Skip to content
/ zoom Public

A blazing-fast datastore and querying engine for Go built on Redis.


Notifications You must be signed in to change notification settings


Repository files navigation


Version Circle CI GoDoc

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

Table of Contents

Development Status

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.

When is Zoom a Good Fit?

Zoom might be a good fit if:

  1. 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.
  2. 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.
  3. 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:

  1. 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.
  2. 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 or OR 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 to pull down the current master branch, or install with the dependency manager of your choice to lock in a specific version.


First, add to your import statement:

import (
	 // ...

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)


What is a Model?

Models in Zoom are just structs which implement the zoom.Model interface:

type Model interface {
  ModelID() 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

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.

Customizing Field Names

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"`

If you don't want a field to be saved in Redis at all, you can use the special struct tag redis:"-".

Creating Collections

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:

  1. The collection name cannot contain a colon.
  2. Queries, as well as the FindAll, DeleteAll, and Count methods will not work if Index is false. 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

Saving Models

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.

Updating Models

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.

Finding a Single Model

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.

Finding Only Certain Fields

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.

Finding All Models

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.

Deleting Models

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.

Counting the Number of Models

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
// 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 ReplyHandlers. 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 Models. You can also write your own custom ReplyHandlers if needed.


The Query Object

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.

Using Query Modifiers

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

A Note About String Indexes

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.

More Information


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:

Concurrent Updates and Optimistic Locking

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
		redis.Args{postKey, "Likes", 1},
	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:

Testing & Benchmarking

Running the Tests

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  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

Running the Benchmarks

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.



Example Usage

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.


A blazing-fast datastore and querying engine for Go built on Redis.









Contributors 4