Redis
Redis
License
The Little Redis Book is licensed under the Attribution-NonCommercial 3.0 Unported license. You should not have paid for this book. You are free to copy, distribute, modify or display the book. However, I ask that you always attribute the book to me, Karl Seguin, and do not use it for commercial purposes. You can see the full text of the license at: http://creativecommons.org/licenses/by-nc/3.0/legalcode
With Thanks To
A special thanks to Perry Neal for lending me his eyes, mind and passion. You provided me with invaluable help. Thank you.
Latest Version
The latest source of this book is available at: http://github.com/karlseguin/the-little-redis-book
Introduction
Over the last couple years, the techniques and tools used for persisting and querying data have grown at an incredible pace. While its safe to say that relational databases arent going anywhere, we can also say that the ecosystem around data is never going to be the same. Of all the new tools and solutions, for me, Redis has been the most exciting. Why? First because its unbelievably easy to learn. Hours is the right unit to use when talking about length of time it takes to get comfortable with Redis. Secondly, it solves a specific set of problems while at the same time being quite generic. What exactly does that mean? Redis doesnt try to be all things to all data. As you get to know Redis, itll become increasingly evident what does and what does not belong in it. And when it does, as a developer, its a great experience. While you can build a complete system using Redis only, I think most people will find that it supplements their more generic data solution - whether that be a traditional relational database, a document-oriented system, or something else. Its the kind of solution you use to implement specific features. In that way, its similar to an indexing engine. You wouldnt build your entire application on Lucene. But when you need good search, its a much better experience for both you and your users. Of course, the similarities between Redis and indexing engines end there. The goal of this book is to build the foundation youll need to master Redis. Well focus on learning Redis five data structures and look at various data modeling approaches. Well also touch on some key administrative details and debugging techniques.
Getting Started
We all learn differently: some like to get their hands dirty, some like to watch videos, and some like to read. Nothing will help you understand Redis more than actually experiencing it. Redis is easy to install and comes with a simple shell thatll give us everything we need. Lets take a couple minutes and get it up and running on our machine.
On Windows
Redis itself doesnt officially support Windows, but there are options available. You wouldnt run these in production, but Ive never experienced any limitations while doing development. A port by Microsoft Open Technologies, Inc. can be found at https://github.com/MSOpenTech/redis. As of this writing the solution is not ready for use in production systems. Another solution, which has been available for some time, can be found at https://github.com/dmajkic/redis/downloads. You can download the most up to date version (which should be at the top of the list). Extract the zip file and, based on your architecture, open either the 64bit or 32bit folder.
(Alternatively, Redis is available via various package managers. For example, MacOSX users with Homebrew installed can simply type brew install redis.) If you built it from source, the binary outputs have been placed in the src directory. Navigate to the src directory by executing cd src.
Next start the Redis console by either double clicking redis-cli (Windows) or running ./redis-cli (*nix/MacOSX). This will connect to the locally-running server on the default port (6379). You can test that everything is working by entering info into the command line interface. Youll hopefully see a bunch of key-value pairs which provide a great deal of insight into the servers status. If you are having problems with the above setup I suggest you seek help in the official Redis support group.
Redis Drivers
As youll soon learn, Redis API is best described as an explicit set of functions. It has a very simple and procedural feel to it. This means that whether you are using the command line tool, or a driver for your favorite language, things are very similar. Therefore, you shouldnt have any problems following along if you prefer to work from a programming language. If you want, head over to the client page and download the appropriate driver.
Databases
Redis has the same basic concept of a database that you are already familiar with. A database contains a set of data. The typical use-case for a database is to group all of an applications data together and to keep it separate from another applications. In Redis, databases are simply identified by a number with the default database being number 0. If you want to change to a different database you can do so via the select command. In the command line interface, type select 1. Redis should reply with an OK message and your prompt should change to something like redis 127.0.0.1:6379[1]>. If you want to switch back to the default database, just enter select 0 in the command line interface..
