Design Amazon's sales ranking by category feature

Step 1: Outline use cases and constraints

Use cases

Service calculates the past week's most popular products by category
User views the past week's most popular products by category
Service has high availability

Out of scope

The general e-commerce site
- Design components only for calculating sales rank

Constraints and assumptions

State assumptions

Traffic is not evenly distributed
Items can be in multiple categories
Items cannot change categories
There are no subcategories ie foo/bar/baz
Results must be updated hourly
- More popular products might need to be updated more frequently
10 million products
1000 categories
1 billion transactions per month
100 billion read requests per month
100:1 read to write ratio

Calculate usage

Size per transaction:
- created_at - 5 bytes
- product_id - 8 bytes
- category_id - 4 bytes
- seller_id - 8 bytes
- buyer_id - 8 bytes
- quantity - 4 bytes
- total_price - 5 bytes
- Total: ~40 bytes
40 GB of new transaction content per month
- 40 bytes per transaction * 1 billion transactions per month
- 1.44 TB of new transaction content in 3 years
- Assume most are new transactions instead of updates to existing ones
400 transactions per second on average
40,000 read requests per second on average

Handy conversion guide:

2.5 million seconds per month
1 request per second = 2.5 million requests per month
40 requests per second = 100 million requests per month
400 requests per second = 1 billion requests per month

Step 2: Create a high level design

Step 3: Design core components

Use case: Service calculates the past week's most popular products by category

We could store the raw Sales API server log files on a managed Object Store such as Amazon S3, rather than managing our own distributed file system.

We'll assume this is a sample log entry, tab delimited:

timestamp   product_id  category_id    qty     total_price   seller_id    buyer_id
t1          product1    category1      2       20.00         1            1
t2          product1    category2      2       20.00         2            2
t2          product1    category2      1       10.00         2            3
t3          product2    category1      3        7.00         3            4
t4          product3    category2      7        2.00         4            5
t5          product4    category1      1        5.00         5            6
...

The Sales Rank Service could use MapReduce, using the Sales API server log files as input and writing the results to an aggregate table sales_rank in a SQL Database. We should discuss the use cases and tradeoffs between choosing SQL or NoSQL.

We'll use a multi-step MapReduce:

Step 1 - Transform the data to (category, product_id), sum(quantity)
Step 2 - Perform a distributed sort

class SalesRanker(MRJob):

    def within_past_week(self, timestamp):
        """Return True if timestamp is within past week, False otherwise."""
        ...

    def mapper(self, _ line):
        """Parse each log line, extract and transform relevant lines.

        Emit key value pairs of the form:

        (category1, product1), 2
        (category2, product1), 2
        (category2, product1), 1
        (category1, product2), 3
        (category2, product3), 7
        (category1, product4), 1
        """
        timestamp, product_id, category_id, quantity, total_price, seller_id, \
            buyer_id = line.split('\t')
        if self.within_past_week(timestamp):
            yield (category_id, product_id), quantity

    def reducer(self, key, value):
        """Sum values for each key.

        (category1, product1), 2
        (category2, product1), 3
        (category1, product2), 3
        (category2, product3), 7
        (category1, product4), 1
        """
        yield key, sum(values)

    def mapper_sort(self, key, value):
        """Construct key to ensure proper sorting.

        Transform key and value to the form:

        (category1, 2), product1
        (category2, 3), product1
        (category1, 3), product2
        (category2, 7), product3
        (category1, 1), product4

        The shuffle/sort step of MapReduce will then do a
        distributed sort on the keys, resulting in:

        (category1, 1), product4
        (category1, 2), product1
        (category1, 3), product2
        (category2, 3), product1
        (category2, 7), product3
        """
        category_id, product_id = key
        quantity = value
        yield (category_id, quantity), product_id

    def reducer_identity(self, key, value):
        yield key, value

    def steps(self):
        """Run the map and reduce steps."""
        return [
            self.mr(mapper=self.mapper,
                    reducer=self.reducer),
            self.mr(mapper=self.mapper_sort,
                    reducer=self.reducer_identity),
        ]

The result would be the following sorted list, which we could insert into the sales_rank table:

(category1, 1), product4
(category1, 2), product1
(category1, 3), product2
(category2, 3), product1
(category2, 7), product3

The sales_rank table could have the following structure:

id int NOT NULL AUTO_INCREMENT
category_id int NOT NULL
total_sold int NOT NULL
product_id int NOT NULL
PRIMARY KEY(id)
FOREIGN KEY(category_id) REFERENCES Categories(id)
FOREIGN KEY(product_id) REFERENCES Products(id)

We'll create an index on id , category_id, and product_id to speed up lookups (log-time instead of scanning the entire table) and to keep the data in memory. Reading 1 MB sequentially from memory takes about 250 microseconds, while reading from SSD takes 4x and from disk takes 80x longer.

Use case: User views the past week's most popular products by category

The Client sends a request to the Web Server, running as a reverse proxy
The Web Server forwards the request to the Read API server
The Read API server reads from the SQL Database sales_rank table

We'll use a public REST API:

$ curl https://amazon.com/api/v1/popular?category_id=1234

Response:

{
    "id": "100",
    "category_id": "1234",
    "total_sold": "100000",
    "product_id": "50",
},
{
    "id": "53",
    "category_id": "1234",
    "total_sold": "90000",
    "product_id": "200",
},
{
    "id": "75",
    "category_id": "1234",
    "total_sold": "80000",
    "product_id": "3",
},

For internal communications, we could use Remote Procedure Calls.

Step 4: Scale the design

Important: Do not simply jump right into the final design from the initial design!

State you would 1) Benchmark/Load Test, 2) Profile for bottlenecks 3) address bottlenecks while evaluating alternatives and trade-offs, and 4) repeat. See Design a system that scales to millions of users on AWS as a sample on how to iteratively scale the initial design.

It's important to discuss what bottlenecks you might encounter with the initial design and how you might address each of them. For example, what issues are addressed by adding a Load Balancer with multiple Web Servers? CDN? Master-Slave Replicas? What are the alternatives and Trade-Offs for each?

We'll introduce some components to complete the design and to address scalability issues. Internal load balancers are not shown to reduce clutter.

To avoid repeating discussions, refer to the following system design topics for main talking points, tradeoffs, and alternatives:

The Analytics Database could use a data warehousing solution such as Amazon Redshift or Google BigQuery.

We might only want to store a limited time period of data in the database, while storing the rest in a data warehouse or in an Object Store. An Object Store such as Amazon S3 can comfortably handle the constraint of 40 GB of new content per month.

To address the 40,000 average read requests per second (higher at peak), traffic for popular content (and their sales rank) should be handled by the Memory Cache instead of the database. The Memory Cache is also useful for handling the unevenly distributed traffic and traffic spikes. With the large volume of reads, the SQL Read Replicas might not be able to handle the cache misses. We'll probably need to employ additional SQL scaling patterns.

400 average writes per second (higher at peak) might be tough for a single SQL Write Master-Slave, also pointing to a need for additional scaling techniques.

SQL scaling patterns include: