Importing relational data into a CubicWeb instance

Introduction

This tutorial explains how to import data from an external source (e.g. a collection of files) into a CubicWeb cube instance.

First, once we know the format of the data we wish to import, we devise a data model, that is, a CubicWeb (Yams) schema which reflects the way the data is structured. This schema is implemented in the schema.py file. In this tutorial, we will describe such a schema for a particular data set, the Diseasome data (see below).

Once the schema is defined, we create a cube and an instance. The cube is a specification of an application, whereas an instance is the application per se.

Once the schema is defined and the instance is created, the import can be performed, via the following steps:

  1. Build a custom parser for the data to be imported. Thus, one obtains a Python memory representation of the data.

  2. Map the parsed data to the data model defined in schema.py.

  3. Perform the actual import of the data. This comes down to “populating” the data model with the memory representation obtained at 1, according to the mapping defined at 2.

This tutorial illustrates all the above steps in the context of relational data stored in the RDF format.

More specifically, we describe the import of Diseasome RDF/OWL data.

Building a data model

The first thing to do when using CubicWeb for creating an application from scratch is to devise a data model, that is, a relational representation of the problem to be modeled or of the structure of the data to be imported.

In such a schema, we define an entity type (EntityType objects) for each type of entity to import. Each such type has several attributes. If the attributes are of known CubicWeb (Yams) types, viz. numbers, strings or characters, then they are defined as attributes, as e.g. attribute = Int() for an attribute named attribute which is an integer.

Each such type also has a set of relations, which are defined like the attributes, except that they represent, in fact, relations between the entities of the type under discussion and the objects of a type which is specified in the relation definition.

For example, for the Diseasome data, we have two types of entities, genes and diseases. Thus, we create two classes which inherit from EntityType:

class Disease(EntityType):
    # Corresponds to http://www.w3.org/2000/01/rdf-schema#label
    label = String(maxsize=512, fulltextindexed=True)
    ...

    #Corresponds to http://www4.wiwiss.fu-berlin.de/diseasome/resource/diseasome/associatedGene
    associated_genes = SubjectRelation('Gene', cardinality='**')
    ...

    #Corresponds to 'http://www4.wiwiss.fu-berlin.de/diseasome/resource/diseasome/chromosomalLocation'
    chromosomal_location = SubjectRelation('ExternalUri', cardinality='?*', inlined=True)


class Gene(EntityType):
    ...

In this schema, there are attributes whose values are numbers or strings. Thus, they are defined by using the CubicWeb / Yams primitive types, e.g., label = String(maxsize=12). These types can have several constraints or attributes, such as maxsize. There are also relations, either between the entity types themselves, or between them and a CubicWeb type, ExternalUri. The latter defines a class of URI objects in CubicWeb. For instance, the chromosomal_location attribute is a relation between a Disease entity and an ExternalUri entity. The relation is marked by the CubicWeb / Yams SubjectRelation method. The latter can have several optional keyword arguments, such as cardinality which specifies the number of subjects and objects related by the relation type specified. For example, the '?*' cardinality in the chromosomal_relation relation type says that zero or more Disease entities are related to zero or one ExternalUri entities. In other words, a Disease entity is related to at most one ExternalUri entity via the chromosomal_location relation type, and that we can have zero or more Disease entities in the data base. For a relation between the entity types themselves, the associated_genes between a Disease entity and a Gene entity is defined, so that any number of Gene entities can be associated to a Disease, and there can be any number of Disease s if a Gene exists.

Of course, before being able to use the CubicWeb / Yams built-in objects, we need to import them:

from yams.buildobjs import EntityType, SubjectRelation, String, Int
from cubicweb.schemas.base import ExternalUri

Building a custom data parser

The data we wish to import is structured in the RDF format, as a text file containing a set of lines. On each line, there are three fields. The first two fields are URIs (“Universal Resource Identifiers”). The third field is either an URI or a string. Each field bares a particular meaning:

  • the leftmost field is an URI that holds the entity to be imported. Note that the entities defined in the data model (i.e., in schema.py) should correspond to the entities whose URIs are specified in the import file.

  • the middle field is an URI that holds a relation whose subject is the entity defined by the leftmost field. Note that this should also correspond to the definitions in the data model.

  • the rightmost field is either an URI or a string. When this field is an URI, it gives the object of the relation defined by the middle field. When the rightmost field is a string, the middle field is interpreted as an attribute of the subject (introduced by the leftmost field) and the rightmost field is interpreted as the value of the attribute.

