A blazing-fast datastore and querying engine for Go built on Redis.
Requires Redis version >= 2.8.9 and Go version >= 1.2. The latest version of both is recommended.
Full documentation is available on godoc.org.
- Development Status
- When is Zoom a Good Fit?
- Installation
- Initialization
- Models
- Transactions
- Queries
- More Information
- Testing & Benchmarking
- Contributing
- Example Usage
- License
Zoom was first started in 2013. It is well-tested and going forward the API will be relatively stable. However, it is not actively maintained and there are some known performance issues with queries that use more than one filter.
At this time, Zoom can be considered safe for use in low-traffic production applications. However, I would recommend that you look at more actively maintained alternatives.
Zoom follows semantic versioning, but offers no guarantees of backwards compatibility until version 1.0. You can also keep an eye on the Releases page to see a full changelog for each release. In addition, starting with version 0.9.0, migration guides will be provided for any non-trivial breaking changes, making it easier to stay up to date with the latest version.
Zoom might be a good fit if:
- You are building a low-latency application. Because Zoom is built on top of Redis and all data is stored in memory, it is typically much faster than datastores/ORMs based on traditional SQL databases. Latency will be the most noticeable difference, although throughput may also be improved.
- You want more out of Redis. Zoom offers a number of features that you don't get by using a Redis driver directly. For example, Zoom supports a larger number of types out of the box (including custom types, slices, maps, complex types, and embedded structs), provides tools for making multi-command transactions easier, and of course, provides the ability to run queries.
- You want an easy-to-use datastore. Zoom has a simple API and is arguably easier to use than some ORMs. For example, it doesn't require database migrations and instead builds up a schema based on your struct types. Zoom also does not typically require any knowledge of Redis in order to use effectively. Just connect it to a database and you're good to go!
Zoom might not be a good fit if:
- You are working with a lot of data. Redis is an in-memory database, and Zoom does not yet support sharding or Redis Cluster. Memory could be a hard constraint for larger applications. Keep in mind that it is possible (if expensive) to run Redis on machines with up to 256GB of memory on cloud providers such as Amazon EC2.
- You need advanced queries. Zoom currently only provides support for basic queries and is
not as powerful or flexible as something like SQL. For example, Zoom currently lacks the
equivalent of the
IN
orOR
SQL keywords. See the documentation for a full list of the types of queries supported.
Zoom is powered by Redis and needs to connect to a Redis database. You can install Redis on the same machine that Zoom runs on, connect to a remote database, or even use a Redis-as-a-service provider such as Redis To Go, RedisLabs, Google Cloud Redis, or Amazon Elasticache.
If you need to install Redis, see the installation instructions on the official Redis website.
To install Zoom itself, run go get -u github.com/albrow/zoom
to pull down the
current master branch, or install with the dependency manager of your choice to
lock in a specific version.
First, add github.com/albrow/zoom to your import statement:
import (
// ...
github.com/albrow/zoom
)
Then, you must create a new pool with
NewPool
. A pool represents
a pool of connections to the database. Since you may need access to the pool in
different parts of your application, it is sometimes a good idea to declare a
top-level variable and then initialize it in the main
or init
function. You
must also call pool.Close
when your application exits, so it's a good idea to
use defer.
var pool *zoom.Pool
func main() {
pool = zoom.NewPool("localhost:6379")
defer func() {
if err := pool.Close(); err != nil {
// handle error
}
}()
// ...
}
The NewPool
function accepts an address which will be used to connect to
Redis, and it will use all the
default values
for the other options. If you need to specify different options, you can use the
NewPoolWithOptions
function.
For convenience, the
PoolOptions
type has
chainable methods for changing each option. Typically you would start with
DefaultOptions
and
call WithX
to change value for option X
.
For example, here's how you could initialize a Pool that connects to Redis using
a unix socket connection on /tmp/unix.sock
:
options := zoom.DefaultPoolOptions.WithNetwork("unix").WithAddress("/tmp/unix.sock")
pool = zoom.NewPoolWithOptions(options)
Models in Zoom are just structs which implement the zoom.Model
interface:
type Model interface {
ModelID() string
SetModelID(string)
}
To clarify, all you have to do to implement the Model
interface is add a getter and setter
for a unique id property.
If you want, you can embed zoom.RandomID
to give your model all the
required methods. A struct with zoom.RandomID
embedded will generate a pseudo-random id for itself
the first time the ModelID
method is called iff it does not already have an id. The pseudo-randomly
generated id consists of the current UTC unix time with second precision, an incremented atomic
counter, a unique machine identifier, and an additional random string of characters. With ids generated
this way collisions are extremely unlikely.
