Google SWE teaches systems design | EP26: Redis and Memcached Explained (While Drunk?)

Summary

In this candid video, the speaker, despite being inebriated, dives into the technical differences between Redis and Memcached, two popular caching solutions for large-scale distributed systems. Redis, an enhanced version of Memcached, offers additional features like data persistence, transactions, and support for various data types. The video also touches on Memcached's lack of built-in replication and Redis's high availability through its cluster feature. The speaker emphasizes the importance of choosing the right tool based on specific system requirements, highlighting the trade-offs between flexibility and built-in functionality.

Takeaways

• 🕒 The video was recorded at 1:39 a.m. on a Saturday by a speaker who was intoxicated, taking advantage of having the house to themselves.
• 📹 The speaker initially recorded a video while drunk but decided to re-record it the next morning for clarity.
• 🔑 Both Memcached and Redis are used for caching in large-scale distributed systems, primarily because they store data in RAM for faster access.
• 🔄 Memcached allows building a distributed hash map but nodes are unaware of each other, requiring consistent hashing to route requests correctly.
• 🚫 Memcached lacks built-in failure handling, replication, or availability measures, often requiring custom solutions like Facebook's 'gutter rat' instance strategy.
• 🌟 Redis offers more features than Memcached, including support for various data types, transactions, range queries, and disk persistence.
• 💾 Redis supports disk persistence through checkpointing or write-ahead logging, offering different trade-offs between speed and data integrity.
• 🔄 Redis Cluster provides high availability and consistency through replication and a gossip protocol for node communication and failover.
• 🔑 Redis uses a fixed number of partitions (16,384) to ensure even load distribution and avoid hotspots in a distributed setup.
• 🤔 Choosing between Redis and Memcached depends on specific needs; Redis offers more out-of-the-box features, but Memcached allows for more customized solutions.
• 💡 The importance of understanding the differences between Redis and Memcached is highlighted for making informed decisions in system design and interviews.

Q & A

What are the primary use cases for Redis and Memcached?

Redis and Memcached are primarily used for caching in large-scale distributed systems. They are beneficial because they store data in RAM, which allows for the creation of a distributed hash map with fast access times.

Why is RAM used for caching in Redis and Memcached?

RAM is used because it provides faster access times compared to databases on disk. This is crucial for high-performance caching in distributed systems.

What is the main difference between Memcached and Redis?

Memcached is more limited in features compared to Redis. While both can build a distributed hash map, Redis offers additional features such as support for various data types, transactions, range queries, and disk persistence.

What is consistent hashing and how is it used in Memcached?

Consistent hashing is a technique used to distribute keys across multiple nodes in a way that minimizes reorganization when nodes are added or removed. In Memcached, it helps in directing requests to the correct node.

What is the LRU (Least Recently Used) cache and its role in Memcached?

LRU cache is an eviction policy used when a cache instance reaches its capacity. It removes the least recently used items to make room for new data. In Memcached, LRU helps manage memory by evicting old data when necessary.

How does Redis differ from Memcached in terms of data types and structures?

Redis supports a variety of data types and structures, such as strings, lists, sets, and sorted sets, along with atomic operations. This is in contrast to Memcached, which is primarily a hash map from strings to strings.

What is the significance of transactions in Redis?

Transactions in Redis allow multiple write operations to be executed serially and as an atomic unit on a single node. This ensures data integrity and consistency.

What is disk persistence in Redis and why is it important?

Disk persistence in Redis is the ability to save the dataset to disk, which makes Redis more viable as a database. It helps prevent data loss and allows for data recovery in case of a system failure.

How does Redis Cluster provide high availability and consistency?

Redis Cluster provides high availability through built-in replication and automatic failover. It ensures consistency through a fixed number of partitions and a gossip protocol that prevents split brain scenarios.

Why might someone choose Memcached over Redis despite its fewer features?

One might choose Memcached over Redis if they require a simpler system or need to implement custom features such as strong consistency, alternate replication patterns, or a coordination service for partition management.

What is the importance of caching in large-scale distributed systems as illustrated by Facebook's use case?

Caching is crucial in large-scale distributed systems to reduce load on databases and improve performance. Facebook's example shows that caching handles 99% of their read requests, demonstrating its significant impact on system efficiency.

