Reindexer

Reindexer is an embeddable, in-memory, document-oriented database with a high-level Query builder interface.

Reindexer's goal is to provide fast search with complex queries. We at Restream weren't happy with Elasticsearch and created Reindexer as a more performant alternative.

The core is written in C++ and the application level API is in Go.

Documents and indices are stored in dense binary C++ structs, so they don't impose any load on Go's garbage collector.
String duplicates are merged.
Memory overhead is about 32 bytes per document + ≈4-16 bytes per each search index.
There is an object cache on the Go level for deserialized documents produced after query execution. Future queries use pre-deserialized documents, which cuts repeated deserialization and allocation costs
The Query interface uses sync.Pool for reusing internal structures and buffers. Combining of these techings let's Reindexer execute most of queries without any allocations.

Full text search

Reindexer has internal full text search engine. Full text search usage documentation and examples are here

Disk Storage

Reindexer can store documents to and load documents from disk via LevelDB. Documents are written to the storage backend asynchronously by large batches automatically in background.

When a namespace is created, all its documents are stored into RAM, so the queries on these documents run entirely in in-memory mode.

Usage

Here is complete example of basic Reindexer usage:

package main

// Import package
import (
	"fmt"
	"math/rand"

	"github.com/restream/reindexer"
	// choose how the Reindexer binds to the app (in this case "builtin," which means link Reindexer as a static library)
	_ "github.com/restream/reindexer/bindings/builtin"
)

// Define struct with reindex tags
type Item struct {
	ID       int64  `reindex:"id,,pk"`    // 'id' is primary key
	Name     string `reindex:"name"`      // add index by 'name' field
	Articles []int  `reindex:"articles"`  // add index by articles 'articles' array
	Year     int    `reindex:"year,tree"` // add sortable index by 'year' field
}

func main() {
	// Init a database instance and choose the binding
	db := reindexer.NewReindex("builtin")

	// Enable persistent storage (optional)
	db.EnableStorage("/tmp/reindex/")

	// Create new namespace with name 'items', which will store structs of type 'Item'
	db.OpenNamespace("items", reindexer.DefaultNamespaceOptions(), Item{})

	// Generate dataset
	for i := 0; i < 100000; i++ {
		err := db.Upsert("items", &Item{
			ID:       int64(i),
			Name:     "Vasya",
			Articles: []int{rand.Int() % 100, rand.Int() % 100},
			Year:     2000 + rand.Int()%50,
		})
		if err != nil {
			panic(err)
		}
	}

	// Query a single document
	elem, found := db.Query("items").
		Where("id", reindexer.EQ, 40).
		Get()

	if found {
		item := elem.(*Item)
		fmt.Println("Found document:", *item)
	}

	// Query multiple documents
	query := db.Query("items").
		Sort("year", false).                          // Sort results by 'year' field in ascending order
		WhereString("name", reindexer.EQ, "Vasya").   // 'name' must be 'Vasya'
		WhereInt("year", reindexer.GT, 2020).         // 'year' must be greater than 2020
		WhereInt("articles", reindexer.SET, 6, 1, 8). // 'articles' must contain one of [6,1,8]
		Limit(10).                                    // Return maximum 10 documents
		Offset(0).                                    // from 0 position
		ReqTotal()                                    // Calculate the total count of matching documents

	// Execute the query and return an iterator
	iterator := query.Exec()
	// Iterator must be closed
	defer iterator.Close()

	fmt.Println("Found", iterator.TotalCount(), "total documents, first", iterator.Count(), "documents:")

	// Iterate over results
	for iterator.Next() {
		// Get the next document and cast it to a pointer
		elem := iterator.Object().(*Item)
		fmt.Println(*elem)
	}
	// Check the error
	if err := iterator.Error(); err != nil {
		panic(err)
	}
}

SQL compatible interface

As alternative to Query builder Reindexer provides SQL compatible query interface. Here is sample of SQL interface usage:

    ...
	iterator := db.ExecSQL ("SELECT * FROM items WHERE name='Vasya' AND year > 2020 AND articles IN (6,1,8) ORDER BY year LIMIT 10")
    ...

Please note, that Query builder interface is prefferable way: It have more features, and faster than SQL interface

Installation

Prerequirements

Reindexer's core is written in C++11 and uses LevelDB as the storage backend, so the C++11 toolchain and LevelDB must be installed before installing Reindexer.

To build Reindexer, g++ 4.8+ or clang 3.3+ is required.

