{"id":"design-sales-rank","title":"Design Amazon's sales rank by category feature","sections":[{"id":"overview","title":"Overview","html":"

Note: This document links directly to relevant areas found in the system design topics to avoid duplication. Refer to the linked content for general talking points, tradeoffs, and alternatives.

","headings":[]},{"id":"step-1-outline-use-cases-and-constraints","title":"Step 1: Outline use cases and constraints","html":"

Step 1: Outline use cases and constraints

\n
\n

Gather requirements and scope the problem.\nAsk questions to clarify use cases and constraints.\nDiscuss assumptions.

\n
\n

Without an interviewer to address clarifying questions, we'll define some use cases and constraints.

\n

Use cases

\n

We'll scope the problem to handle only the following use case

\n\n

Out of scope

\n\n

Constraints and assumptions

\n

State assumptions

\n\n

Calculate usage

\n

Clarify with your interviewer if you should run back-of-the-envelope usage calculations.

\n\n

Handy conversion guide:

\n","headings":[{"id":"use-cases","title":"Use cases","depth":3},{"id":"well-scope-the-problem-to-handle-only-the-following-use-case","title":"We'll scope the problem to handle only the following use case","depth":4},{"id":"out-of-scope","title":"Out of scope","depth":4},{"id":"constraints-and-assumptions","title":"Constraints and assumptions","depth":3},{"id":"state-assumptions","title":"State assumptions","depth":4},{"id":"calculate-usage","title":"Calculate usage","depth":4}]},{"id":"step-2-create-a-high-level-design","title":"Step 2: Create a high level design","html":"

Step 2: Create a high level design

\n
\n

Outline a high level design with all important components.

\n
\n

\"Diagram\"

","headings":[]},{"id":"step-3-design-core-components","title":"Step 3: Design core components","html":"

Step 3: Design core components

\n
\n

Dive into details for each core component.

\n
\n\n

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.

\n

Clarify with your interviewer how much code you are expected to write.

\n

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

\n
timestamp   product_id  category_id    qty     total_price   seller_id    buyer_id\nt1          product1    category1      2       20.00         1            1\nt2          product1    category2      2       20.00         2            2\nt2          product1    category2      1       10.00         2            3\nt3          product2    category1      3        7.00         3            4\nt4          product3    category2      7        2.00         4            5\nt5          product4    category1      1        5.00         5            6\n...
\n

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.

\n

We'll use a multi-step MapReduce:

\n\n
class SalesRanker(MRJob):\n\n    def within_past_week(self, timestamp):\n        \"\"\"Return True if timestamp is within past week, False otherwise.\"\"\"\n        ...\n\n    def mapper(self, _ line):\n        \"\"\"Parse each log line, extract and transform relevant lines.\n\n        Emit key value pairs of the form:\n\n        (category1, product1), 2\n        (category2, product1), 2\n        (category2, product1), 1\n        (category1, product2), 3\n        (category2, product3), 7\n        (category1, product4), 1\n        \"\"\"\n        timestamp, product_id, category_id, quantity, total_price, seller_id, \\\n            buyer_id = line.split('\\t')\n        if self.within_past_week(timestamp):\n            yield (category_id, product_id), quantity\n\n    def reducer(self, key, value):\n        \"\"\"Sum values for each key.\n\n        (category1, product1), 2\n        (category2, product1), 3\n        (category1, product2), 3\n        (category2, product3), 7\n        (category1, product4), 1\n        \"\"\"\n        yield key, sum(values)\n\n    def mapper_sort(self, key, value):\n        \"\"\"Construct key to ensure proper sorting.\n\n        Transform key and value to the form:\n\n        (category1, 2), product1\n        (category2, 3), product1\n        (category1, 3), product2\n        (category2, 7), product3\n        (category1, 1), product4\n\n        The shuffle/sort step of MapReduce will then do a\n        distributed sort on the keys, resulting in:\n\n        (category1, 1), product4\n        (category1, 2), product1\n        (category1, 3), product2\n        (category2, 3), product1\n        (category2, 7), product3\n        \"\"\"\n        category_id, product_id = key\n        quantity = value\n        yield (category_id, quantity), product_id\n\n    def reducer_identity(self, key, value):\n        yield key, value\n\n    def steps(self):\n        \"\"\"Run the map and reduce steps.\"\"\"\n        return [\n            self.mr(mapper=self.mapper,\n                    reducer=self.reducer),\n            self.mr(mapper=self.mapper_sort,\n                    reducer=self.reducer_identity),\n        ]
\n

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

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

The sales_rank table could have the following structure:

\n
id int NOT NULL AUTO_INCREMENT\ncategory_id int NOT NULL\ntotal_sold int NOT NULL\nproduct_id int NOT NULL\nPRIMARY KEY(id)\nFOREIGN KEY(category_id) REFERENCES Categories(id)\nFOREIGN KEY(product_id) REFERENCES Products(id)
\n

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.<a href=https://github.com/donnemartin/system-design-primer#latency-numbers-every-programmer-should-know>1

\n\n\n

We'll use a public REST API:

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

Response:

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

For internal communications, we could use Remote Procedure Calls.

