You are browsing a read-only backup copy of Wikitech. The live site can be found at

User:AndreaWest/Blazegraph Features and Capabilities

From Wikitech-static
< User:AndreaWest
Revision as of 00:19, 26 May 2022 by imported>AndreaWest (→‎SPARQL Functional Extensions: Discussed GeoSPARQL geometries and added detail on migrating the Blazegraph functions)
Jump to navigation Jump to search

The following is a list of Blazegraph-specific features and capabilities used by WDQS and its community. Defining alternative implementations that minimize the user impact is of critical importance.

Overview of Blazegraph-Specific Features and Capabilities

  • SPARQL functionality extensions
    • Typically, SPARQL is extended by new datatypes and functions
      • The current Blazegraph implementation has a mix of function extensions (geof:distance, geof:globe, geof:latitude, geof:longitude and wikibase:decodeUri) and SERVICE extensions (such as wikibase:label, wikibase:around or wikibase:mwapi)
      • Whereas functions return a single value, the WDQS SERVICES provide multiple outputs
      • Each of the current datatypes/functions and SERVICES are discussed below
  • Named subqueries
    • Documentation
    • Note that although support for subqueries is required for SPARQL compliance, naming is not a compliant feature
    • It is likely that subqueries will not be name-able
    • Based on the placement of the subquery in the overall SPARQL and the use of query hints, a subquery's order of execution can be controlled

SPARQL Functional Extensions

In order to support the current Blazegraph functional extensions, creation of similar custom SPARQL functions would be needed. This capability is supported by all the Blazegraph alternative backends.

  • The functions, geof:globe, geof:latitude and geof:longitude, are simple decompositions of the geometry data of a POINT
    • Typically, the "geof:" prefix represents the namespace defined by
    • Note that since these are NOT valid functions in GeoSPARQL, it seems inappropriate to reference them using the "geof:" namespace
      • Another namespace prefix (such as "wdqs:") is recommended
        • "wdqs:" is recommended (as opposed to "wikibase:") to avoid confusion between the functions, such as wdqs:latitude, which would be used in a SPARQL SELECT, BIND or FILTER clause, and the corresponding value property (wikibase:geoLatitude) which is used in the RDF data as a predicate
          • For example, one might query, "?item p:P625 ?coordinate_statement. ?coordinate_statement psv:P625 ?coordinate_node. ?coordinate_node wikibase:geoLatitude ?lat."
    • In both Wikidata and GeoSPARQL, a geometric POINT can be expressed using a WKT (well-known text) representation, which specifies a coordinate system, followed by a longitude and latitude
      • The coordinate system (also known as the spatial reference system) is defined either by WGS84 on Earth or identified by an item ID (within right and left carets, '<' and '>') which specifies a non-Earth/planetary body
      • geof:globe/wdqs:globe retrieves the coordinate system
      • geof:latitude/wdqs:latitude and geof:longitude/wdqs:longitude functions split the POINT data into its component parts
  • The function, wikibase:decodeURI, can be defined using the logic at
    • Recommend that it use the prefix, "wdqs:", for consistency with the geospatial functions

It was mentioned above that the RDF dump contains the properties, wikibase:geoGlobe, wikibase:geoLatitude and wikibase:geoLongitude. These are value properties that can be used to construct GeoSPARQL POINT geometries. The following query shows how this would be done:

@prefix geo: <> .
CONSTRUCT {?item geo:hasGeometry ?geoLoc} 
?item p:P625 ?coordinate_stmt .
?coordinate_stmt psv:P625 ?coordinate_node.
?coordinate_node wikibase:geoLongitude ?lon.
?coordinate_node wikibase:geoLatitude ?lat.  
OPTIONAL {?coordinate_node wikibase:geoGlobe ?globe}
BIND(IF(!BOUND(?globe) || ?globe = <>, 
        CONCAT("Point(", str(?lon), " ", str(?lat), ")"),
        CONCAT("<", str(?globe), "> Point(", str(?lon), " ", str(?lat), ")"))
     AS ?geoLoc)

Geospatial Support Using GeoSPARQL

The last Blazegraph property extension is geof:distance. That is supported by GeoSPARQL, and is identified as geof:distance. Blazegraph's geof:distance takes as input two string-based POINTs and returns the distance between them in kilometers. The GeoSPARQL function, geof:distance, also supports the input of two POINTs and adds a third parameter, units (which could be defaulted in the code base to kilometers). Note that GeoSPARQL 1.0 has only a few basic units of measure defined. These are adequate for use with Wikidata. The proposed GeoSPARQL 1.1 specification indicates the use of the Quantities, Units, Dimensions and Types ontology (QUDT), which is much broader.

