Ebay Auction
Topic: Ebay Auction
Interviewer: Tom (慢慢失败)
Interviewee: shihao (沉)
Level: L5 (Senior)
Topic
Mock System Design Interview Summary
Interview Overview
Date: 1/9/2022
Target level: L5
Duration: 1 hour
Topic covered: Ebay Auction
Drawing tool used: Draw.io/Diagram
Requirements
Functional requirements
Can create products
Can have auctions / bid
Keywords and products
No cheating
Seller report lower PR
User and order service available
Interviewer:
Focus on auction
Interviewee:
Change auction begin and end should be prevented 1 hour before auction end
Interviewer:
Assume auction is immutable
Non-functional requirements
DAU 100M 10% 10 million/day
read/write ratio 88:2:10 readproduct/createproduct/makebid
QPS: 10k/s -> 30k/s
Machine: 1 container ~ 1k QPS. 30 replication
(QPS based on experience)
Data storage: 2 million product/day -> 10KB -> 2GB/day
System Design
System design diagram
External APIs
/products ->
GET /products/
DEL /products/
POST /products
GET /products?keyword=<key_word>&page_size=…
/auctions
POST /auctions?product_id=<product_id>
System Design Diagram
3 layers:
Stateless service layer. Service mesh architecture. Istio and envoy,
Met down, logging, tracing, rate-limiting
Caching layer
Persistence layer
Database schema
Auction
System design diagram
Interviewee: will design search next.
Interviewer: please move on to how user makes a bid
GET /auctions?product_id
Schema
Database choice
Consistency
Hbase
How to handle auction in more details?
Key is to handle high throughput
“Local cache”: cached max. Can be used to reject bids
Redis-cluster: global max
Interviewer: who updates the redis-cluster?
Redis-cluster:
What happens if lots of requests with $10
Reconsolidation service in the end, take the earliest $10
Mutable? Immutable? How to present to the user?
Discussions during the Interview
1 user makes a bid
2 compared against the local cache to ensure it’s higher than cached max
3 insert bid to HBase
4 async process “auction reconciliation” will update global max at redis
Flow for auction:
Atomicity operation:
Compare and swap
Interviewer and Audience Feedback after the Interview
Interviewer:
Strong skill, strong know
炫技
We may not need to guarantee strong consistency
Solution - is not very convincing
Improve the diagrams: present the workflow
Hard to see
Audience
Auction - ebay 产品
Some auctions are different
Close envelope price
Bidding, vs auction
Audience
Gathering Requirement
Lost in functional requirement
Design ebay platform
Suggest to interviewer: which requirement is most important
Requirement 可能走偏
Looks like the interviewee knows the question ahead of time
The interviewee didn’t clarify which feature is the main feature
没有突出重点
For example, how to handle multiple biddings at the same time
Ideally:
Start with user,
search engine, create product,
Interviewer
Functional requirement is fine
Don’t understand why we go into search too much
Audience:
Basic is better
Implement the basic
Try to find key flow
“If we have more time, we can discuss other flows”
Can ask requirement in more details
User can create product, lowest bidding price
What happens if 2 users have the same price
When searching: should I specify a time range?
QPS:
Where we want to use a message queue, where we want to use cache. Tradeoff
Detail design: which one is cassandra, dynamo, memory cache: cassandra
Interviewee
description, dynamic schema: no sql (instead of sql)
Update rather low: cassandra or dynamo
Single line insert
LSM tree
HBase
Audience: sql. Rock-database
Hardskill
Redis, nosql database
Redis: sorted set
Cassandra: price can be a key
Consistency: not in requirement gathering
Is there a consistency issue?
Different people/geography may see different highest price
Multiple db globally
Vs single db?
US vs Australia
Is it global service, or regional service?
Audience
Redis: sorted set-> can it solve consistency?
Audience
No. It depends on where you put Redis. Usually there are multiple instances of Redis.
Redis cluster is usually in one availability zone
Redis cluster: globally. New York, Australia, Paris. There is a long delay.
High consistency globally. Redis may not be able to achieve it.
Global consistency: is difficult
May clarify in requirement whether consistency is required
Which are the key test points?
How to ensure buyers to see the latest price? Pull vs push
Concurrent bid? How to handle
Bid legitimacy
How to summarize highest bid
High concurrency
===
There is no workable solution?
Share a workable solution.
===
Interviewee:
I initially thought the key test point is to protect against high throughput
===
Alternative design:
Put message queue to buffer the incoming bids in front of the HBase
Check the current highest price.
Can use redis to ensure the highest bid. Don’t enter message queue
Handle redis crash through backup - may
Then send to message queue
Append only - doesn’t require lock
Leaderboard is only cassandra
Why need to guarantee FIFO?
Likely 2 people bid the same price. FIFO
Message queue is fine. Redis does not need to be most up to date.
#Bidding product -> global max -> message queue -> streaming process ->
bid –> MQ –> Stream process –> MQ –> Worker –> DB
===
Workable solution:
Message queue
Update cache value. Just delete the cache value.
Every 2 seconds update the max in redis
HBase: source of truth
Eventual consistency
Audience
Key improvements
Web Socket to get the latest
Flink, message queue, get global max, send to message queue
Worker
P50 1 second: bridge
Stream process -> distributed, can spin up more copies