Future versions of Zoom may provide additional id implementations out of the box, e.g. one that assigns auto-incremented ids. You are also free to write your own id implementation as long as it satisfies the interface.
A struct definition serves as a sort of schema for your model. Here's an example of a model for a person:
type Person struct {
Name string
Age int
zoom.RandomID
}
Because of the way Zoom uses reflection, all the fields you want to save need to be exported. Unexported fields (including unexported embedded structs with exported fields) will not be saved. This is a departure from how the encoding/json and encoding/xml packages behave. See issue #25 for discussion.
Almost any type of field is supported, including custom types, slices, maps, complex types, and embedded structs. The only things that are not supported are recursive data structures and functions.
You can change the name used to store the field in Redis with the redis:"<name>"
struct tag. So
for example, if you wanted the fields to be stored as lowercase fields in Redis, you could use the
following struct definition:
type Person struct {
Name string `redis:"name"`
Age int `redis:"age"`
zoom.RandomID
}
If you don't want a field to be saved in Redis at all, you can use the special struct tag redis:"-"
.
You must create a Collection
for each type of model you want to save. A
Collection
is simply a set of all models of a specific type and has methods
for saving, finding, deleting, and querying those models. NewCollection
examines the type of a model and uses reflection to build up an internal schema.
You only need to call NewCollection
once per type. Each pool keeps track of
its own collections, so if you wish to share a model type between two or more
pools, you will need to create a collection for each pool.
// Create a new collection for the Person type.
People, err := pool.NewCollection(&Person{})
if err != nil {
// handle error
}
The convention is to name the Collection
the plural of the corresponding
model type (e.g. "People"), but it's just a variable so you can name it
whatever you want.
NewCollection
will use all the
default options
for the collection.
If you need to specify other options, use the
NewCollectionWithOptions
function. The second argument to NewCollectionWithOptions
is a
CollectionOptions
.
It works similarly to PoolOptions
, so you can start with
DefaultCollectionOptions
and use the chainable WithX
methods to specify a new value for option X
.
Here's an example of how to create a new Collection
which is indexed, allowing
you to use Queries and methods like FindAll
which rely on collection indexing:
options := zoom.DefaultCollectionOptions.WithIndex(true)
People, err = pool.NewCollectionWithOptions(&Person{}, options)
if err != nil {
// handle error
}
There are a few important points to emphasize concerning collections:
- The collection name cannot contain a colon.
- Queries, as well as the
FindAll
,DeleteAll
, andCount
methods will not work ifIndex
isfalse
. This may change in future versions.
If you need to access a Collection
in different parts of
your application, it is sometimes a good idea to declare a top-level variable
and then initialize it in the init
function:
var (
People *zoom.Collection
)
func init() {
var err error
// Assuming pool and Person are already defined.
People, err = pool.NewCollection(&Person{})
if err != nil {
// handle error
}
}
Continuing from the previous example, to persistently save a Person
model to
the database, we use the People.Save
method. Recall that in this example,
"People" is just the name we gave to the Collection
which corresponds to the
model type Person
.
p := &Person{Name: "Alice", Age: 27}
if err := People.Save(p); err != nil {
// handle error
}
When you call Save
, Zoom converts all the fields of the model into a format
suitable for Redis and stores them as a Redis hash. There is a wiki page
describing
how zoom works under the hood in more detail.
Sometimes, it is preferable to only update certain fields of the model instead
of saving them all again. It is more efficient and in some scenarios can allow
safer simultaneous changes to the same model (as long as no two clients update
the same field at the same time). In such cases, you can use UpdateFields
.
if err := People.UpdateFields([]string{"Name"}, person); err != nil {
// handle error
}
UpdateFields
uses "last write wins" semantics, so if another caller updates
the same field, your changes may be overwritten. That means it is not safe for
"read before write" updates. See the section on
Concurrent Updates for more
information.
To retrieve a model by id, use the Find
method:
p := &Person{}
if err := People.Find("a_valid_person_id", p); err != nil {
// handle error
}
The second argument to Find
must be a pointer to a struct which satisfies Model
, and must have a type corresponding to
the Collection
. In this case, we passed in Person
since that is the struct type that corresponds to our People
collection. Find
will mutate p
by setting all its fields. Using Find
in this way allows the caller to maintain type
safety and avoid type casting. If Zoom couldn't find a model of type Person
with the given id, it will return a
ModelNotFoundError
.