Optional dependencies

Doxygen package is also required for building a documentation of the project.
cmake, gtest,gbenchmark for run C++ tests and benchmarks
gperftools for memory and performance profiling

Get Reindexer

go get -a github.com/restream/reindexer
bash $GOPATH/src/github.com/restream/reindexer/dependencies.sh
go generate github.com/restream/reindexer

Advanced Usage

Index Types and Their Capabilites

Internally, structs are split into two parts:

indexed fields, marked with reindex struct tag
tuple of non-indexed fields

Queries are possible only on the indexed fields, marked with reindex tag. The reindex tag contains the index name, type, and additional options:

reindex:"<name>[[,<type>],<opts>]"

name – index name.
type – index type:
- hash – fast select by EQ and SET match. Does not allow sorting results by field. Used by default. Allows slow and uneffecient sorting by field
- tree – fast select by RANGE, GT, and LT matches. A bit slower for EQ and SET matches than hash index. Allows fast sorting results by field.
- text – full text search index. Usage details of full text search is described here
- - – column index. Can't perform fast select because it's implemented with full-scan technic. Has the smallest memory overhead.
opts – additional index options:
- pk – field is part of a primary key. Struct must have at least 1 field tagged with pk
- composite – create composite index. The field type must be an empty struct: struct{}.
- joined – field is a recipient for join. The field type must be []*SubitemType.
- dense - reduce index size. For hash and tree it will save 8 bytes per unique key value. For - it will save 4-8 bytes per each element. Useful for indexes with high sectivity, but for tree and hash indexes with low selectivity can seriously decrease update performance. Also dense will slow down wide fullscan queries on - indexes, due to lack of CPU cache optimization.
- collate_numeric - create string index that provides values order in numeric sequence. The field type must be a string.
- collate_ascii - create case-insensitive string index works with ASCII. The field type must be a string.
- collate_utf8 - create case-insensitive string index works with UTF8. The field type must be a string.

Nested Structs

By default Reindexer scans all nested structs and adds their fields to the namespace (as well as indexes specified).

type Actor struct {
	Name string `reindex:"actor_name"`
}

type BaseItem struct {
	ID int64 `reindex:"id,hash,pk"`
}

type ComplexItem struct {
	BaseItem         // Index fields of BaseItem will be added to reindex
	actor    []Actor // Index fields of Actor will be added to reindex as arrays
	Name     string  `reindex:"name"`
	Year     int     `reindex:"year,tree"`
	parent   *Item   `reindex:"-"` // Index fields of parent will NOT be added to reindex
}

Join

Reindexer can join documents from multiple namespaces into a single result:

type Actor struct {
	ID        int    `reindex:"id"`
	Name      string `reindex:"name"`
	IsVisible bool   `reindex:"is_visible"`
}

type ItemWithJoin struct {
	ID        int      `reindex:"id"`
	Name      string   `reindex:"name"`
	ActorsIDs []int    `reindex:"actors_ids"`
	Actors    []*Actor `reindex:"actors,,joined"`
}
....
    
	query := db.Query("items_with_join").Join(
		db.Query("actors").
			WhereBool("is_visible", reindexer.EQ, true),
		"actors",
	).On("id", reindexer.SET, "actors_ids")

	query.Exec ()

In this example, Reindexer uses reflection under the hood to create Actor slice and copy Actor struct.

Joinable interface

To avoid using reflection, Item can implement Joinable interface. If that implemented, Reindexer uses this instead of the slow reflection-based implementation. This increases overall performance by 10-20%, and reduces the amount of allocations.

// Joinable interface implementation.
// Join adds items from the joined namespace to the `ItemWithJoin` object.
// When calling Joinable interface, additional context variable can be passed to implement extra logic in Join.
func (item *ItemWithJoin) Join(field string, subitems []interface{}, context interface{}) {

	switch field {
	case "actors":
		for _, joinItem := range subitems {
			item.Actors = append(item.Actors, joinItem.(*Actor))
		}
	}
}

Complex Primary Keys and Composite Indices

A Document can have multiple fields as its primary key. Reindexer checks unique composition of all pk fields during upserts:

type Item struct {
	ID    int64 `reindex:"id,,pk"`     // 'id' is a part of primary key
	SubID int   `reindex:"sub_id,,pk"` // 'sub_id' is a part of primary key
	// Fields
	//	....
}

Too complex primary key (>2 fields) can slow down upsert and select operations, because Reindexer has to do separate selects to each index, and intersect results.

Composite index feature exists to speed things up. Reindexer can create composite index using multiple fields and use it instead of several separate indices.

type Item struct {
	ID       int64 `reindex:"id,-,pk"`         // 'id' is part of primary key, WITHOUT personal searchable index
	SubID    int   `reindex:"sub_id,-,pk"`     // 'sub_id' is part of primary key, WITHOUT personal searchable index
	SubSubID int   `reindex:"sub_sub_id,-,pk"` // 'sub_sub_id' is part of primary key WITHOUT personal searchable index

	// Fields
	// ....

	// Composite index
	_ struct{} `reindex:"id+sub_id+sub_sub_id,,composite"`
}

Also composite indices are useful for sorting results by multiple fields:

type Item struct {
	ID     int64 `reindex:"id,,pk"`
	Rating int   `reindex:"rating"`
	Year   int   `reindex:"year"`

	// Composite index
	_ struct{} `reindex:"rating+year,tree,composite"`
}

...
	// Sort query resuls by rating first, then by year
	query := db.Query("items").Sort("rating+year", true)

Aggregations

Reindexer allows to do aggregation queries. Currently Average and Sum aggregations are supported. To support aggregation Query has 2 methods: Aggregate and GetAggreatedValue. Aggregate should be called before Query execution - to ask reindexer calculate aggregation and GetAggreatedValue after Query execution to obtain aggregated value

Direct JSON operations

Upsert data in JSON format

If source data is available in JSON format, then it is possible to improve performance of Upsert/Delete operations by directly passing JSON to reindexer. JSON deserialization will be done by C++ code, without extra allocs/deserialization in Go code.