Outlines

00:00

🍻 Drunk Tech Talk: Redis and Memcache Basics

In this humorous and candid video, the speaker, despite being inebriated, attempts to discuss Redis and Memcache, two popular caching solutions used in large-scale distributed systems. The speaker acknowledges their state and promises a more coherent re-recording the next day. They explain that both systems store data in RAM to create a distributed hash map for fast data access. The main difference is that Memcache is a simpler system without built-in replication or availability features, while Redis offers more complex data structures, transactions, and disk persistence options.

05:01

🔑 Redis vs. Memcache: Features and Use Cases

This paragraph delves deeper into the technical differences between Redis and Memcache. It explains that Memcache operates with a client library directing requests to the appropriate node using consistent hashing and an LRU cache for eviction policies. However, it lacks failure handling. Redis, on the other hand, is described as an enhanced version of Memcache with support for various data types, transactions, range queries, and disk persistence. The paragraph also touches on Redis Cluster for high availability and consistency, using a gossip protocol and fixed partitioning. The speaker concludes by discussing scenarios where one might prefer Memcache over Redis due to its simplicity and flexibility in custom implementations.

Keywords

💡Redis

Redis is an open-source, in-memory data structure store, used as a database, cache, and message broker. It supports various data structures such as strings, hashes, lists, sets, and more. In the video, Redis is discussed as a more feature-rich alternative to Memcached, with capabilities like transactions, range queries, and disk persistence, making it suitable for more complex caching and data storage scenarios.

💡Memcached

Memcached is a high-performance, distributed memory caching system intended for use in speeding up dynamic web applications by alleviating database load. It is simpler than Redis and primarily used for caching purposes. The script mentions that Memcached is like a subset of Redis, lacking some of the advanced features but being sufficient for certain use cases where a simpler system is preferred.

💡Caching

Caching is the process of storing copies of data in a cache, close to the point of use, to reduce the need to repeatedly access the original data from a slower location. In the video, caching is the central theme, with both Redis and Memcached being discussed as solutions for caching in large-scale distributed systems to improve performance by reducing database load.

💡Distributed Hash Map

A distributed hash map is a data structure that provides a key-value store and is spread across multiple nodes in a network. In the script, the concept is used to describe how both Redis and Memcached can be used to create a fast-access data store by utilizing RAM, allowing for quick key-value retrieval and storage.

💡Consistent Hashing

Consistent hashing is a method of distributing a dataset across multiple nodes in a way that minimizes reorganization when nodes are added or removed. The script explains that consistent hashing is used in Memcached to direct requests to the appropriate node in the cache system.

💡LRU Cache

LRU stands for 'Least Recently Used,' a cache eviction policy that discards the least recently used items first. In the context of the video, LRU cache is mentioned as a feature of Memcached, which helps manage memory by evicting the least recently accessed data to make space for new data.

💡Replication

Replication in the context of databases and caching systems refers to the process of copying data to multiple locations to ensure redundancy and fault tolerance. The video script contrasts the lack of built-in replication in Memcached with Redis's support for replication, which enhances data availability and durability.

💡Failover

Failover is the process of switching to a redundant or standby system in the event of a failure. The script discusses Redis's automatic failover in its replication system, which helps maintain high availability by quickly promoting a replica to take over if the primary node fails.

💡Gossip Protocol

A gossip protocol is a method of communication between nodes in a distributed system where each node shares information with a few other nodes, which in turn share it with others. The script explains that Redis Cluster uses a gossip protocol to manage node state and facilitate failover by sharing heartbeats and partition information.

💡Partitioning

Partitioning in databases is the process of dividing a large dataset into smaller, more manageable pieces, often to distribute the load across multiple servers. The video describes Redis Cluster's use of fixed number partitioning, where data is evenly distributed across a set number of partitions to prevent hotspots and balance the load.

💡Checkpointing

Checkpointing is a process used in data storage to save the state of a system at regular intervals, allowing recovery to the last checkpoint in case of a failure. The script mentions checkpointing as one of the methods Redis uses for disk persistence, trading off between speed and data integrity.

💡Write-Ahead Logging

Write-Ahead Logging (WAL) is a technique used in database management systems to ensure that changes are written to a durable storage medium before they are written to the database itself. The video script describes Redis's use of WAL for disk persistence, which ensures data integrity at the cost of slower write operations.