The previous paragraph glossed over a detail which is important when migrating Wikidata to GeoSPARQL compliance. Whereas Wikidata and the Blazegraph implementation use a GlobeCoordinate declaration for a point location, GeoSPARQL uses geometries. Therefore, when loading the RDF dump into the alternative database, the query shown directly above should be performed and the new triples added to the dump. Ideally, this could be done in Wikidata itself.

Beyond the geof:distance function, there are other valuable GeoSPARQL features and functions which could be used in Wikidata queries. These include:

  • Specification of geometries/locations beyond POINTs, such as POLYGONs (which are specified as a group of POINTs that define the geometry's boundary)
  • geof:buffer function, which conceptualizes the space around a geometry (such as a POINT), where the space is defined by a radius given by some units
  • geof:envelope function, which returns the minimal bounding box for an input geometry
    • Given a complex POLYGON, the function would return another POLYGON defining the 4 corners of the minimal bounding box
  • Specification of topology relation functions which compare two geometries and return a boolean indicating if they meet the criteria of the function:
    • geof:sfEquals, returns true if the 2 geometries are equal
    • geof:sfDisjoint, returns true if the 2 geometries are disjoint/separate (inverse of geof:sfEquals)
    • geof:sfIntersects, returns true if any part of the first geometry overlaps with any part of the second
    • geof:sfTouches, returns true if a boundary of the first geometry comes into contact with the boundary of the second (but the interiors of the geometries do NOT intersect)
    • geof:sfCrosses, returns true if the interior of the first geometry comes into contact with the interior or boundary of the second
    • geof:sfWithin, returns true if the second geometry completely encloses the first
    • geof:sfContains, returns true if the first geometry completely encloses the second

Note that some of the above will be used to address the Blazegraph geospatial SERVICEs (wikibase:around and wikibase:box), as explained below.

SERVICE Extensions

This section describes how the WDQS- and Blazegraph-specific SERVICEs (wikibase:label, wikibase:mwapi, wikibase:around, wikibase:box, gas:service, bd:sample and bd:slice) could be supported moving forward.

The geospatial SERVICES, wikibase:around and :box, can (and should) be provided by the use of GeoSPARQL. Details and examples are discussed below.

Unfortunately, there is no straightforward, functional approach for supporting wikibase:mwapi, the GAS service and bd:sample. The problem is that these SERVICEs return multiple (possibly many) results, and some execute based on complex parameters that are defined using unique triple patterns. That combination of requirements does not translate into the standard SPARQL function extensions, which take a set of predefined parameters and return a single result. In order to support these SERVICEs, modifications to the backend code bases will be required - to distinguish local SERVICE IRIs from HTTP federated requests, and then invoke appropriate "handlers".

Note that this discussion did not reference the bd:slice and wikibase:label SERVICEs. bd:slice functionality can be provided by a judicious use of sub-queries. On the other hand, label details can be provided using a SPARQL function extension, although that function will be less convenient than (but with equivalent capabilities to) the existing SERVICE approach. The inconvenience will be due to the need to repeat language preferences. The alternatives for bd:slice and wikibase:label are described in more detail below.

wikibase:around and wikibase:box

It is reasonable to replace the wikibase:around and :box SERVICEs with graph patterns that utilize the GeoSPARQL topology relation functions discussed above. This approach might be most easily explained by using examples.

