cd_motivating_examples

Motivating examples

How many people live in Europe? How many books have been written by Europeans? How many women head cities in Europe? How many distinct objects are in a knowledge graph?

These queries are expressed as SPARQL queries in the queries folder. For instance, the first query about european people is:

PREFIX wd: <http://www.wikidata.org/entity/>
PREFIX wdt: <http://www.wikidata.org/prop/direct/>

SELECT(COUNT(DISTINCT ?person) AS ?cd_persons )
WHERE {
  ?person wdt:P31 wd:Q5 .
  ?person wdt:P19 ?city .
  ?city wdt:P17 ?country .
  ?country wdt:P30 wd:Q46 .
}

Such queries time out on the public SPARQL endpoints of Wikidata that comprises more than 16B triples. However, executing them locally on WDBench that only comprises 1.2B triples, they may not time out anymore.

To execute a query using Blazegraph without running a server, we use the following: java -jar embedded-blazegraph-jar-with-dependencies.jar --database=wdbench-blaze.jnl --file=./queries/books.sparql

Blazegraph using 16 threads evaluates these queries:

• books.sparql: 841ms
• women.sparql: 27.7s
• people.sparql: 39.3s

They do not timeout but are already quite challenging. Therefore we constrain Blazegraph to use only one 1 process using docker run --cpuset-cpus="0" -m=55g.

• books.sparql: 3s
• women.sparql: 105s
• people.sparql: 170s

Therefore, with a timeout threshold set to 60s, the end-user would have had no result for her queries. Using approximate query processing, we are guaranteed to get an approximate result before 60s. Using CRAWD, the approximate results of these queries are available in the results folder. With the configuration involving #tp x 100 scans for each sampled result, and a single thread, CRAWD provides at 60s:

• books.sparql: estimated 3509 vs 3542 expected
• women.sparql: estimated 2849 vs 2877 expected
• people.sparql: estimated 1796949 vs 1804559 expected

While these estimate are already good, a end-user could decide to further refine them by asking for additional rounds of sampling.