times, efficient data processing, and simpler query access, showcasing a successful transition from a complex legacy system to a more agile infrastructure.
![{
"id" : 12345678,
"title" : "Awesome Deal",
"price" : 10000,
"foo" : "bar",
...
"meta" : {
"categories" : [111, 222, 333],
"score" : 2.71828,
"type" : "DEAL"
},
...
"is_awesome" : true,
"start_date" : "2015-08-25 00:00:00",
"end_date" : "2016-01-22 00:00:00"
}
Document Data Modeling:](https://image.slidesharecdn.com/springcamp2016-ikimslidev2static-160429032541/85/Spring-Camp-2016-List-query-performance-improvement-using-Couchbase-38-320.jpg)
![function (doc, meta) {
if (doc.meta.type != "DEAL") {
return;
}
var categories = doc.meta.category;
var start_date = Number(doc.start_date.split(' ')[0].replace(/-/g, ''));
for (var i = 0; i < categories.length; i++) {
emit([categories[i], start_date, doc.id], [doc.is_awesome]);
}
}
Couchbase View - Map Function:](https://image.slidesharecdn.com/springcamp2016-ikimslidev2static-160429032541/85/Spring-Camp-2016-List-query-performance-improvement-using-Couchbase-39-320.jpg)
![import json, httplib
from gcouchbase.bucket import Bucket
template = '/****list/list?cat*=%s&page=%s&page_size=50'
http = httplib.HTTPConnection('****.ticketmonster.co.kr', timeout=500)
bucket = Bucket('couchbase://localhost/prototype', password='****')
sliced = [[…], […],…]
def run_each_slice(sliced_categories):
for category in sliced_categories:
page, total = 0, 0
while(True):
http.request('GET', template % (category, page))
data = http.getresponse().read()
http.close()
for deal in json.loads(data)['lists'][0]['deals']:
bucket_work('deal_replica::%d' % deal['id'], deal, category)
for each in sliced:
run_each_slice(each)
Data Scraper Script(199 LoC):](https://image.slidesharecdn.com/springcamp2016-ikimslidev2static-160429032541/85/Spring-Camp-2016-List-query-performance-improvement-using-Couchbase-42-320.jpg)
![import json
from gcouchbase.bucket import Bucket
bucket = Bucket('couchbase://localhost/prototype', password='****')
results = bucket.query('deals', ‘date_ordered_view',
limit = 100,
skip = 0,
mapkey_range = [[12345, '99999999'], [12345, '00000000']],
descending = True,
include_docs = True,
reduce = False
)
for item in results:
print item
List Query Script(22 LoC):](https://image.slidesharecdn.com/springcamp2016-ikimslidev2static-160429032541/85/Spring-Camp-2016-List-query-performance-improvement-using-Couchbase-44-320.jpg)
Spring Camp 2016 - List query performance improvement using Couchbase
- 1. Asymmetric Replication Architecture Improve Item Listing Performance Using Couchbase, With Asynchronous Replication Approach Intae Kim | Technical Lead/Architect, Ticketmonster
- 2. Company Information: • Service Name: TMON • #1 Social Commerce in Korea • 10,000,000+ user • NOW HIRING(?!) TICKET MONSTER
- 3. let the game begin
- 4. porting legacy API without changing master databases, introducing NoSQL(Couchbase) & Message Queue(RabbitMQ) to enhance data consumption flow abstract:
- 5. Agenda: • Open The Pandora’s Box • Think Outside The Box • Prototyping The Box • Implementing Brand New Box • Retrospective; Wrap It Up
- 6. Open The Pandora’s Box Most BEAUTIFUL Code Ever Seen… (Are You Serious?)
- 7. Mission: • Porting Current PHP into Java • + Migrate Category Mappings(!) • + Resolve Hard-Coded(!!)
- 9. item list
- 10. • title, image, price • discount • sales count • categories • pagination • sort order • tags • …
- 11. but…
- 15. List API - Call Tree:
- 16. List API - Complexity: • 2000+ function call • 20+ call tree depth • cyclic dependency • NO history, NO documentation, … • …
- 17. welcome to hell
- 18. List API - Performance: • 120+ sec response duration1) • 40+ table, 10+ RPC reference • table joins, OUTER joins • O(N4) loop statement • … 1) without cache
- 19. List API - etc: • promiscuous cache coating • side effect friendly functions • black box parameters • lying/obsolete comments • …
- 20. a nightmare
- 21. Think Outside The Box Asynchronous Data Replication & Simple Data Read
- 22. Pandora’s Box Architecture: user service 40+ tables admin console client slave (MySQL)master (MySQL) adjustment C/R/U/D update “list” order fetch “list” 10+ RPCs
- 23. Pandora’s Box: • Typical Web Application • Homogeneous, Symmetric Master - Slave • Synchronous • Primeval Cache • Monolith
- 24. API Result, Mixed Up 40+ Tables:
- 26. Asymmetric Box Architecture: C/R/U/D user service SIMPLE query replication40+ tables admin console NoSQL replica (Couchbase) item “list” composite json client item data master (MySQL) adjustment purchase data 10+ RPCs
- 27. Asymmetric Box: • Heterogeneous, Asymmetric Master - Slave • Asynchronous • Parallel Data Processing • NO Join • NO Cache • MSA-like
- 28. The Key: • READ: frequent, ~109 request / day • WRITE: rare, ~ 106 request / day
- 29. Prototyping The Box Concept Verification for Data Change & On-Demand Listing
- 30. List Query on Legacy: user service 40+ tables client slave (MySQL) admin console master (MySQL) adjustment C/R/U/D update “list” order fetch “list” 10+ RPCs On-Demand List Calculation
- 31. C/R/U/D replication40+ tables admin console composite json master (MySQL) item data adjustment purchase data List Query on Couchbase: user service SIMPLE Query NoSQL replica (Couchbase) client
- 32. The Difference: 40+ tables 10+ RPCs On-Demand Calculation with 40+ Tables, 10+ RPCs Simple, Single Table Access Legacy - User Request Part Alternative - User Request Part
- 33. Replication Datastore Candidates: • redis • MongoDB • Couchbase • RDBMS
- 34. why couchbase?
- 35. Struggle #1 - NoSQL: • how to query for sorted, filtered? • how to design data modeling?
- 36. Couchbase “View Query”: • Map-Reduce • Eventually Consistent • B-Tree • Functional
- 41. Data Scraping for Prototyping: Couchbase Scrape, Scrape, Scrape, …
- 43. Prototyping List Query: Couchbase Fetch “ORDERED” List with: - offset - limit - CATEGORY
- 45. ready to rock n’ roll
- 46. Implementing Brand New Box Asynchronous Data Replication
- 47. Struggle #2 - Read Legacy: • an epic • associative array abuse • fickle parameters
- 48. 2 weeks to read
- 49. Skill Set: • Java, Python, Javascript, PHP • RxJava, jOOQ, SpringMVC, Jersey • Flask, gevent, greenlet • Couchbase, MySQL, RabbitMQ, Tomcat
- 50. Methodology: • Test Driven Development • Pseudo-Agile Process
- 53. Retrospective; Wrap It Up So Far, So Good… So What?
- 55. CENSORED 48ms
- 57. simple & clean code
- 60. Measurement Compare #1/3 - LoC:
- 61. Measurement Compare #2/3 - Debt:
- 62. Measurement Compare #3/3 - Issues:
- 63. is it all?
- 64. what made the monster?
- 65. (…bragging…)