Highlights

Redis and Memcache are both used for caching in large-scale distributed systems, storing data in RAM for fast access.

The speaker is recording while intoxicated, aiming to discuss Redis and Memcache despite the state.

The Balmer curve is mentioned as a way to gauge the speaker's ability to discuss coding topics after drinking.

Memcache allows building a distributed hash map with nodes unaware of each other, using consistent hashing.

Memcache uses LRU cache for eviction when necessary, lacking built-in failure handling or replication.

Facebook's approach to Memcache instance failure involves adding new instances without data replication.

Redis is a modification of Memcache with additional features, supporting various data types and atomic operations.

Redis supports transactions and range queries on a single partition, enhancing its capabilities over Memcache.

Redis offers disk persistence through checkpointing or write-ahead logging, increasing its viability as a database.

Redis Cluster provides high availability and consistency with replication and automatic failover.

Redis uses a gossip protocol for node communication and heartbeat sharing, crucial for failover processes.

Redis prevents split brain using a quorum of master nodes and epoch numbers during failover.

Redis employs fixed partitioning with 16,384 partitions, unlike other databases using consistent hashing.

Redis's built-in features can make it harder to diverge from its design pattern compared to Memcache.

The choice between Redis and Memcache depends on specific needs like strong consistency or replication patterns.

Caching is crucial for large-scale distributed systems, with Facebook handling 99% of read requests through cache.

The video concludes by emphasizing the importance of understanding the differences between Redis and Memcache.

