Chapter 10 checkpoint

Chapter 10 checkpoint: Twitter timeline and search

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

Users post 500 million tweets a day, and each is delivered to 10 followers on average. About how many deliveries per second is that?

Your answerdeliveries/s

Answers within 5% count.

Question 2 of 17

Each home-timeline entry is 17 bytes. Keeping 800 entries for each of 100 million active users, how much RAM is that before replication, in TB?

Your answer

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

Question 3 of 17

Twitter replicated each timeline entry on 3 machines. How much RAM does the 1.36 TB of timelines need once replicated, in TB?

Your answer

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

Question 4 of 17

Tweets average 10 KB including media, and 500 million arrive each day. How much new content accumulates in 3 years, in PB?

Your answer

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

Question 5 of 17

What is the main trade-off between fan-out on write (push) and fan-out on read (pull)?

Push is strongly consistent and pull is notThey use the same memory; only latency differsPush does the work per follower at write time; pull does it per followed account at read timePush makes reads slow; pull makes writes slow

Question 6 of 17

How does the hybrid design deal with accounts that have millions of followers?

It skips fan-out for them and merges their recent tweets in when timelines are readIt switches every user to fan-out on readIt pushes their tweets faster with more workersIt delivers their tweets only to active followers

Question 7 of 17

Followers of a celebrity sometimes see a reply to the celebrity’s tweet before the tweet itself. Why, and what mitigates it?

The search index reordered themThe celebrity’s fan-out takes minutes, so the reply lands first; re-order at serve timeThe reply’s author has a fast clockThe celebrity’s tweet missed the cache

Question 8 of 17

Match each timeline strategy to the work it does.

Fan-out on writeFan-out on readHybrid

Question 9 of 17

A user returns after 40 days away, and the cache holds timelines only for users active in the last 30 days. What does the primer do?

Show an empty timeline until new tweets arriveRun a search for each followed accountKeep every user’s timeline in the cache foreverRebuild their timeline on demand from the user graph and the SQL database

Question 10 of 17

Snowflake gives each generator a 12-bit sequence per millisecond. At most how many IDs can one generator produce per second?

Your answerIDs/s

Answers within 1% count.

Question 11 of 17

How does Snowflake make IDs that are roughly time ordered without a central counter?

It generates random 64-bit numbersA central database hands out increasing numbersIt guarantees exact time order across all generatorsIt puts a millisecond timestamp in the high bits, then a machine id and a sequence

Question 12 of 17

You shard the tweet store by user_id. What do you gain and what do you risk?

Cross-shard joins become freeNothing changes, because the key does not matterPerfectly even writes, but slower timelinesA user’s timeline is one-shard; very active users can make a hot shard

Question 13 of 17

After an NTP correction moves a server’s clock back by two seconds, its Snowflake generator stops returning IDs for about two seconds. Why?

It refuses to generate IDs until the clock passes the last time it usedTwo generators have the same machine idThe 41-bit timestamp has run outIts 12-bit sequence has overflowed

Question 14 of 17

Earlybird hash-partitions tweets across its servers. Why must each search query go to every server?

Every server holds a full copy of the indexTo spread load evenlyEach server holds one time rangeA matching tweet could be on any shard

Question 15 of 17

Put the steps of the primer’s search request in order.

Send the query to every server in the search cluster
Parse and tokenize the query, normalizing terms
Return the results to the user
Each server returns its best matches
Merge, rank and sort the results

Question 16 of 17

The posting list for “raft” is [1042, 1089, 2104], and for “consensus” it is [1042, 1520, 2104]. Which documents match raft AND consensus?

1042 only1042, 1089, 1520 and 21041042 and 2104None

Question 17 of 17

Each of 8 shards returns its top 10 matches, and the front end merges them. Assume a document scores the same whichever shard holds it. Is the merged top 10 the same as searching one big index would give?

Yes: anything in the global top 10 is in its own shard’s top 10No: it is only an approximationOnly if each shard returns its top 80Only if all matches are on one shard