While Redis is more than just a key-value store, at its core, every one of Redis five data structures has at least a key and a value. Its imperative that we understand keys and values before moving on to other available pieces of information. Keys are how you identify pieces of data. Well be dealing with keys a lot, but for now, its good enough to know that a key might look like users:leto. One could reasonably expect such a key to contain information about a user named
leto. The colon doesnt have any special meaning, as far as Redis is concerned, but using a separator is a common
approach people use to organize their keys. Values represent the actual data associated with the key. They can be anything. Sometimes youll store strings, sometimes integers, sometimes youll store serialized objects (in JSON, XML or some other format). For the most part, Redis treats values as a byte array and doesnt care what they are. Note that different drivers handle serialization differently (some leave it up to you) so in this book well only talk about string, integer and JSON. Lets get our hands a little dirty. Enter the following command:
set users:leto {name: leto, planet: dune, likes: [spice]}
This is the basic anatomy of a Redis command. First we have the actual command, in this case set. Next we have its parameters. The set command takes two parameters: the key we are setting and the value we are setting it to. Many, but not all, commands take a key (and when they do, its often the first parameter). Can you guess how to retrieve this value? Hopefully you said (but dont worry if you werent sure!):
get users:leto
Go ahead and play with some other combinations. Keys and values are fundamental concepts, and the get and set commands are the simplest way to play with them. Create more users, try different types of keys, try different values. 7
Querying
As we move forward, two things will become clear. As far as Redis is concerned, keys are everything and values are nothing. Or, put another way, Redis doesnt allow you to query an objects values. Given the above, we cant find the user(s) which live on planet dune. For many, this is will cause some concern. Weve lived in a world where data querying is so flexible and powerful that Redis approach seems primitive and unpragmatic. Dont let it unsettle you too much. Remember, Redis isnt a one-size-fits-all solution. Therell be things that just dont belong in there (because of the querying limitations). Also, consider that in some cases youll find new ways to model your data. Well look at more concrete examples as we move on, but its important that we understand this basic reality of Redis. It helps us understand why values can be anything - Redis never needs to read or understand them. Also, it helps us get our minds thinking about modeling in this new world.
We mentioned before that Redis is an in-memory persistent store. With respect to persistence, by default, Redis snapshots the database to disk based on how many keys have changed. You configure it so that if X number of keys change, then save the database every Y seconds. By default, Redis will save the database every 60 seconds if 1000 or more keys have changed all the way to 15 minutes if 9 or less keys has changed. Alternatively (or in addition to snapshotting), Redis can run in append mode. Any time a key changes, an append-only file is updated on disk. In some cases its acceptable to lose 60 seconds worth of data, in exchange for performance, should there be some hardware or software failure. In some cases such a loss is not acceptable. Redis gives you the option. In chapter 6 well see a third option, which is offloading persistence to a slave. With respect to memory, Redis keeps all your data in memory. The obvious implication of this is the cost of running Redis: RAM is still the most expensive part of server hardware. I do feel that some developers have lost touch with how little space data can take. The Complete Works of William Shakespeare takes roughly 5.5MB of storage. As for scaling, other solutions tend to be IO- or CPU-bound. Which limitation (RAM or IO) will require you to scale out to more machines really depends on the type of data and how you are storing and querying it. Unless youre storing large multimedia files in Redis, the in-memory aspect is probably a non-issue. For apps where it is an issue youll likely be trading being IO-bound for being memory bound. Redis did add support for virtual memory. However, this feature has been seen as a failure (by Redis own developers) and its use has been deprecated. (On a side note, that 5.5MB file of Shakespeares complete works can be compressed down to roughly 2MB. Redis doesnt do auto-compression but, since it treats values as bytes, theres no reason you cant trade processing time for RAM by compressing/decompressing the data yourself.) 8
Putting It Together
Weve touched on a number of high level topics. The last thing I want to do before diving into Redis is bring some of those topics together. Specifically, query limitations, data structures and Redis way to store data in memory. When you add those three things together you end up with something wonderful: speed. Some people think Of course Redis is fast, everythings in memory. But thats only part of it. The real reason Redis shines versus other solutions is its specialized data structures. How fast? It depends on a lot of things - which commands you are using, the type of data, and so on. But Redis performance tends to be measured in tens of thousands, or hundreds of thousands of operations per second. You can run redis-benchmark (which is in the same folder as the redis-server and redis-cli) to test it out yourself. I once changed code which used a traditional model to using Redis. A load test I wrote took over 5 minutes to finish using the relational model. It took about 150ms to complete in Redis. You wont always get that sort of massive gain, but it hopefully gives you an idea of what we are talking about. Its important to understand this aspect of Redis because it impacts how you interact with it. Developers with an SQL background often work at minimizing the number of round trips they make to the database. Thats good advice for any system, including Redis. However, given that we are dealing with simpler data structures, well sometimes need to hit the Redis server multiple times to achieve our goal. Such data access patterns can feel unnatural at first, but in reality it tends to be an insignificant cost compared to the raw performance we gain.