If you only want to find certain fields in the model instead of retrieving all
of them, you can use FindFields
, which works similarly to UpdateFields
.
p := &Person{}
if err := People.FindFields("a_valid_person_id", []string{"Name"}, p); err != nil {
// handle error
}
fmt.Println(p.Name, p.Age)
// Output:
// Alice 0
Fields that are not included in the given field names will not be mutated. In
the above example, p.Age
is 0
because p
was just initialized and that's
the zero value for the int
type.
To find all models of a given type, use the FindAll
method:
people := []*Person{}
if err := People.FindAll(&people); err != nil {
// handle error
}
FindAll
expects a pointer to a slice of some registered type that implements Model
. It grows or shrinks the slice as needed,
filling in all the fields of the elements inside of the slice. So the result of the call is that people
will be a slice of
all models in the People
collection.
FindAll
only works on indexed collections. To index a collection, you need to
include Index: true
in the CollectionOptions
.
To delete a model, use the Delete
method:
// ok will be true iff a model with the given id existed and was deleted
if ok, err := People.Delete("a_valid_person_id"); err != nil {
// handle err
}
Delete
expects a valid id as an argument, and will attempt to delete the model with the given id. If there was no model
with the given type and id, the first return value will be false.
You can also delete all models in a collection with the DeleteAll
method:
numDeleted, err := People.DeleteAll()
if err != nil {
// handle error
}
DeleteAll
will return the number of models that were successfully deleted.
DeleteAll
only works on indexed collections. To index a collection, you need
to include Index: true
in the CollectionOptions
.
You can get the number of models in a collection using the Count
method:
count, err := People.Count()
if err != nil {
// handle err
}
Count
only works on indexed collections. To index a collection, you need
to include Index: true
in the CollectionOptions
.
Zoom exposes a transaction API which you can use to run multiple commands efficiently and atomically. Under the hood,
Zoom uses a single Redis transaction to perform all the commands in a single
round trip. Transactions feature delayed execution, so nothing touches the database until you call Exec
. A transaction
also remembers its errors to make error handling easier on the caller. The first error that occurs (if any) will be
returned when you call Exec
.
Here's an example of how to save two models and get the new number of models in
the People
collection in a single transaction.
numPeople := 0
t := pool.NewTransaction()
t.Save(People, &Person{Name: "Foo"})
t.Save(People, &Person{Name: "Bar"})
// Count expects a pointer to an integer, which it will change the value of
// when the transaction is executed.
t.Count(People, &numPeople)
if err := t.Exec(); err != nil {
// handle error
}
// numPeople will now equal the number of `Person` models in the database
fmt.Println(numPeople)
// Output:
// 2
You can execute custom Redis commands or run custom Lua scripts inside a
Transaction
using the
Command
and
Script
methods.
Both methods expect a
ReplyHandler
as an
argument. A ReplyHandler
is simply a function that will do something with the
reply from Redis. ReplyHandler
's are executed in order when you call Exec
.
Right out of the box, Zoom exports a few useful ReplyHandler
s. These include
handlers for the primitive types int
, string
, bool
, and float64
, as well
as handlers for scanning a reply into a Model
or a slice of Model
s. You can
also write your own custom ReplyHandler
s if needed.
Zoom provides a useful abstraction for querying the database. You create queries by using the NewQuery
constructor, where you must pass in the name corresponding to the type of model you want to query. For now,
Zoom only supports queries on a single collection at a time.
You can add one or more query modifiers to the query, such as Order
, Limit
, and Filter
. These methods
return the query itself, so you can chain them together. The first error (if any) that occurs due to invalid
arguments in the query modifiers will be remembered and returned when you attempt to run the query.
Finally, you run the query using a query finisher method, such as Run
or Count
. Queries feature delayed
execution, so nothing touches the database until you execute the query with a finisher method.
You can chain a query object together with one or more different modifiers. Here's a list of all the available modifiers:
You can run a query with one of the following query finishers:
Here's an example of a more complicated query using several modifiers:
people := []*Person{}
q := People.NewQuery().Order("-Name").Filter("Age >=", 25).Limit(10)
if err := q.Run(&people); err != nil {
// handle error
}
Full documentation on the different modifiers and finishers is available on godoc.org.
Because Redis does not allow you to use strings as scores for sorted sets, Zoom relies on a workaround
to store string indexes. It uses a sorted set where all the scores are 0 and each member has the following
format: value\x00id
, where \x00
is the NULL character. With the string indexes stored this way, Zoom
can issue the ZRANGEBYLEX command and related commands to filter models by their string values. As a consequence,
here are some caveats to keep in mind:
- Strings are sorted by ASCII value, exactly as they appear in an ASCII table, not alphabetically. This can have surprising effects, for example 'Z' is considered less than 'a'.
