Neo4j (CSV) Batch Importer

This software is licensed under the GPLv3 for now. You can ask Neo Technology about a different licensing agreement.

For support for Neo4j and Labels see the 2.0 branch.

You provide one tab separated csv file for nodes and one for relationships (optionally more for indexes)

Example data for the files is a small family network

File format

• tab separated csv files
• Property names in first row.
• If only one file is initially imported, the row number corresponds to the node-id (node 0 is the reference node)
• Property values not listed will not be set on the nodes or relationships.
• Optionally property fields can have a type (defaults to String) indicated with name:type where type is one of (int, long, float, double, boolean, byte, short, char, string). The string value is then converted to that type. Conversion failure will result in abort of the import operation.
• There is a separate "label" type, which should be used for relationship types and/or node labels (when we add 2.0 support)
• Property fields may also be arrays by adding "_array" to the types above and separating the data with commas.
• for non-ascii characters make sure to add -Dfile.encoding=UTF-8 to the commandline arguments
• Optionally automatic indexing of properties can be configured with a header like name:string:users and a configured index in batch.properties like batch_import.node_index=exact then the property name will be indexed in the users index for each row with a value there
• multiple files for nodes and rels, comma separated, without spaces like "node1.csv,node2.csv"
• you can specify concrete node-id's with: i:id
• csv files can be zipped individually as *.gz or *.zip

Examples

There is also a sample directory, please run from the main directory sh sample/import.sh

nodes.csv

name    age works_on
Michael 37  neo4j
Selina  14
Rana    6
Selma   4

rels.csv

start	end	type	    since   counter:int
1     2   FATHER_OF	1998-07-10  1
1     3   FATHER_OF 2007-09-15  2
1     4   FATHER_OF 2008-05-03  3
3     4   SISTER_OF 2008-05-03  5
2     3   SISTER_OF 2007-09-15  7

Execution

mvn clean compile exec:java -Dexec.mainClass="org.neo4j.batchimport.Importer" -Dexec.args="neo4j/data/graph.db nodes.csv rels.csv"

or

java -server -Dfile.encoding=UTF-8 -Xmx4G -jar target/batch-import-jar-with-dependencies.jar neo4j/data/graph.db nodes.csv rels.csv


ynagzet:batchimport mh$ rm -rf target/db
ynagzet:batchimport mh$ mvn clean compile assembly:single
[INFO] Scanning for projects...
[INFO] ------------------------------------------------------------------------
[INFO] Building Simple Batch Importer
[INFO]    task-segment: [clean, compile, assembly:single]
[INFO] ------------------------------------------------------------------------
...
[INFO] Building jar: /Users/mh/java/neo/batchimport/target/batch-import-jar-with-dependencies.jar
[INFO] ------------------------------------------------------------------------
[INFO] BUILD SUCCESSFUL
[INFO] ------------------------------------------------------------------------
ynagzet:batchimport mh$ java -server -Xmx4G -jar target/batch-import-jar-with-dependencies.jar target/db nodes.csv rels.csv
Physical mem: 16384MB, Heap size: 3640MB
use_memory_mapped_buffers=false
neostore.propertystore.db.index.keys.mapped_memory=5M
neostore.propertystore.db.strings.mapped_memory=100M
neostore.propertystore.db.arrays.mapped_memory=215M
neo_store=/Users/mh/java/neo/batchimport/target/db/neostore
neostore.relationshipstore.db.mapped_memory=1000M
neostore.propertystore.db.index.mapped_memory=5M
neostore.propertystore.db.mapped_memory=1000M
dump_configuration=true
cache_type=none
neostore.nodestore.db.mapped_memory=200M
...........................................................................
Importing 7500000 Nodes took 17 seconds
....................................................................................................35818 ms
....................................................................................................39343 ms
....................................................................................................41788 ms
....................................................................................................48897 ms
............
Importing 41246740 Relationships took 170 seconds
212 seconds
ynagzet:batchimport mh$ du -sh target/db/
3,2G	target/db/

Parameters

First parameter MIGHT be the property-file name then it has to end with .properties, then this file will be used and all other parameters are consumed as usual

First parameter is the graph database directory, a new db will be created in the directory except when batch_import.keep_db=true is set in batch.properties.

Second parameter supply a comma separated list of nodes-files

Third parameter supply a comma separated list of relationship-files

It is also possible to specifiy those two file-lists in the config:

batch_import.nodes_files=nodes1.csv[,nodes2.csv]
batch_import.rels_files=rels1.csv[,rels2.csv]

Fourth parameter set of 4 values: node_index users fulltext nodes_index.csv or more generally: node-or-rel-index index-name index-type index-file This parameter set can be repeatedly used, see below. It is also possible to configure this in the config (batch.properties)

batch_import.node_index.users=exact

Indexing

Automatic Indexing

You can automatically index properties of nodes and relationships by adding ":indexName" to the property-header. Just configure the indexes in batch.properties like so:

batch_import.node_index.users=exact

name:string:users    age works_on
Michael 37  neo4j
Selina  14
Rana    6
Selma   4

In the relationships-file you can optionally specify that the start and end-node should be looked up from the index in the same way

name:string:users	name:string:users	type	    since   counter:int
Michael     Selina   FATHER_OF	1998-07-10  1
Michael     Rana   FATHER_OF 2007-09-15  2
Michael     Selma   FATHER_OF 2008-05-03  3
Rana     Selma   SISTER_OF 2008-05-03  5
Selina     Rana   SISTER_OF 2007-09-15  7

Explicit Indexing

Optionally you can add nodes and relationships to indexes.