In This Chapter
Although we barely got to play with Redis, we did cover a wide range of topics. Dont worry if something isnt crystal clear - like querying. In the next chapter well go hands-on and any questions you have will hopefully answer themselves. The important takeaways from this chapter are: Keys are strings which identify pieces of data (values) Values are arbitrary byte arrays that Redis doesnt care about Redis exposes (and is implemented as) five specialized data structures Combined, the above make Redis fast and easy to use, but not suitable for every scenario
Theres nothing more important than having fun and trying things out. You can always erase all the values in your database by entering flushdb, so dont be shy and try doing crazy things!
Strings
Strings are the most basic data structures available in Redis. When you think of a key-value pair, you are thinking of strings. Dont get mixed up by the name, as always, your value can be anything. I prefer to call them scalars, but maybe thats just me. We already saw a common use-case for strings, storing instances of objects by key. This is something that youll make heavy use of:
set users:leto {name: leto, planet: dune, likes: [spice]}
Additionally, Redis lets you do some common operations. For example strlen <key> can be used to get the length of a keys value; getrange <key> <start> <end> returns the specified range of a value; append <key> <value> appends the value to the existing value (or creates it if it doesnt exist already). Go ahead and try those out. This is what I get:
> strlen users:leto (integer) 42
Now, you might be thinking, thats great, but it doesnt make sense. You cant meaningfully pull a range out of JSON or append a value. You are right, the lesson here is that some of the commands, especially with the string data structure, only make sense given specific type of data. Earlier we learnt that Redis doesnt care about your values. Most of the time thats true. However, a few string commands are specific to some types or structure of values. As a vague example, I could see the above append and 10
getrange commands being useful in some custom space-efficient serialization. As a more concrete example I give you
the incr, incrby, decr and decrby commands. These increment or decrement the value of a string:
> incr stats:page:about (integer) 1 > incr stats:page:about (integer) 2
As you can imagine, Redis strings are great for analytics. Try incrementing users:leto (a non-integer value) and see what happens (you should get an error). A more advanced example is the setbit and getbit commands. Theres a wonderful post on how Spool uses these two commands to efficiently answer the question how many unique visitors did we have today. For 128 million users a laptop generates the answer in less than 50ms and takes only 16MB of memory. It isnt important that you understand how bitmaps work, or how Spool uses them, but rather to understand that Redis strings are more powerful than they initially seem. Still, the most common cases are the ones we gave above: storing objects (complex or not) and counters. Also, since getting a value by key is so fast, strings are often used to cache data.
Hashes
Hashes are a good example of why calling Redis a key-value store isnt quite accurate. You see, in a lot of ways, hashes are like strings. The important difference is that they provide an extra level of indirection: a field. Therefore, the hash equivalents of set and get are:
hset users:goku powerlevel 9000 hget users:goku powerlevel
We can also set multiple fields at once, get multiple fields at once, get all fields and values, list all the fields or delete a specific field:
hmset users:goku race saiyan age 737 hmget users:goku race powerlevel hgetall users:goku hkeys users:goku hdel users:goku age
As you can see, hashes give us a bit more control over plain strings. Rather than storing a user as a single serialized value, we could use a hash to get a more accurate representation. The benefit would be the ability to pull and update/delete specific pieces of data, without having to get or write the entire value. 11
Looking at hashes from the perspective of a well-defined object, such as a user, is key to understanding how they work. And its true that, for performance reasons, more granular control might be useful. However, in the next chapter well look at how hashes can be used to organize your data and make querying more practical. In my opinion, this is where hashes really shine.
Lists
Lists let you store and manipulate an array of values for a given key. You can add values to the list, get the first or last value and manipulate values at a given index. Lists maintain their order and have efficient index-based operations. We could have a newusers list which tracks the newest registered users to our site:
lpush newusers goku ltrim newusers 0 50
First we push a new user at the front of the list, then we trim it so that it only contains the last 50 users. This is a common pattern. ltrim is an O(N) operation, where N is the number of values we are removing. In this case, where we always trim after a single insert, itll actually have a constant performance of O(1) (because N will always be equal to 1). This is also the first time that we are seeing a value in one key referencing a value in another. If we wanted to get the details of the last 10 users, wed do the following combination:
keys = redis.lrange(newusers, 0, 10) redis.mget(*keys.map {|u| users:#{u}})
The above is a bit of Ruby which shows the type of multiple roundtrips we talked about before. Of course, lists arent only good for storing references to other keys. The values can be anything. You could use lists to store logs or track the path a user is taking through a site. If you were building a game, you might use it to track a queued user actions.