Upsert or Delete functions can process JSON just by passing []byte argument with json

	json := []byte (`{"id":1,"name":"test"}`)
	db.UpsertJSON ("items",json)

It is just faster equalent of:

	item := &Item{}
	json.Unmarshal ([]byte (`{"id":1,"name":"test"}`),item)
	db.Upsert ("items",item)

Get Query results in JSON format

In case of requiment to serialize results of Query in JSON format, then it is possible to improve performance by directly obtaining results in JSON format from reindexer. JSON serialization will be done by C++ code, without extra allocs/serialization in Go code.

...		
	iterator := db.Query("items").
		Select ("id","name").        // Filter output JSON: Select only "id" and "name" fields of items, another fields will be ommited
		Limit (1).
		ExecToJson ("root_object")   // Name of root object of output JSON

	json,err := iterator.JsonAll()
	// Check the error
	if err != nil {
		panic(err)
	}
	fmt.Printf ("%s\n",string (json))
...

This code will print something like:

{"root_object":[{"id":1,"name":"test"}]}

Using object cache

To avoid race conditions, by default object cache is turned off and all objects are allocated and deserialized from reindexer internal format (called CJSON) per each query. The deserialization is uses reflection, so it's speed is not optimal (in fact CJSON deserialization is ~3-10x faster than JSON, and ~1.2x faster than GOB), but perfrormance is still seriously limited by reflection overhead.

There are 2 ways to enable object cache:

Provide DeepCopy interface
Ask query return shared objects from cache

DeepCopy interface

If object is implements DeepCopy intreface, then reindexer will turn on object cache and use DeepCopy interface to copy objects from cache to query results. The DeepCopy interface is responsible to make deep copy of source object.

Here is sample of DeepCopy interface implementation

func (item *Item) DeepCopy () interface {} {
	copyItem := &Item{
		ID: item.ID,
		Name: item.Name,
		Articles: make ([]int,cap (item.Articles),len (item.Articles)),
		Year: item.Year,
	}
	copy (copyItem.Articles,item.Articles)
	return copyItem
}

There are availbale code generation tool gencopy, which can automatically generate DeepCopy interface for structs.

Get shared objects from object cache (USE WITH CAUTION)

To speed up queries and do not allocate new objects per each query it is possible ask query return objects directly from object cache. For enable this behaviour, call AllowUnsafe(true) on Iterator.

WARNING: when used AllowUnsafe(true) queries returns shared pointers to structs in object cache. Therefore application MUST NOT modify returned objects.

	res, err := db.Query("items").WhereInt ("id",reindexer.EQ,1).Exec().AllowUnsafe(true).FetchAll()
	if err != nil {
		panic (err)
	}

	if len (res) > 1 {
		// item is SHARED pointer to struct in object cache
		item = res[0].(*Item)

		// It's OK - fmt.Printf will not modify item
		fmt.Printf ("%v",item)

		// It's WRONG - can race, and will corrupt data in object cache
		item.Name = "new name"
	}

Logging, debug and profiling

Turn on logger

Reindexer logger can be turned on by db.SetLogger() method, just like in this snippet of code:

type Logger struct {
}
func (Logger) Printf(level int, format string, msg ...interface{}) {
	log.Printf(format, msg...)
}
...
	db.SetLogger (Logger{})

Debug queries

Another useful feature is debug print of processed Queries. To debug print queries details there are 2 methods:

db.SetDefaultQueryDebug(namespace string,level int) - it globally enables print details of all queries by namespace
query.Debug(level int) - print details of query execution

level is level of verbosity:

reindexer.INFO - will print only query conditions
reindexer.TRACE - will print query conditions and execution details with timings

Profiling

Because reindexer core is written in C++ all calls to reindexer and their memory consumption are not visible for go profiler. To profile reindexer core there are cgo profiler available. cgo profiler now is part of reindexer, but it can be used with any another cgo code.

Usage of cgo profiler is very similar with usage of go profiler.

Add import:

import _ "github.com/restream/reindexer/pprof"

If your application is not already running an http server, you need to start one. Add "net/http" and "log" to your imports and the following code to your main function:

go func() {
	log.Println(http.ListenAndServe("localhost:6060", nil))
}()

Run application with envirnoment variable HEAPPROFILE=/tmp/pprof
Then use the pprof tool to look at the heap profile:

pprof -symbolize remote http://localhost:6060/debug/cgo/pprof/heap

Limitations and known issues

Currently Reindexer is stable and production ready, but it is still a work in progress, so there are some limitations and issues:

There is no standalone server mode. Only embeded (builtin) binding is supported for now.
Internal C++ API is not stabilized and is subject to change.

VLVLKY / reindexer