Add four arguments per each index to command line:

To create a full text node index called users using nodes_index.csv:

node_index users fulltext nodes_index.csv

To create an exact relationship index called worked using rels_index.csv:

rel_index worked exact rels_index.csv

Example command line:

java -server -Xmx4G -jar ../batch-import/target/batch-import-jar-with-dependencies.jar neo4j/data/graph.db nodes.csv rels.csv node_index users fulltext nodes_index.csv rel_index worked exact rels_index.csv

Using Neo4j's Automatic Indexing

The auto-indexing elsewhere in this file pertains to the batch inserter's ability to automatically index. If you want to use this cool feature from the batch inserter, there's a little gotcha. You still need to enable the batch inserter's feature with batch_import.node_index but then instead of specifying the name of a regular index, specify the auto index's name like so:

batch_import.node_index.node_auto_index=exact

Examples

nodes_index.csv

id	name	language
1	Victor Richards	West Frisian
2	Virginia Shaw	Korean
3	Lois Simpson	Belarusian
4	Randy Bishop	Hiri Motu
5	Lori Mendoza	Tok Pisin

rels_index.csv

id	property1	property2
0	cwqbnxrv	rpyqdwhk
1	qthnrret	tzjmmhta
2	dtztaqpy	pbmcdqyc

Configuration

The Importer uses a supplied batch.properties file to be configured:

Memory Mapping I/O Config

Most important is the memory config, you should try to have enough RAM map as much of your store-files to memory as possible.

At least the node-store and large parts of the relationship-store should be mapped. The property- and string-stores are mostly append only so don't need that much RAM. Below is an example for about 6GB RAM, to leave room for the heap and also OS and OS caches.

cache_type=none
use_memory_mapped_buffers=true
# 9 bytes per node
neostore.nodestore.db.mapped_memory=200M
# 33 bytes per relationships
neostore.relationshipstore.db.mapped_memory=3G
# 38 bytes per property
neostore.propertystore.db.mapped_memory=500M
# 60 bytes per long-string block
neostore.propertystore.db.strings.mapped_memory=500M
neostore.propertystore.db.index.keys.mapped_memory=5M
neostore.propertystore.db.index.mapped_memory=5M

Indexes (experimental)

batch_import.node_index.users=exact
batch_import.node_index.articles=fulltext
batch_import.relationship_index.friends=exact

CSV (experimental)

batch_import.csv.quotes=true // default, set to false for faster, experimental csv-reader
batch_import.csv.delim=,

Index-Cache (experimental)

batch_import.mapdb_cache.disable=true

Keep Database (experimental)

batch_import.keep_db=true

Parallel Batch Inserter with Neo4j

Uses the LMAX Disruptor to parallelize operations during batch-insertion.

The 6 operations are:

1. property encoding
2. property-record creation
3. relationship-id creation and forward handling of reverse relationship chains
4. writing node-records
5. writing relationship-records
6. writing property-records

Dependencies:

(1)<--(2)<--(6)
(2)<--(5)-->(3)   
(2)<--(4)-->(3)   

It uses the above dependency setup of disruptor handlers to execute the different concerns in parallel. A ringbuffer of about 2^18 elements is used and a heap size of 5-20G, MMIO configuration within the heap limits.

Execution:

MAVEN_OPTS="-Xmx5G -Xms5G -server -d64 -XX:NewRatio=5" mvn clean test-compile exec:java -Dexec.mainClass=org.neo4j.batchimport.DisruptorTest -Dexec.classpathScope=test

current limitations, constraints:

• have to know max # of rels per node, properties per node and relationship
• relationships have to be pre-sorted by min(start,end)

measurements

We successfully imported 2bn nodes (2 properties) and 20bn relationships (1 property) in 11 hours on an EC2 high-IO instance, with 35 ECU, 60GB RAM, 2TB SSD writing up to 500MB/s, resulting in a store of 1.4 TB. That makes around 500k elements per second.

future improvements:

• stripe writes across store-files (i.e. strip the relationship-record file over 10 handlers, according to CPUs)
• parallelize writing to dynamic string and arraystore too
• change relationship-record updates for backwards pointers to run in a separate handler that is RandomAccessFile-based (or nio2) and just writes the 2 int values directly at file-pos
• add a csv analyser / sorter that
• add support & parallelize index addition
• good support for index based lookup for relationship construction (kv-store, better in-memory structure, e.g. a collection of long[])
• use id-compression internally to save memory in structs (write a CompressedLongArray)
• reuse PropertyBlock, PropertyRecords, RelationshipRecords, NodeRecords, probably subclass them and override getId() etc. or copy the code from the Store's to work with interfaces