Sets
Set are used to store unique values and provide a number of set-based operations, like unions. Sets arent ordered but they provide efficient value-based operations. A friends list is the classic example of using a set:
sadd friends:leto ghanima paul chani jessica sadd friends:duncan paul jessica alia
Regardless of how many friends a user has, we can efficiently tell (O(1)) whether userX is a friend of userY or not:
sismember friends:leto jessica sismember friends:leto vladimir
Furthermore we can see what two or more people share the same friends:
sinter friends:leto friends:duncan
12
Sets are great for tagging or tracking any other properties of a value for which duplicates dont make any sense (or where we want to apply set operations such as intersections and unions).
Sorted Sets
The last and most powerful data structure are sorted sets. If hashes are like strings but with fields, then sorted sets are like sets but with a score. The score provides sorting and ranking capabilities. If we wanted a ranked list of friends, we might do:
zadd friends:duncan 70 ghanima 95 paul 95 chani 75 jessica 1 vladimir
Want to find out how many friends duncan has with a score of 90 or over?
zcount friends:duncan 90 100
We use zrevrank instead of zrank since Redis default sort is from low to high (but in this case we are ranking from high to low). The most obvious use-case for sorted sets is a leaderboard system. In reality though, anything you want sorted by an some integer, and be able to efficiently manipulate based on that score, might be a good fit for a sorted set.
In This Chapter
Thats a high level overview of Redis five data structures. One of the neat things about Redis is that you can often do more than you first realize. There are probably ways to use string and sorted sets that no one has thought of yet. As long as you understand the normal use-case though, youll find Redis ideal for all types of problems. Also, just because Redis exposes five data structures and various methods, dont think you need to use all of them. It isnt uncommon to build a feature while only using a handful of commands.
13
In the previous chapter we talked about the five data structures and gave some examples of what problems they might solve. Now its time to look at a few more advanced, yet common, topics and design patterns.
Big O Notation
Throughout this book weve made references to the Big O notation in the form of O(n) or O(1). Big O notation is used to explain how something behaves given a certain number of elements. In Redis, its used to tell us how fast a command is based on the number of items we are dealing with. Redis documentation tells us the Big O notation for each of its commands. It also tells us what the factors are that influence the performance. Lets look at some examples. The fastest anything can be is O(1) which is a constant. Whether we are dealing with 5 items or 5 million, youll get the same performance. The sismember command, which tells us if a value belongs to a set, is O(1). sismember is a powerful command, and its performance characteristics are a big reason for that. A number of Redis commands are O(1). Logarithmic, or O(log(N)), is the next fastest possibility because it needs to scan through smaller and smaller partitions. Using this type of divide and conquer approach, a very large number of items quickly gets broken down in a few iterations. zadd is a O(log(N)) command, where N is the number of elements already in the set. Next we have linear commands, or O(N). Looking for a non-indexed column in a table is an O(N) operation. So is using the ltrim command. However, in the case of ltrim, N isnt the number of elements in the list, but rather the elements being removed. Using ltrim to remove 1 item from a list of millions will be faster than using ltrim to remove 10 items from a list of thousands. (Though theyll probably both be so fast that you wouldnt be able to time it.)
zremrangebyscore which removes elements from a sorted set with a score between a minimum and a maximum value
has a complexity of O(log(N)+M). This makes it a mix. By reading the documentation we see that N is the number of total elements in the set and M is the number of elements to be removed. In other words, the number of elements thatll get removed is probably going to be more significant, in terms of performance, than the total number of elements in the set. The sort command, which well discuss in greater detail in the next chapter has a complexity of O(N+M*log(M)). From its performance characteristic, you can probably tell that this is one of Redis most complex commands. There are a number of other complexities, the two remaining common ones are O(N2) and O(CN). The larger N is, the worse these perform relative to a smaller N. None of Redis commands have this type of complexity. Its worth pointing out that the Big O notation deals with the worst case. When we say that something takes O(N), we might actually find it right away or it might be the last possible element. 14
This is bad because its a nightmare to manage and it takes twice the amount of memory. It would be nice if Redis let you link one key to another, but it doesnt (and it probably never will). A major driver in Redis development is to keep the code and API clean and simple. The internal implementation of linking keys (theres a lot we can do with keys that we havent talked about yet) isnt worth it when you consider that Redis already provides a solution: hashes. Using a hash, we can remove the need for duplication:
set users:9001 {id: 9001, email: leto@dune.gov, ...} hset users:lookup:email leto@dune.gov 9001
What we are doing is using the field as a pseudo secondary index and referencing the single user object. To get a user by id, we issue a normal get:
get users:9001
This is something that youll likely end up doing often. To me, this is where hashes really shine, but it isnt an obvious use-case until you see it.