- Indexed string values may not contain the NULL or DEL characters (the characters with ASCII codepoints of 0 and 127 respectively). Zoom uses NULL as a separator and DEL as a suffix for range queries.
Zoom is as persistent as the underlying Redis database. If you intend to use Redis as a permanent
datastore, it is recommended that you turn on both AOF and RDB persistence options and set fsync
to
everysec
. This will give you good performance while making data loss highly unlikely.
If you want greater protections against data loss, you can set fsync
to always
. This will hinder performance
but give you persistence guarantees
very similar to SQL databases such as PostgreSQL.
Read more about Redis persistence
All methods and functions in Zoom that touch the database do so atomically. This is accomplished using Redis transactions and Lua scripts when necessary. What this means is that Zoom will not put Redis into an inconsistent state (e.g. where indexes to not match the rest of the data).
However, it should be noted that there is a caveat with Redis atomicity guarantees. If Redis crashes
in the middle of a transaction or script execution, it is possible that your AOF file can become
corrupted. If this happens, Redis will refuse to start until the AOF file is fixed. It is relatively
easy to fix the problem with the redis-check-aof
tool, which will remove the partial transaction
from the AOF file.
If you intend to issue Redis commands directly or run custom scripts, it is highly recommended that
you also make everything atomic. If you do not, Zoom can no longer guarantee that its indexes are
consistent. For example, if you change the value of a field which is indexed, you should also
update the index for that field in the same transaction. The keys that Zoom uses for indexes
and models are provided via the ModelKey
,
AllIndexKey
, and
FieldIndexKey
methods.
Read more about:
Zoom 0.18.0 introduced support for basic optimistic locking. You can use optimistic locking to safely implement concurrent "read before write" updates.
Optimistic locking utilizes the WATCH
, MULTI
, and EXEC
commands in Redis
and only works in the context of transactions. You can use the
Transaction.Watch
method to watch a model for changes. If the model changes after you call Watch
but before you call Exec
, the transaction will not be executed and instead
will return a
WatchError
. You can
also use the WatchKey
method, which functions exactly the same but operates on
keys instead of models.
To understand why optimistic locking is useful, consider the following code:
// likePost increments the number of likes for a post with the given id.
func likePost(postID string) error {
// Find the Post with the given postID
post := &Post{}
if err := Posts.Find(postID, post); err != nil {
return err
}
// Increment the number of likes
post.Likes += 1
// Save the post
if err := Posts.Save(post); err != nil {
return err
}
}
The line post.Likes += 1
is a "read before write" operation. That's because
the +=
operator implicitly reads the current value of post.Likes
and then
adds to it.
This can cause a bug if the function is called across multiple goroutines or
multiple machines concurrently, because the Post
model can change in between
the time we retrieved it from the database with Find
and saved it again with
Save
.
You can use optimistic locking to avoid this problem. Here's the revised code:
// likePost increments the number of likes for a post with the given id.
func likePost(postID string) error {
// Start a new transaction and watch the post key for changes. It's important
// to call Watch or WatchKey *before* finding the model.
tx := pool.NewTransaction()
if err := tx.WatchKey(Posts.ModelKey(postID)); err != nil {
return err
}
// Find the Post with the given postID
post := &Post{}
if err := Posts.Find(postID, post); err != nil {
return err
}
// Increment the number of likes
post.Likes += 1
// Save the post in a transaction
tx.Save(Posts, post)
if err := tx.Exec(); err != nil {
// If the post was modified by another goroutine or server, Exec will return
// a WatchError. You could call likePost again to retry the operation.
return err
}
}
Optimistic locking is not appropriate for models which are frequently updated,
because you would almost always get a WatchError
. In fact, it's called
"optimistic" locking because you are optimistically assuming that conflicts will
be rare. That's not always a safe assumption.
Don't forget that Zoom allows you to run Redis commands directly. This
particular problem might be best solved by the HINCRBY
command.
// likePost atomically increments the number of likes for a post with the given
// id and then returns the new number of likes.
func likePost(postID string) (int, error) {
// Get the key which is used to store the post in Redis
postKey := Posts.ModelKey(postID, post)
// Start a new transaction
tx := pool.NewTransaction()
// Add a command to increment the number of Likes. The HINCRBY command returns
// an integer which we will scan into numLikes.
var numLikes int
tx.Command(
"HINCRBY",
redis.Args{postKey, "Likes", 1},
zoom.NewScanIntHandler(&numLikes),
)
if err := tx.Exec(); err != nil {
return 0, err
}
return numLikes, nil
}
Finally, if optimistic locking is not appropriate and there is no built-in Redis
command that offers the functionality you need, Zoom also supports custom Lua
scripts via the
Transaction.Script
method. Redis is single-threaded and scripts are always executed atomically, so
you can perform complicated updates without worrying about other clients
changing the database.
