-
Memcachedb is basically memcached done persistent. Redis is a different evolution
+ = Why I need Redis if there is already memcachedb, Tokyo Cabinet, ...? =
Memcachedb is basically memcached done persistent. Redis is a different evolution
path in the key-value DBs, the idea is that the main advantages of key-value DBs
are retained even without a so severe loss of comfort of plain key-value DBs.
So Redis offers more features:
- Keys can store different data types, not just strings. Notably Lists and Sets. For example if you want to use Redis as a log storage system for different computers every computer can just
RPUSH data to the computer_ID key. Don't want to save more than 1000 log lines per computer? Just issue a LTRIM computer_ID 0 999 command to trim the list after every push.
@@ -34,7 +34,31 @@ So Redis offers more features:
- Keys can store different data t
- We wrote a simple Twitter Clone using just Redis as database. Download the source code from the download section and imagine to write it with a plain key-value DB without support for lists and sets... it's much harder.
- Multiple DBs. Using the SELECT command the client can select different datasets. This is useful because Redis provides a MOVE atomic primitive that moves a key form a DB to another one, if the target DB already contains such a key it returns an error: this basically means a way to perform locking in distributed processing.
- So what is Redis really about? The User interface with the programmer. Redis aims to export to the programmer the right tools to model a wide range of problems. Sets, Lists with O(1) push operation, lrange and ltrim, server-side fast intersection between sets, are primitives that allow to model complex problems with a key value database.
-I imagine key-value DBs, in the short term future, to be used like you use memory in a program, with lists, hashes, and so on. With Redis it's like this, but this special kind of memory containing your data structures is shared, atomic, persistent.
When we write code it is obvious, when we take data in memory, to use the most sensible data structure for the work, right? Incredibly when data is put inside a relational DB this is no longer true, and we create an absurd data model even if our need is to put data and get this data back in the same order we put it inside (an ORDER BY is required when the data should be already sorted. Strange, dont' you think?).
Key-value DBs bring this back at home, to create sensible data models and use the right data structures for the problem we are trying to solve.Yes you can. When Redis saves the DB it actually creates a temp file, then rename(2) that temp file name to the destination file name. So even while the server is working it is safe to save the database file just with the cp unix command. Note that you can use master-slave replication in order to have redundancy of data, but if all you need is backups, cp or scp will do the work pretty well.Worst case scenario: 1 Million keys with the key being the natural numbers from 0 to 999999 and the string "Hello World" as value use 100MB on my Intel macbook (32bit). Note that the same data stored linearly in an unique string takes something like 16MB, this is the norm because with small keys and values there is a lot of overhead. Memcached will perform similarly.
With large keys/values the ratio is much better of course.
64 bit systems will use much more memory than 32 bit systems to store the same keys, especially if the keys and values are small, this is because pointers takes 8 bytes in 64 bit systems. But of course the advantage is that you can have a lot of memory in 64 bit systems, so to run large Redis servers a 64 bit system is more or less required.The whole key-value hype started for a reason: performances. Redis takes the whole dataset in memory and writes asynchronously on disk in order to be very fast, you have the best of both worlds: hyper-speed and persistence of data, but the price to pay is exactly this, that the dataset must fit on your computers RAM.
If the data is larger then memory, and this data is stored on disk, what happens is that the bottleneck of the disk I/O speed will start to ruin the performances. Maybe not in benchmarks, but once you have real load from multiple clients with distributed key accesses the data must come from disk, and the disk is damn slow. Not only, but Redis supports higher level data structures than the plain values. To implement this things on disk is even slower.