Each member of this set is a reference to a Redis string value containing details on the actual user. What if chani changes her name, or deletes her account? Maybe it would make sense to also track the inverse relationships: 15
Maintenance cost aside, if you are anything like me, you might cringe at the processing and memory cost of having these extra indexed values. In the next section well talk about ways to reduce the performance cost of having to do extra round trips (we briefly talked about it in the first chapter). If you actually think about it though, relational databases have the same overhead. Indexes take memory, must be scanned or ideally seeked and then the corresponding records must be looked up. The overhead is neatly abstracted away (and they do a lot of optimizations in terms of the processing to make it very efficient). Again, having to manually deal with references in Redis is unfortunate. But any initial concerns you have about the performance or memory implications of this should be tested. I think youll find it a non-issue.
Redis also supports pipelining. Normally when a client sends a request to Redis it waits for the reply before sending the next request. With pipelining you can send a number of requests without waiting for their responses. This reduces the networking overhead and can result in significant performance gains. Its worth noting that Redis will use memory to queue up the commands, so its a good idea to batch them. How large a batch you use will depend on what commands you are using, and more specifically, how large the parameters are. But, if you are issuing commands against ~50 character keys, you can probably batch them in thousands or tens of thousands. Exactly how you execute commands within a pipeline will vary from driver to driver. In Ruby you pass a block to the
pipelined method: redis.pipelined do 9001.times do redis.incr(powerlevel) end end
As you can probably guess, pipelining can really speed up a batch import! 16
Transactions
Every Redis command is atomic, including the ones that do multiple things. Additionally, Redis has support for transactions when using multiple commands. You might not know it, but Redis is actually single-threaded, which is how every command is guaranteed to be atomic. While one command is executing, no other command will run. (Well briefly talk about scaling in a later chapter.) This is particularly useful when you consider that some commands do multiple things. For example:
incr is essentially a get followed by a set getset sets a new value and returns the original setnx first checks if the key exists, and only sets the value if it does not
Although these commands are useful, youll inevitably need to run multiple commands as an atomic group. You do so by first issuing the multi command, followed by all the commands you want to execute as part of the transaction, and finally executing exec to actually execute the commands or discard to throw away, and not execute the commands. What guarantee does Redis make about transactions? The commands will be executed in order The commands will be executed as a single atomic operation (without another clients command being executed halfway through) That either all or none of the commands in the transaction will be executed You can, and should, test this in the command line interface. Also note that theres no reason why you cant combine pipelining and transactions.
multi hincrby groups:1percent balance -9000000000 hincrby groups:99percent balance 9000000000 exec
Finally, Redis lets you specify a key (or keys) to watch and conditionally apply a transaction if the key(s) changed. This is used when you need to get values and execute code based on those values, all in a transaction. With the code above, we wouldnt be able to implement our own incr command since they are all executed together once exec is called. From code, we cant do:
redis.multi() current = redis.get(powerlevel) redis.set(powerlevel, current + 1) redis.exec()
That isnt how Redis transactions work. But, if we add a watch to powerlevel, we can do:
redis.watch(powerlevel) current = redis.get(powerlevel) redis.multi()
17
If another client changes the value of powerlevel after weve called watch on it, our transaction will fail. If no client changes the value, the set will work. We can execute this code in a loop until it works.
Keys Anti-Pattern
In the next chapter well talk about commands that arent specifically related to data structures. Some of these are administrative or debugging tools. But theres one Id like to talk about in particular: the keys command. This command takes a pattern and finds all the matching keys. This command seems like its well suited for a number of tasks, but it should never be used in production code. Why? Because it does a linear scan through all the keys looking for matches. Or, put simply, its slow. How do people try and use it? Say you are building a hosted bug tracking service. Each account will have an id and you might decide to store each bug into a string value with a key that looks like bug:account_id:bug_id. If you ever need to find all of an accounts bugs (to display them, or maybe delete them if they delete their account), you might be tempted (as I was!) to use the keys command:
keys bug:1233:*
The better solution is to use a hash. Much like we can use hashes to provide a way to expose secondary indexes, so too can we use them to organize our data:
hset bugs:1233 1 {id:1, account: 1233, subject: ...} hset bugs:1233 2 {id:2, account: 1233, subject: ...}
To get all the bug ids for an account we simply call hkeys bugs:1233. To delete a specific bug we can do hdel bugs
:1233 2 and to delete an account we can delete the key via del bugs:1233.