Read more about:
- Redis Commands
- Redigo, the Redis Driver used by Zoom
ReplyHandler
s provided by Zoom- How Zoom works Under the Hood
To run the tests, make sure you're in the root directory for Zoom and run:
go test
If everything passes, you should see something like:
ok github.com/albrow/zoom 2.267s
If any of the tests fail, please open an issue and describe what happened.
By default, tests and benchmarks will run on localhost:6379 and use database #9. You can change the address, network, and database used with flags. So to run on a unix socket at /tmp/redis.sock and use database #3, you could use:
go test -network=unix -address=/tmp/redis.sock -database=3
To run the benchmarks, make sure you're in the root directory for the project and run:
go test -run=none -bench .
The -run=none
flag is optional, and just tells the test runner to skip the tests and run only the benchmarks
(because no test function matches the pattern "none"). You can also use the same flags as above to change the
network, address, and database used.
You should see some runtimes for various operations. If you see an error or if the build fails, please open an issue.
Here are the results from my laptop (2.8GHz quad-core i7 CPU, 16GB 1600MHz RAM) using a socket connection with Redis set to append-only mode:
BenchmarkConnection-8 5000000 318 ns/op
BenchmarkPing-8 100000 15146 ns/op
BenchmarkSet-8 100000 18782 ns/op
BenchmarkGet-8 100000 15556 ns/op
BenchmarkSave-8 50000 29307 ns/op
BenchmarkSave100-8 3000 546427 ns/op
BenchmarkFind-8 50000 24767 ns/op
BenchmarkFind100-8 5000 374947 ns/op
BenchmarkFindAll100-8 5000 383919 ns/op
BenchmarkFindAll10000-8 30 47267433 ns/op
BenchmarkDelete-8 50000 29902 ns/op
BenchmarkDelete100-8 3000 530866 ns/op
BenchmarkDeleteAll100-8 2000 730934 ns/op
BenchmarkDeleteAll1000-8 200 9185093 ns/op
BenchmarkCount100-8 100000 16411 ns/op
BenchmarkCount10000-8 100000 16454 ns/op
BenchmarkQueryFilterInt1From1-8 20000 82152 ns/op
BenchmarkQueryFilterInt1From10-8 20000 83816 ns/op
BenchmarkQueryFilterInt10From100-8 10000 144206 ns/op
BenchmarkQueryFilterInt100From1000-8 2000 1010463 ns/op
BenchmarkQueryFilterString1From1-8 20000 87347 ns/op
BenchmarkQueryFilterString1From10-8 20000 88031 ns/op
BenchmarkQueryFilterString10From100-8 10000 158968 ns/op
BenchmarkQueryFilterString100From1000-8 2000 1088961 ns/op
BenchmarkQueryFilterBool1From1-8 20000 82537 ns/op
BenchmarkQueryFilterBool1From10-8 20000 84556 ns/op
BenchmarkQueryFilterBool10From100-8 10000 149463 ns/op
BenchmarkQueryFilterBool100From1000-8 2000 1017342 ns/op
BenchmarkQueryOrderInt100-8 3000 386156 ns/op
BenchmarkQueryOrderInt10000-8 30 50011375 ns/op
BenchmarkQueryOrderString100-8 2000 1004530 ns/op
BenchmarkQueryOrderString10000-8 20 77855970 ns/op
BenchmarkQueryOrderBool100-8 3000 387056 ns/op
BenchmarkQueryOrderBool10000-8 30 49116863 ns/op
BenchmarkComplexQuery-8 20000 84614 ns/op
The results of these benchmarks can vary widely from system to system, and so the benchmarks here are really only useful for comparing across versions of Zoom, and for identifying possible performance regressions or improvements during development. You should run your own benchmarks that are closer to your use case to get a real sense of how Zoom will perform for you. High performance is one of the top priorities for this project.
See CONTRIBUTING.md.
albrow/people is an example HTTP/JSON API which uses the latest version of Zoom. It is a simple example that doesn't use all of Zoom's features, but should be good enough for understanding how Zoom can work in a real application.
Zoom is licensed under the MIT License. See the LICENSE file for more information.