Redis will always continue to hold the whole dataset in memory because this days scalability requires to use RAM as storage media, and RAM is getting cheaper and cheaper. Today it is common for an entry level server to have 16 GB of RAM! And in the 64-bit era there are no longer limits to the amount of RAM you can have in theory.You may try to load a dataset larger than your memory in Redis and see what happens, basically if you are using a modern Operating System, and you have a lot of data in the DB that is rarely accessed, the OS's virtual memory implementation will try to swap rarely used pages of memory on the disk, to only recall this pages when they are needed. If you have many large values rarely used this will work. If your DB is big because you have tons of little values accessed at random without a specific pattern this will not work (at low level a page is usually 4096 bytes, and you can have different keys/values stored at a single page. The OS can't swap this page on disk if there are even few keys used frequently).
Another possible solution is to use both MySQL and Redis at the same time, basically take the state on Redis, and all the things that get accessed very frequently: user auth tokens, Redis Lists with chronologically ordered IDs of the last N-comments, N-posts, and so on. Then use MySQL as a simple storage engine for larger data, that is just create a table with an auto-incrementing ID as primary key and a large BLOB field as data field. Access MySQL data only by primary key (the ID). The application will run the high traffic queries against Redis but when there is to take the big data will ask MySQL for specific resources IDs.This may happen and it's prefectly ok. Redis objects are small C structures allocated and freed a lot of times. This costs a lot of CPU so instead of being freed, released objects are taken into a free list and reused when needed. This memory is taken exactly by this free objects ready to be reused.With modern operating systems malloc() returning NULL is not common, usually the server will start swapping and Redis performances will be disastrous so you'll know it's time to use more Redis servers or get more RAM.
However it is planned to add a configuration directive to tell Redis to stop accepting queries but instead to SAVE the latest data and quit if it is using more than a given amount of memory. Also the new INFO command (work in progress in this days) will report the amount of memory Redis is using so you can write scripts that monitor your Redis servers checking for critical conditions.
Update: redis SVN is able to know how much memory it is using and report it via the INFO command.Just an example on normal hardware: It takes about 45 seconds to restore a 2 GB database on a fairly standard system, no RAID. This can give you some kind of feeling about the order of magnitude of the time needed to load data when you restart the server.Short answer: echo 1 > /proc/sys/vm/overcommit_memory :)
And now the long one:
Redis background saving schema relies on the copy-on-write semantic of fork in modern operating systems: Redis forks (creates a child process) that is an exact copy of the parent. The child process dumps the DB on disk and finally exits. In theory the child should use as much memory as the parent being a copy, but actually thanks to the copy-on-write semantic implemented by most modern operating systems the parent and child process will share the common memory pages. A page will be duplicated only when it changes in the child or in the parent. Since in theory all the pages may change while the child process is saving, Linux can't tell in advance how much memory the child will take, so if the overcommit_memory setting is set to zero fork will fail unless there is as much free RAM as required to really duplicate all the parent memory pages, with the result that if you have a Redis dataset of 3 GB and just 2 GB of free memory it will fail.
Setting overcommit_memory to 1 says Linux to relax and perform the fork in a more optimistic allocation fashion, and this is indeed what you want for Redis.Simply start multiple instances of Redis in different ports in the same box and threat them as different servers! Given that Redis is a distributed database anyway in order to scale you need to think in terms of multiple computational units. At some point a single box may not be enough anyway.
In general key-value databases are very scalable because of the property that different keys can stay on different servers independently.
In Redis there are client libraries such Redis-rb (the Ruby client) that are able to handle multiple servers automatically using consistent hashing. We are going to implement consistent hashing in all the other major client libraries. If you use a different language you can implement it yourself otherwise just hash the key before to SET / GET it from a given server. For example imagine to have N Redis servers, server-0, server-1, ..., server-N. You want to store the key "foo", what's the right server where to put "foo" in order to distribute keys evenly among different servers? Just perform the crc = CRC32("foo"), then servernum = crc % N (the rest of the division for N). This will give a number between 0 and N-1 for every key. Connect to this server and store the key. The same for gets.