Note however that some attributes (i.e. relations whose objects are strings) have their objects defined as strings followed by ^^ and by another URI; we ignore this part.

Let us show some examples:

  • of line holding an attribute definition: <http://www4.wiwiss.fu-berlin.de/diseasome/resource/genes/CYP17A1> <http://www.w3.org/2000/01/rdf-schema#label> "CYP17A1" . The line contains the definition of the label attribute of an entity of type gene. The value of label is ‘CYP17A1’.

  • of line holding a relation definition: <http://www4.wiwiss.fu-berlin.de/diseasome/resource/diseases/1> <http://www4.wiwiss.fu-berlin.de/diseasome/resource/diseasome/associatedGene> <http://www4.wiwiss.fu-berlin.de/diseasome/resource/genes/HADH2> . The line contains the definition of the associatedGene relation between a disease subject entity identified by 1 and a gene object entity defined by HADH2.

Thus, for parsing the data, we can (:note: see the diseasome_parser module):

  1. define a couple of regular expressions for parsing the two kinds of lines, RE_ATTS for parsing the attribute definitions, and RE_RELS for parsing the relation definitions.

  2. define a function that iterates through the lines of the file and retrieves (yield s) a (subject, relation, object) tuple for each line. We called it _retrieve_structure in the diseasome_parser module. The function needs the file name and the types for which information should be retrieved.

Alternatively, instead of hand-making the parser, one could use the RDF parser provided in the dataio cube.

Once we get to have the (subject, relation, object) triples, we need to map them into the data model.

Mapping the data to the schema

In the case of diseasome data, we can just define two dictionaries for mapping the names of the relations as extracted by the parser, to the names of the relations as defined in the schema.py data model. In the diseasome_parser module they are called MAPPING_ATTS and MAPPING_RELS. Given that the relation and attribute names are given in CamelCase in the original data, mappings are necessary if we follow the PEP08 when naming the attributes in the data model. For example, the RDF relation chromosomalLocation is mapped into the schema relation chromosomal_location.

Once these mappings have been defined, we just iterate over the (subject, relation, object) tuples provided by the parser and we extract the entities, with their attributes and relations. For each entity, we thus have a dictionary with two keys, attributes and relations. The value associated to the attributes key is a dictionary containing (attribute: value) pairs, where “value” is a string, plus the cwuri key / attribute holding the URI of the entity itself. The value associated to the relations key is a dictionary containing (relation: value) pairs, where “value” is an URI. This is implemented in the entities_from_rdf interface function of the module diseasome_parser. This function provides an iterator on the dictionaries containing the attributes and relations keys for all entities.

However, this is a simple case. In real life, things can get much more complicated, and the mapping can be far from trivial, especially when several data sources (which can follow different formatting and even structuring conventions) must be mapped into the same data model.

Importing the data

The data import code should be placed in a Python module. Let us call it diseasome_import.py. Then, this module should be called via cubicweb-ctl, as follows:

cubicweb-ctl shell diseasome_import.py -- <other arguments e.g. data file>

In the import module, we should use a store for doing the import. A store is an object which provides three kinds of methods for importing data:

  • a method for importing the entities, along with the values of their attributes.

  • a method for importing the relations between the entities.

  • a method for committing the imports to the database.

In CubicWeb, we have four stores:

  1. ObjectStore base class for the stores in CubicWeb. It only provides a skeleton for all other stores and provides the means for creating the memory structures (dictionaries) that hold the entities and the relations between them.

  2. RQLObjectStore: store which uses the RQL language for performing database insertions and updates. It relies on all the CubicWeb hooks machinery, especially for dealing with security issues (database access permissions).

  1. NoHookRQLObjectStore: store which uses the RQL language for performing database insertions and updates, but for which all hooks are deactivated. This implies that certain checks with respect to the CubicWeb / Yams schema (data model) are not performed. However, all SQL queries obtained from the RQL ones are executed in a sequential manner, one query per inserted entity.

  1. SQLGenObjectStore: store which uses the SQL language directly. It inserts entities either sequentially, by executing an SQL query for each entity, or directly by using one PostGRES COPY FROM query for a set of similarly structured entities.

For really massive imports (millions or billions of entities), there is a cube dataio which contains another store, called MassiveObjectStore. This store is similar to SQLGenObjectStore, except that anything related to CubicWeb is bypassed. That is, even the CubicWeb EID entity identifiers are not handled. This store is the fastest, but has a slightly different API from the other four stores mentioned above. Moreover, it has an important limitation, in that it doesn’t insert inlined 1 relations in the database.

