I would use a stream-based xslt transformation to import all the records. Here is a simple stylesheet storing one XML record in a key/value table:

<xsl:stylesheet xmlns:xsl="&lt;a href="http://www.w3.org/1999/XSL/Transform" "="" rel="nofollow">http://www.w3.org/1999/XSL/Transform' version='1.0' 
    xmlns:c="http://www.ncbi.nlm.nih.gov"
    > 
<xsl:output method="xml" omit-xml-declaration="yes"/>
<xsl:template match="/">
<xsl:apply-templates select="//c:PC-Compound"/>
</xsl:template>



<xsl:template match="c:PC-Compound">
insert into table(id,content)
values(
<xsl:value-of select="c:PC-Compound_id/c:PC-CompoundType/c:PC-CompoundType_id/c:PC-CompoundType_id_cid"/>
,
'<xsl:copy-of select="."/>'
);
</xsl:template>

</xsl:stylesheet>

(note: the quotes in the XML should be escaped)

Here is an example using my old tool xsltstream: ( http://plindenbaum.blogspot.fr/2010/02/processing-large-xml-documents-with.html )

$ curl -s "ftp://ftp.ncbi.nlm.nih.gov/pubchem/Compound/CURRENT-Full/XML/Compound_000000001_000025000.xml.gz" | gunzip -c |  java -jar xsltstream.jar -x stylesheet.xsl -q "PC-Compound"

insert into table(id,content)
values(
1
,
'<PC-Compound xmlns="&lt;a href=" http:="" www.ncbi.nlm.nih.gov"="" rel="nofollow">http://www.ncbi.nlm.nih.gov" xmlns:xs="http://www.w3.org/2001/XMLSchema-instance">
    <PC-Compound_id>
      <PC-CompoundType>
        <PC-CompoundType_id>
          <PC-CompoundType_id_cid>1</PC-CompoundType_id_cid>
        </PC-CompoundType_id>
      </PC-CompoundType>
     (...)
        <PC-Count_isotope-atom>0</PC-Count_isotope-atom>
        <PC-Count_covalent-unit>1</PC-Count_covalent-unit>
        <PC-Count_tautomers>1</PC-Count_tautomers>
      </PC-Count>
    </PC-Compound_count>
  </PC-Compound>'
);

insert into table(id,content)
values(
2
,
'<PC-Compound xmlns="&lt;a href=" http:="" www.ncbi.nlm.nih.gov"="" rel="nofollow">http://www.ncbi.nlm.nih.gov">
    <PC-Compound_id>