Chapter 15 checkpoint

Chapter 15 checkpoint: sales rank by category

Every question from this chapter’s lessons. Your first answer to each is the one that counts; afterward you can practice again as often as you like. Passing needs 75% right on the first attempt.

0 of 17 answered0 right on the first attempt

Question 1 of 17

The service receives 1 billion purchase transactions a month. In a 30-day month, how many transactions per second is that on average?

Your answerTPS

Answers within 5% count.

Question 2 of 17

Each transaction log line is about 40 bytes, and there are 1 billion a month. How much new data is that per month, in GB?

Your answer

Answers within 5% count, in any unit you choose.

Question 3 of 17

Using 400 requests per second ≈ 1 billion a month, how many reads per second do 100 billion reads a month average out to?

Your answerreads/s

Answers within 5% count.

Question 4 of 17

The ranking itself, the top 1,000 products in each of 1,000 categories, is only a million rows. So what is actually hard about this feature?

Ranking subcategoriesProducts changing categoryStoring the million rowsComputing it from a billion transactions a month, and serving 40,000 reads a second

Question 5 of 17

Why does recomputing the ranking from the raw log in an hourly batch job give trustworthy results?

Batch jobs update rankings faster than streamsIt needs no storageThe log is complete and immutable, so every transaction is counted, late ones included, and runs can be repeatedIt ranks every purchase within a second

Question 6 of 17

Match each stream-processing term to its meaning.

Tumbling windowSliding windowWatermarkAllowed lateness

Question 7 of 17

The streaming counts for a window are a little lower than the hourly batch recount of the same window. What is the likely cause?

Late events arrived after the watermark passed and were droppedThe CDN is caching old countsThe batch job counts every purchase twiceTumbling windows overlap

Question 8 of 17

Purchases are weighted by

e^{-\lambda \cdot \text{age}}

with a half-life of 24 hours. How much does a sale from 48 hours ago count?

Your answer

Answers within 2% count.

Question 9 of 17

In Kreps’s Kappa architecture, how do you reprocess data after fixing a bug?

It cannot; streams only see new dataEdit the stored results by handRun a second instance of the stream job from the start of the retained log, into a new tableRun a separate batch system in parallel

Question 10 of 17

Why do Redis sorted sets suit a continuously updated leaderboard?

They use a B-tree, like SQL indexesThey recompute the ranking on every readA hash table and a skip list give fast score updates and fast ranked rangesThey keep members in a sorted array

Question 11 of 17

You keep the top 1,000 products for each of 10,000 categories in sorted sets. Assuming about 128 bytes per member, how much RAM is that, in GB?

Your answer

Answers within 5% count, in any unit you choose.

Question 12 of 17

How does a Redis skip list decide how many levels a member’s node has?

Every node is on every levelAt random when the node is inserted, independent of its scoreHigher scores get more levelsBy the member’s current rank

Question 13 of 17

Match each sorted-set command to its cost, for

N

members and

M

returned.

ZSCOREZINCRBYZRANGE … REV for the top

M

Question 14 of 17

A category page is requested 2 times a second, and each miss caches it for 60 seconds. Using

\frac{\lambda T}{1 + \lambda T}

, what is its hit ratio, as a percentage?

Your answer%

Answers within 0.5% count.

Question 15 of 17

A rarely viewed category gets 0.01 requests per second, with the same 60-second cache. What is its hit ratio, as a percentage?

Your answer%

Answers within 2% count.

Question 16 of 17

On a big shopping day, millions of shoppers open the Electronics ranking, and all its reads land on the one Redis node that holds it. What relieves that node?

Recompute the ranking more oftenCache the page in front of Redis (CDN, in-process) and add Redis replicas for readsRead directly from the databaseAdd more Redis shards

Question 17 of 17

A hot category’s writes are split so that each shard holds partial counts for every product. The top 10 is built by merging each shard’s top 10, and sometimes it names the wrong winner. Why?

Replication lag between shardsA product can be outside every shard’s top 10 yet first overall; sum counts first, then rankTies are broken differently on each shardEach shard should return its top 11