In This Chapter
This chapter, combined with the previous one, has hopefully given you some insight on how to use Redis to power real features. There are a number of other patterns you can use to build all types of things, but the real key is to understand the fundamental data structures and to get a sense for how they can be used to achieve things beyond your initial perspective.
18
Expiration
Redis allows you to mark a key for expiration. You can give it an absolute time in the form of a Unix timestamp (seconds since January 1, 1970) or a time to live in seconds. This is a key-based command, so it doesnt matter what type of data structure the key represents.
expire pages:about 30 expireat pages:about 1356933600
The first command will delete the key (and associated value) after 30 seconds. The second will do the same at 12:00 a.m. December 31st, 2012. This makes Redis an ideal caching engine. You can find out how long an item has to live until via the ttl command and you can remove the expiration on a key via the persist command:
ttl pages:about persist pages:about
Finally, theres a special string command, setex which lets you set a string and specify a time to live in a single atomic command (this is more for convenience than anything else):
setex pages:about 30 <h1>about us</h1>....
The reply is the information of your subscription. Now, in the other window, publish a message to the warnings channel:
publish warnings its over 9000!
If you go back to your first window you should have received the message to the warnings channel. 19
You can subscribe to multiple channels (subscribe channel1 channel2 ...), subscribe to a pattern of channels (psubscribe warnings:*) and use the unsubscribe and punsubscribe commands to stop listening to one or more channels, or a channel pattern. Finally, notice that the publish command returned the value 1. This indicates the number of clients that received the message.
Next, issue a few commands. Finally you can retrieve all of the logs, or the most recent logs via:
slowlog get slowlog get 10
You can also get the number of items in the slow log by entering slowlog len For each command you entered you should see four parameters: An auto-incrementing id A Unix timestamp for when the command happened The time, in microseconds, it took to run the command The command and its parameters The slow log is maintained in memory, so running it in production, even with a low threshold, shouldnt be a problem. By default it will track the last 1024 logs.
Sort
One of Redis most powerful commands is sort. It lets you sort the values within a list, set or sorted set (sorted sets are ordered by score, not the members within the set). In its simplest form, it allows us to do: 20
Which will return the values sorted from lowest to highest. Heres a more advanced example:
sadd friends:ghanima leto paul chani jessica alia duncan sort friends:ghanima limit 0 3 desc alpha
The above command shows us how to page the sorted records (via limit), how to return the results in descending order (via desc) and how to sort lexicographically instead of numerically (via alpha). The real power of sort is its ability to sort based on a referenced object. Earlier we showed how lists, sets and sorted sets are often used to reference other Redis objects. The sort command can dereference those relations and sort by the underlying value. For example, say we have a bug tracker which lets users watch issues. We might use a set to track the issues being watched:
sadd watch:leto 12339 1382 338 9338
It might make perfect sense to sort these by id (which the default sort will do), but wed also like to have these sorted by severity. To do so, we tell Redis what pattern to sort by. First, lets add some more data so we can actually see a meaningful result:
set severity:12339 3 set severity:1382 2 set severity:338 5 set severity:9338 4
Redis will substitute the * in our pattern (identified via by) with the values in our list/set/sorted set. This will create the key name that Redis will query for the actual values to sort by. Although you can have millions of keys within Redis, I think the above can get a little messy. Thankfully sort can also work on hashes and their fields. Instead of having a bunch of top-level keys you can leverage hashes:
hset bug:12339 severity 3 hset bug:12339 priority 1 hset bug:12339 details {id: 12339, ....}
hset bug:1382 severity 2 hset bug:1382 priority 2 hset bug:1382 details {id: 1382, ....}
hset bug:338 severity 5 hset bug:338 priority 3 hset bug:338 details {id: 338, ....}
21
hset bug:9338 severity 4 hset bug:9338 priority 2 hset bug:9338 details {id: 9338, ....}
Not only is everything better organized, and we can sort by severity or priority, but we can also tell sort what field to retrieve:
sort watch:leto by bug:*->priority get bug:*->details
The same value substitution occurs, but Redis also recognizes the -> sequence and uses it to look into the specified field of our hash. Weve also included the get parameter, which also does the substitution and field lookup, to retrieve bug details. Over large sets, sort can be slow. The good news is that the output of a sort can be stored:
sort watch:leto by bug:*->priority get bug:*->details store watch_by_priority:leto
Combining the store capabilities of sort with the expiration commands weve already seen makes for a nice combo.