Let us first examine a query using the wikibase:around SERVICE, which finds airports within 100km of Berlin:

SELECT ?place ?location ?dist WHERE {
  wd:Q64 wdt:P625 ?berlinLoc .       # Berlin coordinates
  SERVICE wikibase:around { 
      ?place wdt:P625 ?location . 
      bd:serviceParam wikibase:center ?berlinLoc . 
      bd:serviceParam wikibase:radius "100" . 
      bd:serviceParam wikibase:distance ?dist.
  FILTER EXISTS { ?place wdt:P31/wdt:P279* wd:Q1248784 }    # Is an airport
} ORDER BY ASC(?dist)

This could be written as:

prefix uom: <>
SELECT ?place ?location ?dist WHERE {
  wd:Q64 wdt:P625 ?berlinLoc .             # Berlin location
  ?place wdt:P31/wdt:P279* wd:Q1248784 ;   # Get airports
         wdt:P625 ?location .              # And their coordinates
  BIND (geof:distance(?berlinLoc, ?location, uom:meter) as ?dist) .
  FILTER (?dist <= 100000)
} ORDER BY ASC(?dist)

But, that approach may have poor performance if there are a large number of "?places" (here, airports) that are retrieved. As written, the query is retrieving all relevant places and their locations, then calculating distance and filtering out the results.

Alternately, the check could be accomplished by the following query:

prefix uom: <>
SELECT ?place ?location ?dist WHERE {
  {  # Create a geometry/spatial object of the area surrounding Berlin
     SELECT ?berlinLoc ?aroundBerlinLoc WHERE {
        wd:Q64 wdt:P625 ?berlinLoc .       # Berlin location
        BIND (geof:buffer(?berlinLoc, 100000, uom:meter) as ?aroundBerlinLoc) }   
  ?place wdt:P31/wdt:P279* wd:Q1248784 ;   # Get airports
         wdt:P625 ?location .              # And their coordinates
  # Filter if the airport location is within the specified geometry
  FILTER (geof:sfWithin(?location, ?aroundBerlinLoc)) .
  BIND (geof:distance(?berlinLoc, ?location, uom:meter) as ?dist) .    # Get the actual distance after filtering
} ORDER BY ASC(?dist)

In order to support the wikibase:box functionality, a similar approach is taken - although geof:buffer is replaced by a custom geof:box function. For example, this query using wikibase:box finds all schools between San Jose and San Francisco CA:

SELECT ?place ?location WHERE {
  wd:Q62 wdt:P625 ?point1 .     # San Francisco location
  wd:Q16553 wdt:P625 ?point2 .  # San Jose location
  SERVICE wikibase:box {
    ?place wdt:P625 ?location .
    bd:serviceParam wikibase:cornerWest ?point1 .
    bd:serviceParam wikibase:cornerEast ?point2 .
  FILTER EXISTS { ?place wdt:P31/wdt:P279* wd:Q3914 }   # Get schools

It becomes:

prefix uom: <>
SELECT ?place ?location WHERE {
     SELECT ?boundingBox WHERE {
        wd:Q62 wdt:P625 ?westPoint .        # San Francisco location
        wd:Q16553 wdt:P625 ?eastPoint .     # San Jose location
        BIND (geof:box(?westPoint, ?eastPoint) as ?boundingBox) }  
  ?place wdt:P31/wdt:P279* wd:Q3914 ;   # Get schools
         wdt:P625 ?location .           # And their coordinates
  # Filter if the school location is within the specified bounding box
  FILTER (geof:sfWithin(?location, ?boundingBox))   

Note that the above proposes a new function (geof:box) that constructs a bounding polygon based on two POINT locations - where the first parameter is the western-most point and the second parameter is the eastern-most point. The latter is simply a decomposition of the two POINTs into their latitudes and longitudes, and then the creation of a POLYGON using the SPARQL STRDT function. The function could be provided for convenience. If not provided, the functionality is implemented using the following graph patterns:

BIND (geof:latitude(?westPoint) as ?westLat) .
BIND (geof:longitude(?westPoint) as ?westLong).
BIND (geof:latitude(?eastPoint) as ?eastLat) .
BIND (geof:longitude(?eastPoint) as ?eastLong) .
# Note that a POLYGON must be closed (e.g., begin and end at the same POINT)
BIND (CONCAT("POLYGON(", STR(?westLong), " ", STR(?westLat), ", ", STR(?eastLong), " ", STR(?westLat), ", ",
             STR(?eastLong), " ", STR(?eastLat), ", ", STR(?westLong), " ", STR(?eastLat), ", ",
             STR(?westLong), " ", STR(?westLat), ")") as ?polygonString ) .   
BIND (STRDT(?polygonString, geo:wktLiteral) as ?boundingBox) .