This is a basic way of performing key partitioning, consistent hashing is much better and this is why after Redis 1.0 will be released we'll try to implement this in every widely used client library starting from Python and PHP (Ruby already implements this support).With SORT BY you need that all the weight keys are in the same Redis instance of the list/set you are trying to sort. In order to make this possible we developed a concept called key tags. A key tag is a special pattern inside a key that, if preset, is the only part of the key hashed in order to select the server for this key. For example in order to hash the key "foo" I simply perform the CRC32 checksum of the whole string, but if this key has a pattern in the form of the characters {...} I only hash this substring. So for example for the key "foo{bared}" the key hashing code will simply perform the CRC32 of "bared". This way using key tags you can ensure that related keys will be stored on the same Redis instance just using the same key tag for all this keys. Redis-rb already implements key tags.The latest versions of Redis in the Git repository are able to handle at least 150 million of keys per instance. We are working in order to experiment with larger values.Redis means two things:
+I imagine key-value DBs, in the short term future, to be used like you use memory in a program, with lists, hashes, and so on. With Redis it's like this, but this special kind of memory containing your data structures is shared, atomic, persistent.
When we write code it is obvious, when we take data in memory, to use the most sensible data structure for the work, right? Incredibly when data is put inside a relational DB this is no longer true, and we create an absurd data model even if our need is to put data and get this data back in the same order we put it inside (an ORDER BY is required when the data should be already sorted. Strange, dont' you think?).
Key-value DBs bring this back at home, to create sensible data models and use the right data structures for the problem we are trying to solve.Yes you can. When Redis saves the DB it actually creates a temp file, then rename(2) that temp file name to the destination file name. So even while the server is working it is safe to save the database file just with the cp unix command. Note that you can use master-slave replication in order to have redundancy of data, but if all you need is backups, cp or scp will do the work pretty well.Worst case scenario: 1 Million keys with the key being the natural numbers from 0 to 999999 and the string "Hello World" as value use 100MB on my Intel macbook (32bit). Note that the same data stored linearly in an unique string takes something like 16MB, this is the norm because with small keys and values there is a lot of overhead. Memcached will perform similarly.
With large keys/values the ratio is much better of course.
64 bit systems will use much more memory than 32 bit systems to store the same keys, especially if the keys and values are small, this is because pointers takes 8 bytes in 64 bit systems. But of course the advantage is that you can have a lot of memory in 64 bit systems, so to run large Redis servers a 64 bit system is more or less required.Short answer: If you are using a Redis client that supports consistent hashing you can distribute the dataset across different nodes. For instance the Ruby clients supports this feature. There are plans to develop redis-cluster that basically is a dummy Redis server that is only used in order to distribute the requests among N different nodes using consistent hashing.Redis takes the whole dataset in memory and writes asynchronously on disk in order to be very fast, you have the best of both worlds: hyper-speed and persistence of data, but the price to pay is exactly this, that the dataset must fit on your computers RAM.
If the data is larger then memory, and this data is stored on disk, what happens is that the bottleneck of the disk I/O speed will start to ruin the performances. Maybe not in benchmarks, but once you have real load from multiple clients with distributed key accesses the data must come from disk, and the disk is damn slow. Not only, but Redis supports higher level data structures than the plain values. To implement this things on disk is even slower.
Redis will always continue to hold the whole dataset in memory because this days scalability requires to use RAM as storage media, and RAM is getting cheaper and cheaper. Today it is common for an entry level server to have 16 GB of RAM! And in the 64-bit era there are no longer limits to the amount of RAM you can have in theory.
Amazon EC2 now provides instances with 32 or 64 GB of RAM.You may try to load a dataset larger than your memory in Redis and see what happens, basically if you are using a modern Operating System, and you have a lot of data in the DB that is rarely accessed, the OS's virtual memory implementation will try to swap rarely used pages of memory on the disk, to only recall this pages when they are needed. If you have many large values rarely used this will work. If your DB is big because you have tons of little values accessed at random without a specific pattern this will not work (at low level a page is usually 4096 bytes, and you can have different keys/values stored at a single page. The OS can't swap this page on disk if there are even few keys used frequently).