In This Chapter
This chapter focused on non-data structure-specific commands. Like everything else, their use is situational. It isnt uncommon to build an app or feature that wont make use of expiration, publication/subscription and/or sorting. But its good to know that they are there. Also, we only touched on some of the commands. There are more, and once youve digested the material in this book its worth going through the full list.
22
Why?
Before looking at how to use Lua scripting, you might be wondering why youd want to use it. Many developers dislike traditional stored procedures, is this any different? The short answer is no. Improperly used, Redis Lua scripting can result in harder to test code, business logic tightly coupled with data access or even duplicated logic. Properly used however, its a feature that can simplify code and improve performance. Both of these benefits are largely achieved by grouping multiple commands, along with some simple logic, into a custom-build cohesive function. Code is made simpler because each invocation of a Lua script is run without interruption and thus provides a clean way to create your own atomic commands (essentially eliminating the need to use the cumbersome watch command). It can improve performance by removing the need to return intermediary results - the final output can be calculated within the script. The examples in the following sections will better illustrate these points.
Eval
The eval command takes a Lua script (as a string), the keys well be operating against, and an optional set of arbitrary arguments. Lets look at a simple example (executed from Ruby, since running multi-line Redis commands from its command-line tool isnt fun):
script = <<-eos local friend_names = redis.call(smembers, KEYS[1]) local friends = {} for i = 1, #friend_names do local friend_key = user: .. friend_names[i] local gender = redis.call(hget, friend_key, gender) if gender == ARGV[1] then table.insert(friends, redis.call(hget, friend_key, details)) end end return friends eos Redis.new.eval(script, [friends:leto], [m])
23
The above code gets the details for all of Letos male friends. Notice that to call Redis commands within our script we use the redis.call(command, ARG1, ARG2, ...) method. If you are new to Lua, you should go over each line carefully. It might be useful to know that {} creates an empty
table (which can act as either an array or a dictionary), #TABLE gets the number of elements in the TABLE, and .. is
driver automatically creates this for us. Why is this needed? Consider how the above looks like when executed from the CLI:
eval ..... friends:leto m vs eval ..... 1 friends:leto m
In the first (incorrect) case, how does Redis know which of the parameters are keys and which are simply arbitrary arguments? In the second case, there is no ambiguity. This brings up a second question: why must keys be explicitly listed? Every command in Redis knows, at execution time, which keys are going to needed. This will allow future tools, like Redis Cluster, to distribute requests amongst multiple Redis servers. You might have spotted that our above example actually reads from keys dynamically (without having them passed to eval). An hget is issued on all of Letos male friends. Thats because the need to list keys ahead of time is more of a suggestion than a hard rule. The above code will run fine in a single-instance setup, or even with replication, but wont in the yet-released Redis Cluster.
Script Management
Even though scripts executed via eval are cached by Redis, sending the body every time you want to execute something isnt ideal. Instead, you can register the script with Redis and execute its key. To do this you use the script load command, which returns the SHA1 digest of the script:
redis = Redis.new script_key = redis.script(:load, THE_SCRIPT)
Once weve loaded the script, we can use evalsha to execute it:
redis.evalsha(script_key, [friends:leto], [m])
script kill, script flush and script exists are the other commands that you can use to manage Lua scripts.
They are used to kill a running script, removing all scripts from the internal cache and seeing if a script already exists within the cache.
Libraries
Redis Lua implementation ships with a handful of useful libraries. While table.lib, string.lib and math.lib are quite useful, for me, cjson.lib is worth singling out. First, if you find yourself having to pass multiple arguments to a 24
Which you could then deserialize within the Lua script as:
local arguments = cjson.decode(ARGV[1])
Of course, the JSON library can also be used to parse values stored in Redis itself. Our above example could potentially be rewritten as such:
local friend_names = redis.call(smembers, KEYS[1]) local friends = {} for i = 1, #friend_names do local friend_raw = redis.call(get, user: .. friend_names[i]) local friend_parsed = cjson.decode(friend_raw) if friend_parsed.gender == ARGV[1] then table.insert(friends, friend_raw) end end return friends
Instead of getting the gender from specific hash field, we could get it from the stored friend data itself. (This is a much slower solution, and I personally prefer the original, but it does show whats possible).
