N3 dump is very slow since rdflib >= 5.0.0

[chabotsi]

Hi,

We are using rdflib to generate rdf dump in various format (xml, nt, n3) and we have noticed a huuuuge slow down with rdflib >= 5.0.0 when n3 serializing.

For the purpose of this issue, we have limited the export, and we have the following results. It took less than 12s with rdflib 4.2.2 and more than 10 minutes (!) with rdflib 5.0.0 (same results with rdflib 6).

The profiling suggests that with rdflib 5.0.0 the compute_qname() function costs a lot. On our small example, the startswith() method has been called more than 2 billion times !

The corresponding code may have been introduced by: #649

Any help or suggestion would be appreciated !
Thank you,

---

One can find the profiling results below.

• with rdflib 4.2.2
  

• with rdflib 5.0.0
  

[chabotsi]

Hi,

The issue occurs because with have a lot (millions) of resources whose URI are like "http://exemple.com/ark:/XXXX/YYYYY#about". The split_uri method returns http://exemple.com/ark:/XXXX/YYYYY# as the namespace. Therefore, we have millions of namespaces in the __trie and __strie dictionnaries, and make millions of comparisons.

We are wondering about what would be the proper way to tackling this issue − knowing it was working with rdflib 4.2.2. Redefining a new NamespaceManager ? Adapting split_uri for our specific case ? other ?

Thanks for you help,