1

An inlined relation is a relation defined in the schema with the keyword argument inlined=True. Such a relation is inserted in the database as an attribute of the entity whose subject it is.

In the following section we will see how to import data by using the stores in CubicWeb’s dataimport module.

Using the stores in dataimport

ObjectStore is seldom used in real life for importing data, since it is only the base store for the other stores and it doesn’t perform an actual import of the data. Nevertheless, the other three stores, which import data, are based on ObjectStore and provide the same API.

All three stores RQLObjectStore, NoHookRQLObjectStore and SQLGenObjectStore provide exactly the same API for importing data, that is entities and relations, in an SQL database.

Before using a store, one must import the dataimport module and then initialize the store, with the current session as a parameter:

import cubicweb.dataimport as cwdi
...

store = cwdi.RQLObjectStore(session)

Each such store provides three methods for data import:

  1. create_entity(Etype, **attributes), which allows us to add an entity of the Yams type Etype to the database. This entity’s attributes are specified in the attributes dictionary. The method returns the entity created in the database. For example, we add two entities, a person, of Person type, and a location, of Location type:

    person = store.create_entity('Person', name='Toto', age='18', height='190')
    
    location = store.create_entity('Location', town='Paris', arrondissement='13')
    
  2. relate(subject_eid, r_type, object_eid), which allows us to add a relation of the Yams type r_type to the database. The relation’s subject is an entity whose EID is subject_eid; its object is another entity, whose EID is object_eid. For example 2:

    store.relate(person.eid(), 'lives_in', location.eid(), **kwargs)
    

    kwargs is only used by the SQLGenObjectStore’s relate method and is here to allow us to specify the type of the subject of the relation, when the relation is defined as inlined in the schema.

2
The eid method of an entity defined via create_entity returns

the entity identifier as assigned by CubicWeb when creating the entity. This only works for entities defined via the stores in the CubicWeb’s dataimport module.

The keyword argument that is understood by SQLGenObjectStore is called subjtype and holds the type of the subject entity. For the example considered here, this comes to having 3:

store.relate(person.eid(), 'lives_in', location.eid(), subjtype=person.cw_etype)

If subjtype is not specified, then the store tries to infer the type of the subject. However, this doesn’t always work, e.g. when there are several possible subject types for a given relation type.

3
The cw_etype attribute of an entity defined via create_entity holds

the type of the entity just created. This only works for entities defined via the stores in the CubicWeb’s dataimport module. In the example considered here, person.cw_etype holds 'Person'.

All the other stores but SQLGenObjectStore ignore the kwargs parameters.

  1. flush(), which allows us to perform the actual commit into the database, along with some cleanup operations. Ideally, this method should be called as often as possible, that is after each insertion in the database, so that database sessions are kept as atomic as possible. In practice, we usually call this method twice: first, after all the entities have been created, second, after all relations have been created.

    Note however that before each commit the database insertions have to be consistent with the schema. Thus, if, for instance, an entity has an attribute defined through a relation (viz. a SubjectRelation) with a "1" or "+" object cardinality, we have to create the entity under discussion, the object entity of the relation under discussion, and the relation itself, before committing the additions to the database.

    The flush method is simply called as:

    store.flush().
    

Using the MassiveObjectStore in the dataio cube

This store, available in the dataio cube, allows us to fully dispense with the CubicWeb import mechanisms and hence to interact directly with the database server, via SQL queries.

Moreover, these queries rely on PostGreSQL’s COPY FROM instruction to create several entities in a single query. This brings tremendous performance improvements with respect to the RQL-based data insertion procedures.

However, the API of this store is slightly different from the API of the stores in CubicWeb’s dataimport module.

Before using the store, one has to import the dataio cube’s dataimport module, then initialize the store by giving it the session parameter:

from cubicweb_dataio import dataimport as mcwdi
...

store = mcwdi.MassiveObjectStore(session)

The MassiveObjectStore provides six methods for inserting data into the database:

  1. init_rtype_table(SubjEtype, r_type, ObjEtype), which specifies the creation of the tables associated to the relation types in the database. Each such table has three column, the type of the subject entity, the type of the relation (that is, the name of the attribute in the subject entity which is defined via the relation), and the type of the object entity. For example:

    store.init_rtype_table('Person', 'lives_in', 'Location')
    

    Please note that these tables can be created before the entities, since they only specify their types, not their unique identifiers.