Another possible solution is to use both MySQL and Redis at the same time, basically take the state on Redis, and all the things that get accessed very frequently: user auth tokens, Redis Lists with chronologically ordered IDs of the last N-comments, N-posts, and so on. Then use MySQL as a simple storage engine for larger data, that is just create a table with an auto-incrementing ID as primary key and a large BLOB field as data field. Access MySQL data only by primary key (the ID). The application will run the high traffic queries against Redis but when there is to take the big data will ask MySQL for specific resources IDs.
Update: it could be interesting to test how Redis performs with datasets larger than memory if the OS swap partition is in one of this very fast Intel SSD disks.Yes, in order to support datasets bigger than RAM there is the plan to
+implement transparent Virtual Memory in Redis, that is, the ability to
+transfer large values associated to keys rarely used on Disk, and
+reload them transparently in memory when this values are requested in
+some way.
So you may ask why don't let the operating system VM do the work for
+us. There are two main reasons: in Redis even a large value stored at
+a given key, for instance a 1 million elements list, is not allocated
+in a contiguous piece of memory. It's actually very fragmented since
+Redis uses quite aggressive object sharing and allocated Redis Objects
+structures reuse.
So you can imagine the memory layout composed of 4096 bytes pages that
+actually contain different parts of different large values. Not only,
+but a lot of values that are large enough for us to swap out to disk,
+like a 1024k value, is just one quarter the size of a memory page, and
+likely in the same page there are other values that are not rarely
+used. So this value wil never be swapped out by the operating system.
+This is the first reason for implementing application-level virtual
+memory in Redis.
There is another one, as important as the first. A complex object in
+memory like a list or a set is something 10 times bigger than the
+same object serialized on disk. Probably you already noticed how Redis
+snapshots on disk are damn smaller compared to the memory usage of
+Redis for the same objects. This happens because when data is in
+memory is full of pointers, reference counters and other metadata. Add
+to this malloc fragmentation and need to return word-aligned chunks of
+memory and you have a clear picture of what happens. So this means to
+have 10 times the I/O between memory and disk than otherwise needed.Yes, try to compile it with 32 bit target if you are using a 64 bit box.
If you are using Redis >= 1.3, try using the Hash data type, it can save a lot of memory.
If you are using hashes or any other type with values bigger than 128 bytes try also this to lower the RSS usage (Resident Set Size): EXPORT MMAP_THRESHOLD=4096This may happen and it's prefectly ok. Redis objects are small C structures allocated and freed a lot of times. This costs a lot of CPU so instead of being freed, released objects are taken into a free list and reused when needed. This memory is taken exactly by this free objects ready to be reused.With modern operating systems malloc() returning NULL is not common, usually the server will start swapping and Redis performances will be disastrous so you'll know it's time to use more Redis servers or get more RAM.
The INFO command (work in progress in this days) will report the amount of memory Redis is using so you can write scripts that monitor your Redis servers checking for critical conditions.
You can also use the "maxmemory" option in the config file to put a limit to the memory Redis can use. If this limit is reached Redis will start to reply with an error to write commands (but will continue to accept read-only commands).Redis uses a lot more memory when compiled for 64 bit target, especially if the dataset is composed of many small keys and values. Such a database will, for instance, consume 50 MB of RAM when compiled for the 32 bit target, and 80 MB for 64 bit! That's a big difference.
You can run 32 bit Redis binaries in a 64 bit Linux and Mac OS X system without problems. For OS X just use make 32bit. For Linux instead, make sure you have libc6-dev-i386 installed, then use make 32bit if you are using the latest Git version. Instead for Redis <= 1.2.2 you have to edit the Makefile and replace "-arch i386" with "-m32".