The wikibase:label SERVICE provides an easy means to retrieve rdfs:label, skos:altLabel and schema:description values for an entity. Its main uses are to simplify the SPARQL query and to provide language preferences for the text that is returned. The latter is the more significant aspect of the SERVICE and is the main focus of the functions defined here.

The label SERVICE could be implemented in 3 new SPARQL functions, each returning a string literal:

string literal wikibase:label (variable var, "string_of_language_codes")
string literal wikibase:altLabel (variable var, "string_of_language_codes")
string literal wikibase:description (variable var, "string_of_language_codes")

These functions would be used in BIND statements to associate specific variable names to the returned texts, which would then be referenced in the query's SELECT clause or used later in the query, for example in a FILTER statement.

As an example of the use of the current label SERVICE, the following query lists the US presidents and their spouses:

SELECT ?p ?pLabel ?w ?wLabel WHERE {
   wd:Q30 p:P6/ps:P6 ?p .
   ?p wdt:P26 ?w .
   SERVICE wikibase:label {
    bd:serviceParam wikibase:language "en" .

This would be (re)written as:

SELECT ?p ?pLabel ?w ?wLabel WHERE {
   wd:Q30 p:P6/ps:P6 ?p .
   ?p wdt:P26 ?w .
   BIND (wikibase:label(?p, "en") as ?pLabel)
   BIND (wikibase:label(?w, "en") as ?wLabel)

As another example, consider this query which uses the manual mode of the label SERVICE:

     SERVICE wikibase:label {
       bd:serviceParam wikibase:language "fr,de,en" .
       wd:Q123 rdfs:label ?q123Label .
       wd:Q123 skos:altLabel ?q123Alt .
       wd:Q123 schema:description ?q123Desc .
       wd:Q321 rdfs:label ?q321Label .

This would be written as:

     BIND (wikibase:label(wd:Q123, "fr,de,en") as ?q123Label) .
     BIND (wikibase:altLabel(wd:Q123, "fr,de,en") as ?q123Alt) .
     BIND (wikibase:description(wd:Q123, "fr,de,en") as ?q123Desc) .
     BIND (wikibase:label(wd:Q321, "fr,de,en") as ?q321Label) .

The downside of this approach is the need to repeat the language preferences in each function call.


The functionality of bd:slice is discussed in the code, Slice Service Factory documentation. In its simplest form, it provides a means to get a subset of results. However, the same functionality can be provided by using a sub-query with a limit/offset.

Let's illustrate this with an example. The query below returned 3743 results in 37074 ms. (The query without bd:slice, with the WHERE clause, "?item wdt:P31 wd:Q13442814. MINUS {?item wdt:P577 ?date}", timed out.)

# Work-around for query for scholarly articles with no date of publication (which times out without bd:slice)
SELECT ?item ?itemLabel 
  SERVICE bd:slice {
    ?item wdt:P31 wd:Q13442814.
    bd:serviceParam bd:slice.limit 1000000   # 1M items returned
  minus {
    ?item wdt:P577 ?date.
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en". }

The same functionality can be achieved by using a sub-query, as follows:

SELECT ?item ?itemLabel 
    SELECT ?item WHERE { ?item wdt:P31 wd:Q13442814 } LIMIT 1000000
  minus {
    ?item wdt:P577 ?date.
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en". }

Running this query returned 3743 results but took 49671 ms.

The other use of bd:slice is to return a count (using the bd:slice.range predicate). There no queries in the February 2022 set that used this predicate, and one query in March (shown below). It returned 1 result (the count of triples = 151295) in 111 ms.