  2. create_entity(Etype, **attributes), which allows us to add new entities, whose attributes are given in the attributes dictionary. Please note however that, by default, this method does not return the created entity. The method is called, for example, as in:

    store.create_entity('Person', name='Toto', age='18', height='190',
                        uri='http://link/to/person/toto_18_190')
    store.create_entity('Location', town='Paris', arrondissement='13',
                        uri='http://link/to/location/paris_13')
    

    In order to be able to link these entities via the relations when needed, we must provide ourselves a means for uniquely identifying the entities. In general, this is done via URIs, stored in attributes like uri or cwuri. The name of the attribute is irrelevant as long as its value is unique for each entity.

  3. relate_by_iid(subject_iid, r_type, object_iid) allows us to actually relate the entities uniquely identified by subject_iid and object_iid via a relation of type r_type. For example:

    store.relate_by_iid('http://link/to/person/toto_18_190',
                        'lives_in',
                        'http://link/to/location/paris_13')
    

    Please note that this method does not work for inlined relations!

  4. convert_relations(SubjEtype, r_type, ObjEtype, subj_iid_attribute, obj_iid_attribute) allows us to actually insert the relations in the database. At one call of this method, one inserts all the relations of type rtype between entities of given types. subj_iid_attribute and object_iid_attribute are the names of the attributes which store the unique identifiers of the entities, as assigned by the user. These names can be identical, as long as their values are unique. For example, for inserting all relations of type lives_in between People and Location entities, we write:

    store.convert_relations('Person', 'lives_in', 'Location', 'uri', 'uri')
    
  5. flush() performs the actual commit in the database. It only needs to be called after create_entity and relate_by_iid calls. Please note that relate_by_iid does not perform insertions into the database, hence calling flush() for it would have no effect.

  6. cleanup() performs database cleanups, by removing temporary tables. It should only be called at the end of the import.

Application to the Diseasome data

Import setup

We define an import function, diseasome_import, which does basically four things:

  1. creates and initializes the store to be used, via a line such as:

    store = cwdi.SQLGenObjectStore(session)
    

    where cwdi is the imported cubicweb.dataimport or cubicweb_dataio.dataimport.

  2. calls the diseasome parser, that is, the entities_from_rdf function in the diseasome_parser module and iterates on its result, in a line such as:

    for entity, relations in parser.entities_from_rdf(filename, ('gene', 'disease')):
    

    where parser is the imported diseasome_parser module, and filename is the name of the file containing the data (with its path), e.g. ../data/diseasome_dump.nt.

  3. creates the entities to be inserted in the database; for Diseasome, there are two kinds of entities:

    1. entities defined in the data model, viz. Gene and Disease in our case.

    2. entities which are built in CubicWeb / Yams, viz. ExternalUri which define URIs.

    As we are working with RDF data, each entity is defined through a series of URIs. Hence, each “relational attribute” 4 of an entity is defined via an URI, that is, in CubicWeb terms, via an ExternalUri entity. The entities are created, in the loop presented above, as such:

    ent = store.create_entity(etype, **entity)
    

    where etype is the appropriate entity type, either Gene or Disease.

4
By “relational attribute” we denote an attribute (of an entity) which

is defined through a relation, e.g. the chromosomal_location attribute of Disease entities, which is defined through a relation between a Disease and an ExternalUri.

The ExternalUri entities are as many as URIs in the data file. For them, we define a unique attribute, uri, which holds the URI under discussion:

extu = store.create_entity('ExternalUri', uri="http://path/of/the/uri")
  1. creates the relations between the entities. We have relations between:

    1. entities defined in the schema, e.g. between Disease and Gene entities, such as the associated_genes relation defined for Disease entities.

    2. entities defined in the schema and ExternalUri entities, such as gene_id.

    The way relations are added to the database depends on the store:

    • for the stores in the CubicWeb dataimport module, we only use store.relate, in another loop, on the relations (that is, a loop inside the preceding one, mentioned at step 2):

      for rtype, rels in relations.iteritems():
          ...
      
          store.relate(ent.eid(), rtype, extu.eid(), **kwargs)
      

      where kwargs is a dictionary designed to accommodate the need for specifying the type of the subject entity of the relation, when the relation is inlined and SQLGenObjectStore is used. For example:

      ...
      store.relate(ent.eid(), 'chromosomal_location', extu.eid(), subjtype='Disease')
      
    • for the MassiveObjectStore in the dataio cube’s dataimport module, the relations are created in three steps:

      1. first, a table is created for each relation type, as in:

        ...
        store.init_rtype_table(ent.cw_etype, rtype, extu.cw_etype)
        

        which comes down to lines such as:

        store.init_rtype_table('Disease', 'associated_genes', 'Gene')
        store.init_rtype_table('Gene', 'gene_id', 'ExternalUri')
        
      2. second, the URI of each entity will be used as its identifier, in the relate_by_iid method, such as:

        disease_uri = 'http://www4.wiwiss.fu-berlin.de/diseasome/resource/diseases/3'
        gene_uri = '<http://www4.wiwiss.fu-berlin.de/diseasome/resource/genes/HSD3B2'
        store.relate_by_iid(disease_uri, 'associated_genes', gene_uri)
        
      3. third, the relations for each relation type will be added to the database, via the convert_relations method, such as in:

        store.convert_relations('Disease', 'associated_genes', 'Gene', 'cwuri', 'cwuri')
        

        and:

        store.convert_relations('Gene', 'hgnc_id', 'ExternalUri', 'cwuri', 'uri')
        

        where cwuri and uri are the attributes which store the URIs of the entities defined in the data model, and of the ExternalUri entities, respectively.

  2. flushes all relations and entities:

    store.flush()
    

    which performs the actual commit of the inserted entities and relations in the database.

If the MassiveObjectStore is used, then a cleanup of temporary SQL tables should be performed at the end of the import:

store.cleanup()

Timing benchmarks

In order to time the import script, we just decorate the import function with the timed decorator:

from logilab.common.decorators import timed
...

@timed
def diseasome_import(session, filename):
    ...

After running the import function as shown in the “Importing the data” section, we obtain two time measurements:

diseasome_import clock: ... / time: ...

Here, the meanings of these measurements are 5:

  • clock is the time spent by CubicWeb, on the server side (i.e. hooks and data pre- / post-processing on SQL queries),

  • time is the sum between clock and the time spent in PostGreSQL.

5

The meanings of the clock and time measurements, when using the @timed decorators, were taken from a blog post on massive data import in CubicWeb.

The import function is put in an import module, named diseasome_import here. The module is called directly from the CubicWeb shell, as follows:

cubicweb-ctl shell diseasome_instance diseasome_import.py \
-- -df diseasome_import_file.nt -st StoreName

The module accepts two arguments:

  • the data file, introduced by -df [--datafile], and

  • the store, introduced by -st [--store].

The timings (in seconds) for different stores are given in the following table, for importing 4213 Disease entities and 3919 Gene entities with the import module just described:

Store

CubicWeb time (clock)

PostGreSQL time (time - clock)

Total time

RQLObjectStore

225.98

62.05

288.03

NoHookRQLObjectStore

62.73

51.38

114.11

SQLGenObjectStore

20.41

11.03

31.44

MassiveObjectStore

4.84

6.93

11.77

Conclusions

In this tutorial we have seen how to import data in a CubicWeb application instance. We have first seen how to create a schema, then how to create a parser of the data and a mapping of the data to the schema. Finally, we have seen four ways of importing data into CubicWeb.

Three of those are integrated into CubicWeb, namely the RQLObjectStore, NoHookRQLObjectStore and SQLGenObjectStore stores, which have a common API:

  • RQLObjectStore is by far the slowest, especially its time spent on the CubicWeb side, and so it should be used only for small amounts of “sensitive” data (i.e. where security is a concern).

  • NoHookRQLObjectStore slashes by almost four the time spent on the CubicWeb side, but is also quite slow; on the PostGres side it is as slow as the previous store. It should be used for data where security is not a concern, but consistency (with the data model) is.

  • SQLGenObjectStore slashes by three the time spent on the CubicWeb side and by five the time spent on the PostGreSQL side. It should be used for relatively great amounts of data, where security and data consistency are not a concern. Compared to the previous store, it has the disadvantage that, for inlined relations, we must specify their subjects’ types.

For really huge amounts of data there is a fourth store, MassiveObjectStore, available from the dataio cube. It provides a blazing performance with respect to all other stores: it is almost 25 times faster than RQLObjectStore and almost three times faster than SQLGenObjectStore. However, it has a few usage caveats that should be taken into account:

  1. it cannot insert relations defined as inlined in the schema,

  2. no security or consistency check is performed on the data,

  3. its API is slightly different from the other stores.

Hence, this store should be used when security and data consistency are not a concern, and there are no inlined relations in the schema.