","headings":[{"id":"use-case-service-calculates-the-past-weeks-most-popular-products-by-category","title":"Use case: Service calculates the past week's most popular products by category","depth":3},{"id":"use-case-user-views-the-past-weeks-most-popular-products-by-category","title":"Use case: User views the past week's most popular products by category","depth":3}]},{"id":"step-4-scale-the-design","title":"Step 4: Scale the design","html":"

Step 4: Scale the design

\n
\n

Identify and address bottlenecks, given the constraints.

\n
\n

\"Diagram\"

\n

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

\n

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.

\n

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?

\n

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

\n

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

\n\n

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

\n

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.

\n

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.

\n

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.

\n

SQL scaling patterns include:

\n\n

We should also consider moving some data to a NoSQL Database.

","headings":[]},{"id":"additional-talking-points","title":"Additional talking points","html":"

Additional talking points

\n
\n

Additional topics to dive into, depending on the problem scope and time remaining.

\n
\n

NoSQL

\n\n

Caching

\n\n

Asynchronism and microservices

\n\n

Communications

\n\n

Security

\n

Refer to the security section.

\n

Latency numbers

\n

See Latency numbers every programmer should know.

\n

Ongoing

\n","headings":[{"id":"nosql","title":"NoSQL","depth":4},{"id":"caching","title":"Caching","depth":3},{"id":"asynchronism-and-microservices","title":"Asynchronism and microservices","depth":3},{"id":"communications","title":"Communications","depth":3},{"id":"security","title":"Security","depth":3},{"id":"latency-numbers","title":"Latency numbers","depth":3},{"id":"ongoing","title":"Ongoing","depth":3}]}],"sourceFiles":[{"name":"sales_rank_mapreduce.py","url":"https://github.com/donnemartin/system-design-primer/blob/master/solutions/system_design/sales_rank/sales_rank_mapreduce.py","html":"
# -*- coding: utf-8 -*-\n\nfrom mrjob.job import MRJob\n\n\nclass SalesRanker(MRJob):\n\n    def within_past_week(self, timestamp):\n        \"\"\"Return True if timestamp is within past week, False otherwise.\"\"\"\n        ...\n\n    def mapper(self, _, line):\n        \"\"\"Parse each log line, extract and transform relevant lines.\n\n        Emit key value pairs of the form:\n\n        (foo, p1), 2\n        (bar, p1), 2\n        (bar, p1), 1\n        (foo, p2), 3\n        (bar, p3), 10\n        (foo, p4), 1\n        \"\"\"\n        timestamp, product_id, category, quantity = line.split('\\t')\n        if self.within_past_week(timestamp):\n            yield (category, product_id), quantity\n\n    def reducer(self, key, values):\n        \"\"\"Sum values for each key.\n\n        (foo, p1), 2\n        (bar, p1), 3\n        (foo, p2), 3\n        (bar, p3), 10\n        (foo, p4), 1\n        \"\"\"\n        yield key, sum(values)\n\n    def mapper_sort(self, key, value):\n        \"\"\"Construct key to ensure proper sorting.\n\n        Transform key and value to the form:\n\n        (foo, 2), p1\n        (bar, 3), p1\n        (foo, 3), p2\n        (bar, 10), p3\n        (foo, 1), p4\n\n        The shuffle/sort step of MapReduce will then do a\n        distributed sort on the keys, resulting in:\n\n        (category1, 1), product4\n        (category1, 2), product1\n        (category1, 3), product2\n        (category2, 3), product1\n        (category2, 7), product3\n        \"\"\"\n        category, product_id = key\n        quantity = value\n        yield (category, quantity), product_id\n\n    def reducer_identity(self, key, value):\n        yield key, value\n\n    def steps(self):\n        \"\"\"Run the map and reduce steps.\"\"\"\n        return [\n            self.mr(mapper=self.mapper,\n                    reducer=self.reducer),\n            self.mr(mapper=self.mapper_sort,\n                    reducer=self.reducer_identity),\n        ]\n\n\nif __name__ == '__main__':\n    SalesRanker.run()
"}]}