Transcripts

hello everyone it is uh i think 1 39 a.m

on a saturday morning if you can't tell

i am actively hammered

why am i doing this right now it's

because my roommate is at his

girlfriend's house and uh i have the

free house to record so you know what

screw it let's get this done um

i'm gonna talk about redis and memcache

today and assuming my brain can support

it because i guess the balmer curve

tells me that i can have a couple drinks

and and you know talk about code then uh

hopefully this video will be coherent

that being said i'm also about eight or

nine drinks in so i'm a little bit past

the balmer curve but

you know what i can still talk about

computer science so let's get it done

hey yeah so this is me from the next

morning uh

i actually did record that video and uh

i looked back on it and i was like damn

i am slurring a lot of words right now

so

um we're gonna do a re-recording and uh

hopefully i'll explain a little bit more

concisely this time so let's do memcache

and redis

so basically in terms of what memcache

and redis are they're both solutions

that are typically used for caching in

large-scale distributed systems the

reason for this is that they generally

speaking store their data in ram random

access memory and that allows you to

basically build something called a

distributed hash map which can access

keys and set keys at of one time which

is great because obviously databases on

disk can't do that

as we can see though there are pretty

significant differences between the two

services themselves

mainly in the sense that

memcache is basically like a subset of

redis but you know sometimes it's good

to have a more limited feature set

because you can build out more

so

what is memcache like i said memcache

allows you to build a distributed hash

map amongst a bunch of nodes however the

nodes basically don't know about one

another so generally speaking you're

actually kind of just using this client

library to go ahead and wire all those

requests to the proper node how do we

wire each request to the right node well

generally speaking we're using

consistent hashing so basically you give

all of your you know application servers

that are going to be using memcache

instances a list of all the nodes that

are running memcached and then that will

allow them to create a consistent

hashing ring

additionally you have an lru cache so if

each memcache instance gets too big or

basically there's too much data in there

and you need to evict something in order

to make room for a new element you're

using the lru algorithm and then you

know a couple of other features that are

built in are

you know basically just like compare and

set um generally speaking there's not

really any failure handling for memcache

they don't have any built-in like

replication or availability measures so

um what actually a company like facebook

did and i'll link this lecture

from like mit basically in the video

description is that they use something

called like a gutter rattus instance

where every time um a red is not a redis

instance um a memcache instance failed

they basically go ahead and throw a new

memcache instance in there to just take

its place and then eventually it'll get

repopulated over time but there's no

like you know copying over of the data

from one instance to another which is

interesting

okay so what is redis well at least on a

single node

redis is a modification of memcache that

basically has the following features so

obviously it's an lru distributed hash

map but it's got some other stuff too so

instead of just being a hash map from

strings to strings you can actually have

other data types as values in the hash

maps so those can be sets

you know strings but with atomic

operations like appending to the string

maps and lists and you can also even

have transactions so if you want to make

multiple writes to a single node and

make sure that those are

executed both serially and also as an

atomic unit you can do that you can also

even make range queries on a single

partition and there's also a way to kind

of

go ahead and hash keys using only part

of the key

doing something called the hashtag which

allows you to kind of have some control

over where each key is going to be sent

to in a partitioning scheme

then finally there's also a concept of

actual disk persistence which makes

redis a little bit more viable as an

actual database for your application as

well and you can do disk persistence via

checkpointing which is probably faster

but obviously comes with the cost of

losing some rights if you don't

checkpoint everything or you can use a

write ahead log

where basically every single write is

going to the disk before it's written in

memory and this obviously comes at the

cost of slower writes

in terms of redis cluster redis cluster

is basically what redis calls

you know running a bunch of radis nodes

in a distributed manner so basically the

point here is that this provides both

high availability and consistency so how

do they provide high ability well unlike

memcache they actually support

replication out of the box so that uses

single leader replication with an

automatic failover

the thing with single leader replication

here is that some rights can be lost so

if the leader has some rights and it's

replicating them asynchronously just

some of its replicas and then the leader

goes ahead and fails before all of those

rights get properly sent out to all of

the followers then those rights are

probably just going to be lost um so how

do we actually do a failover well

there's a gossip protocol between all

the nodes where you're basically sharing

heartbeats that convey the you know kind

of the state of the node as well as the

partitions that it's holding and then uh

in order now to basically you know put a

new replica to the master what needs to

happen is that a quorum of masternodes

amongst all partitions need to basically

go ahead and agree that this new

follower is going to become the leader

and they use an epoch number to do that

and so this basically allows us to

prevent split brain because obviously a

quorum of nodes can't make a

conflicting decision for a single epoch

so that's pretty smart there how they

make sure to prevent split brain and

then finally

in order to kind of do partitioning

unlike a bunch of other database

solutions which most of which we've kind

of just seen using

consistent hashing so far this actually

uses the fixed number of partition

solution that i discussed from ddia

where there are exactly 16 384 fixed

partitions with fixed ranges and then

your job is basically to make sure that

you know you're putting the right number

of partitions on each node such that

there aren't too many hotspots and

obviously you know due to the fact that

there's replication if there are certain

hotkeys hopefully the load won't be too

great on them because you'll actually be

able to serve a lot of requests from

those replicas as opposed to just having

to serve them all from one instance

okay so in terms of a comparison between

memcached and redis

as you can see um redis is basically

just memcached with a bunch of

additional features built in out of the

box so why would you ever not want to

use it well

by virtue of having all these features

built in it makes it harder to kind of

diverge from that design pattern so

let's say you needed strongly consistent

data maybe you'd be better off just

using memcache and then kind of building

out your own system using something with

like a coordination service in order to

ensure strong consistency

maybe you want alternate replication

patterns like a leaderless replication

schema which you know kind of resembles

a dynamo database then perhaps you'd be

better off using memcache and then you

know kind of implementing that yourself

or maybe even you want to use just a

coordination service in general for all

of that partition management as opposed

to just using a gossip protocol

because gossip protocol even though it

does generally work you know it's just a

little bit harder to reason about

sometimes

you know you can do that too

so in conclusion both redis and memcache

are super useful systems for

implementing caching and in an interview

i imagine it probably won't really come

up if you just said yeah i'm going to

use redis or memcache in order to

implement my cache but i think it's

important to know the subtle differences

between the two of them because that's

kind of what this channel is all about

is being able to you know hear all of

the names of these different technology

services and basically go and say okay

well actually now i understand why one

is different from the other

obviously caching is hugely beneficial

for any large scale distributed system

and as a result of that you should

basically be using it whenever you can

until it pretty much prices you out so

you know as long as you can afford it

you should be using caching

that video i mentioned about caching at

facebook basically says that pretty much

99 of their read requests are handled by

their cash in order to take load off

their databases and allow them to keep

operating so okay i hope this video was

useful sorry i couldn't post the drunk

one but it was probably double as long

and very incoherent but uh have a good

one guys