Atomic
Since Redis is single-threaded, you dont have to worry about your Lua script being interrupted by another Redis command. One of the most obvious benefits of this is that keys with a TTL wont expire half-way through execution. If a key is present at the start of the script, itll be present at any point thereafter - unless you delete it.
Administration
The next chapter will talk about Redis administration and configuration in more detail. For now, simply know that the
lua-time-limit defines how long a Lua script is allowed to execute before being terminated. The default is generous
In This Chapter
This chapter introduced Redis Lua scripting capabilities. Like anything, this feature can be abused. However, used prudently in order to implement your own custom and focused commands, it wont only simplify your code, but will likely improve performance. Lua scripting is like almost every other Redis feature/command: you make limited, if any, use of it at first only to find yourself using it more and more every day.
25
Chapter 6 - Administration
Our last chapter is dedicated to some of the administrative aspects of running Redis. In no way is this a comprehensive guide on Redis administration. At best well answer some of the more basic questions new users to Redis are most likely to have.
Configuration
When you first launched the Redis server, it warned you that the redis.conf file could not be found. This file can be used to configure various aspects of Redis. A well-documented redis.conf file is available for each release of Redis. The sample file contains the default configuration options, so its useful to both understand what the settings do and what their defaults are. You can find it at https://github.com/antirez/redis/raw/2.4.6/redis.conf.
This is the config file for Redis 2.4.6. You should replace 2.4.6 in the above URL with your version. You can find your version by running the info command and looking at the first value.
Since the file is well-documented, we wont be going over the settings. In addition to configuring Redis via the redis.conf file, the config set command can be used to set individual values. In fact, we already used it when setting the slowlog-log-slower-than setting to 0. Theres also the config get command which displays the value of a setting. This command supports pattern matching. So if we want to display everything related to logging, we can do:
config get *log*
Authentication
Redis can be configured to require a password. This is done via the requirepass setting (set through either the redis
.conf file or the config set command). When requirepass is set to a value (which is the password to use), clients
will need to issue an auth password command. Once a client is authenticated, they can issue any command against any database. This includes the flushall command which erases every key from every database. Through the configuration, you can rename commands to achieve some security through obfuscation:
rename-command CONFIG 5ec4db169f9d4dddacbfb0c26ea7e5ef rename-command FLUSHALL 1041285018a942a4922cbf76623b741e
Or you can disable a command by setting the new name to an empty string.
Size Limitations
As you start using Redis, you might wonder how many keys can I have? You might also wonder how many fields can a hash have (especially when you use it to organize your data), or how many elements can lists and sets have? Per 26
instance, the practical limits for all of these is in the hundreds of millions.
Replication
Redis supports replication, which means that as you write to one Redis instance (the master), one or more other instances (the slaves) are kept up-to-date by the master. To configure a slave you use either the slaveof configuration setting or the slaveof command (instances running without this configuration are or can be masters). Replication helps protect your data by copying to different servers. Replication can also be used to improve performance since reads can be sent to slaves. They might respond with slightly out of date data, but for most apps thats a worthwhile tradeoff. Unfortunately, Redis replication doesnt yet provide automated failover. If the master dies, a slave needs to be manually promoted. Traditional high-availability tools that use heartbeat monitoring and scripts to automate the switch are currently a necessary headache if you want to achieve some sort of high availability with Redis.
Backups
Backing up Redis is simply a matter of copying Redis snapshot to whatever location you want (S3, FTP, ). By default Redis saves its snapshot to a file named dump.rdb. At any point in time, you can simply scp, ftp or cp (or anything else) this file. It isnt uncommon to disable both snapshotting and the append-only file (aof) on the master and let a slave take care of this. This helps reduce the load on the master and lets you set more aggressive saving parameters on the slave without hurting overall system responsiveness.
In This Chapter
Given the number of projects and sites using Redis already, there can be no doubt that Redis is production-ready, and has been for a while. However, some of the tooling, especially around security and availability is still young. Redis Cluster, which well hopefully see soon, should help address some of the current management challenges.
28
Conclusion
In a lot of ways, Redis represents a simplification in the way we deal with data. It peels away much of the complexity and abstraction available in other systems. In many cases this makes Redis the wrong choice. In others it can feel like Redis was custom-built for your data. Ultimately it comes back to something I said at the very start: Redis is easy to learn. There are many new technologies and it can be hard to figure out whats worth investing time into learning. When you consider the real benefits Redis has to offer with its simplicity, I sincerely believe that its one of the best investments, in terms of learning, that you and your team can make.
29