  SERVICE bd:slice
    ?item wdt:P6039 ?o .
    bd:serviceParam bd:slice.range ?range .

This query can be rewritten using a simple SPARQL COUNT feature. It also returns 1 result (151295) but in 205 ms.

SELECT (COUNT(*) as ?range) WHERE
  ?item wdt:P6039 ?o .  

Note that the timings above do vary based on caching of results.

wikibase:mwapi, gas:service and bd:sample

The remainder of the Blazegraph SERVICEs are each described on the following pages:

In order to provide similar functionality, each of the backend code bases would have to be modified to distinguish a SERVICE invocation addressed to a local IRI (e.g., with the prefix, "urn:", "wikibase:" or similar) and an actual, external HTTP endpoint. That checking could occur:

  1. When the SPARQL is being parsed (its algebra/semantics are being defined)
  2. While iterating through/executing the component clauses of the query
  3. By modifying the SERVICE processing itself

The latter two options are likely preferable - since the backend infrastructure would already account for variable bindings and combining results into the final solution.

When executing a local IRI/SERVICE, it is most logical to check a registry of possible "handlers" and then invoke the appropriate code or return an error. The graph patterns of the SERVICE clause and current variable bindings would be passed to the "handler" code, as is done for all SERVICEs. Results would have to be returned consistent with the SPARQL 1.1 Federated Query specification, meaning that they would be an array of variable-RDF term bindings.

It is likely that the current SERVICE implementations would need to be adapted to the design points of the specific backends, but the majority of the processing logic should be able to be reused.

Note that one of the backend alternatives (Apache Jena) already has hooks for providing custom SERVICES. This implementation takes the approach of invoking the custom SERVICE while iterating through the query processing (bullet #2, above). Unfortunately, at the time of writing (late April 2022), there is no documentation related to its use. There is, however, a simple test scenario defined.

Frequency of Use of the Blazegraph SERVICE Extensions

Modifying the existing implementations to support local SERVICE extensions and adjusting the logic of those extensions to execute in the particular backend environment may be costly and/or introduce errors. In addition, the Wikidata documentation related to the Blazegraph-specific extensions (gas:service and bd:sample) states that support may be discontinued at some time in the future. Making such a call will require discussion with the community.

To inform discussion, the following shows the usage statistics of these extensions (across all queries issued in February 2022):

Usage of Custom SERVICE Extensions
SERVICE Percentage of queries
wikibase:mwapi 11.88%
gas:service 0.024%
bd:sample 0.002%

As an aside, the bd:slice SERVICE (discussed above) is used in 0.04% of the February 2022 queries (significantly higher than gas:service or bd:sample).

Of these extensions, the MWAPI SERVICE is the most critical to support.

More Detail on the Use of the GAS SERVICE

The GAS SERVICE is used infrequently to count the number of "hops" between two items and/or to find the shortest path between them (based on a breadth-first or shortest path algorithm). Examples of these types of queries are:

A review by James Heald (noted in found that "All of these queries ... use either the BFS "breadth first search" or the SSSP "single source shortest path" gas service, with the two seemingly completely interchangeable -- the only difference seems to be that BFS returns a whole number for the number of hops found to each node, whereas SSSP seems to return a real number. But the performance of the two seems to be entirely similar for the queries they are being used on, with identical results if BFS is replaced by SSSP or vice-versa."

Note that there is a possible alternative to implement counting hops. For example, to find the ancestors of the composer Bach, see this query. It works by first finding the ancestors and then by counting the generations (per the Phabricator ticket listed above, "hat-tip to Tony Bowden and Andrew Gray, 2018"). However, the approach is not usable in general because it is inefficient, cannot be stopped at a certain depth/count or when a specific item is found, and fails when traversing symmetric properties (e.g., it could "circle back" on itself). In addition, the approach assumes that there is only one path from the starting point, which may not be true. Due to the latter restriction, it would NOT work to find a shortest path given different alternatives.

Since there is no alternative for the GAS SERVICE, it should be supported. However, the priority of providing that support will be influenced by the low frequency of its use.

As another example of a similar capability, but different implementation, see Stardog's PATH query syntax.