Chapter 14 checkpoint

Chapter 14 checkpoint: a distributed key-value cache

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

Each cache entry is 270 bytes, and there are 10 billion queries a month. If every query were unique and kept, how much would a month add, in TB?

Your answer

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

Question 2 of 17

Using 400 requests per second ≈ 1 billion a month, how many requests per second do 10 billion queries a month average out to?

Your answerRPS

Answers within 5% count.

Question 3 of 17

A client sends one Redis command, waits for the reply, then sends the next, over a link with a 0.5 ms round-trip time. At most how many commands per second can it complete?

Your answercommands/s

Answers within 2% count.

Question 4 of 17

What does Redis pipelining change?

The commands run as one atomic transactionThe client sends many commands before reading their repliesCommands are compressed on the networkThe server runs the commands on many threads

Question 5 of 17

Which data structure does the primer’s LRU cache use?

A sorted array of keysA B-treeA min-heap of access timesA hash table plus a doubly linked list

Question 6 of 17

A Memcached server’s workload switched from small items to larger ones. Now it evicts recent large items constantly, even though plenty of memory sits in pages assigned to the small-item class. What is happening, and what fixes it?

The growth factor is too largeLRU is evicting the wrong items; switch to LFUSlab calcification; move pages between classes with slab reassign or automoveA memory leak; restart the server

Question 7 of 17

With Memcached’s default growth factor of 1.25, about what fraction of a chunk can be wasted in the worst case, as a percentage?

Your answer%

Answers within 5% count.

Question 8 of 17

A 1 MB Memcached page (1,048,576 bytes) belongs to the class with 120-byte chunks. How many chunks does it hold?

Your answerchunks

Answers within 1% count.

Question 9 of 17

Why is a Redis string of up to 44 bytes cheaper to store than a longer one?

It is stored with its header in a single allocation (embstr)It goes into Redis’s smallest slab classShort strings have no headerRedis compresses short strings

Question 10 of 17

Keys are placed with

\text{hash}(k) \bmod N

over 10 servers. One fails, leaving 9. What percentage of keys move to a different server?

Your answer%

Answers within 3% count.

Question 11 of 17

A cache uses consistent hashing with many points per server. Growing from 10 servers to 12, about what percentage of keys move?

Your answer%

Answers within 5% count.

Question 12 of 17

Why does libketama hash each server to 100–200 points on the ring instead of one?

To make lookups fasterSo each server owns many small arcs, evening out the loadTo hide which server holds a keySo each key is stored on 100 servers

Question 13 of 17

In Redis Cluster you need one multi-key command over a user’s profile and orders. How should you name the keys?

user:1001:profile and user:1001:ordersStore both in one key named user:1001Pick slots yourself with hash mod 16384{user:1001}.profile and {user:1001}.orders

Question 14 of 17

Match each Redis Cluster reply to what the client should do.

-MOVED 3999 127.0.0.1:6381-ASK 3999 127.0.0.1:6381A normal reply

Question 15 of 17

Three Sentinels watch a Redis primary, with the quorum set to 2. What does the quorum decide?

How many replicas must acknowledge each writeWhether to fail over, by itselfOnly whether the primary is objectively down; failing over also needs a majority voteWhich node owns each hash slot

Question 16 of 17

A partition cuts the Redis primary off together with some clients, which keep writing to it. Sentinel promotes a replica on the other side. When the partition heals, those writes are gone. What bounds this loss?

WAIT makes Redis a CP systemA bigger replication backlogA larger Sentinel quorummin-replicas-to-write, so an isolated primary refuses writes

Question 17 of 17

Put the steps of a Sentinel failover in order.

It promotes that replica, and clients get +switch-master
A Sentinel marks the primary subjectively down
It picks a replica by priority, then replication offset, then run ID
Enough Sentinels agree: objectively down
A Sentinel wins a majority vote to lead the failover