*** 2,7 **** --- 2,13 ---- ---- + *Reduce overhead of bulk inserts* + + As a test, I modified fts2.c to collect multiple inserts with increasing rowids into in-memory structures, then flush them out as a single level-0 segment. When it flushes every 100 documents, performance is increased by up to 40%. This is even better than the gains from doing bulk inserts in the first place. + + ---- + *Prefix-encode first terms* At this time, interior and leaf nodes fully-encode their first terms, and prefix-encode later terms. If we passed the key which decided which child to take down from the interior node, we could prefix-encode the first term against that. Thus, if we have runs of very long terms, we could save the initial hit from fully-encoding the term.
Small. Fast. Reliable.
Choose any three.
Choose any three.
|
|