If your application is already able to perform application-level sharding, it is very advisable to run N instances of Redis 32bit against a big 64 bit Redis box (with more than 4GB of RAM) instead than a single 64 bit instance, as this is much more memory efficient. Just an example on normal hardware: It takes about 45 seconds to restore a 2 GB database on a fairly standard system, no RAID. This can give you some kind of feeling about the order of magnitude of the time needed to load data when you restart the server.Short answer: echo 1 > /proc/sys/vm/overcommit_memory :)
And now the long one:
Redis background saving schema relies on the copy-on-write semantic of fork in modern operating systems: Redis forks (creates a child process) that is an exact copy of the parent. The child process dumps the DB on disk and finally exits. In theory the child should use as much memory as the parent being a copy, but actually thanks to the copy-on-write semantic implemented by most modern operating systems the parent and child process will share the common memory pages. A page will be duplicated only when it changes in the child or in the parent. Since in theory all the pages may change while the child process is saving, Linux can't tell in advance how much memory the child will take, so if the overcommit_memory setting is set to zero fork will fail unless there is as much free RAM as required to really duplicate all the parent memory pages, with the result that if you have a Redis dataset of 3 GB and just 2 GB of free memory it will fail.
Setting overcommit_memory to 1 says Linux to relax and perform the fork in a more optimistic allocation fashion, and this is indeed what you want for Redis.
A good source to understand how Linux Virtual Memory work and other alternatives for overcommit_memory and overcommit_ratio is this classic from Red Hat Magaize, "Understanding Virtual Memory": http://www.redhat.com/magazine/001nov04/features/vm/ Yes, redis background saving process is always fork(2)ed when the server is outside of the execution of a command, so every command reported to be atomic in RAM is also atomic from the point of view of the disk snapshot.Simply start multiple instances of Redis in different ports in the same box and threat them as different servers! Given that Redis is a distributed database anyway in order to scale you need to think in terms of multiple computational units. At some point a single box may not be enough anyway.
In general key-value databases are very scalable because of the property that different keys can stay on different servers independently.
In Redis there are client libraries such Redis-rb (the Ruby client) that are able to handle multiple servers automatically using consistent hashing. We are going to implement consistent hashing in all the other major client libraries. If you use a different language you can implement it yourself otherwise just hash the key before to SET / GET it from a given server. For example imagine to have N Redis servers, server-0, server-1, ..., server-N. You want to store the key "foo", what's the right server where to put "foo" in order to distribute keys evenly among different servers? Just perform the crc = CRC32("foo"), then servernum = crc % N (the rest of the division for N). This will give a number between 0 and N-1 for every key. Connect to this server and store the key. The same for gets.
This is a basic way of performing key partitioning, consistent hashing is much better and this is why after Redis 1.0 will be released we'll try to implement this in every widely used client library starting from Python and PHP (Ruby already implements this support).With SORT BY you need that all the weight keys are in the same Redis instance of the list/set you are trying to sort. In order to make this possible we developed a concept called key tags. A key tag is a special pattern inside a key that, if preset, is the only part of the key hashed in order to select the server for this key. For example in order to hash the key "foo" I simply perform the CRC32 checksum of the whole string, but if this key has a pattern in the form of the characters {...} I only hash this substring. So for example for the key "foo{bared}" the key hashing code will simply perform the CRC32 of "bared". This way using key tags you can ensure that related keys will be stored on the same Redis instance just using the same key tag for all this keys. Redis-rb already implements key tags.In theory Redis can handle up to 232 keys, and was tested in practice to handle at least 150 million of keys per instance. We are working in order to experiment with larger values.
Every list, set, and ordered set, can hold 232 elements.
Actually Redis internals are ready to allow up to 264 elements but the current disk dump format don't support this, and there is a lot time to fix this issues in the future as currently even with 128 GB of RAM it's impossible to reach 232 elements.Redis means two things:
- it's a joke on the word Redistribute (instead to use just a Relational DB redistribute your workload among Redis servers)
- it means REmote DIctionary Server
In order to scale LLOOGG. But after I got the basic server working I liked the idea to share the work with other guys, and Redis was turned into an open source project.