searchengines-web/MeiliSearch
searchengines-web/MeiliSearch
1
0
Fork 0
mirror of https://github.com/meilisearch/MeiliSearch synced 2024-12-26 22:50:07 +01:00
Code Issues Releases Wiki Activity

Merge #776

Browse Source 
776: Reduce incremental indexing time of `words_prefix_position_docids` DB r=curquiza a=loiclec

Fixes partially https://github.com/meilisearch/milli/issues/605

The `words_prefix_position_docids` can easily contain millions of entries. Thus, iterating
over it can be very expensive. But we do so needlessly for every document addition tasks.

It can sometimes cause indexing performance issues when :
- a user sends many `documentAdditionOrUpdate` tasks that cannot be all batched together (for example if they are interspersed with `documentDeletion` tasks)
- the documents contain long, diverse text fields, thus increasing the number of entries in `words_prefix_position_docids`
- the index has accumulated many soft-deleted documents, further increasing the size of `words_prefix_position_docids`
- the machine running Meilisearch does not have great IO performance (e.g. slow SSD, or quota-limited by the cloud provider)

Note, before approving  the PR: the only changed file should be `milli/src/update/words_prefix_position_docids.rs`.

Co-authored-by: Loïc Lecrenier <loic.lecrenier@me.com>
This commit is contained in:
bors[bot] 2023-01-31 15:52:28 +00:00 committed by GitHub
commit 758b4acea7
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 11 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
milli/src/update/words_prefix_position_docids.rs
View File

@ -140,16 +140,20 @@ impl<'t, 'u, 'i> WordPrefixPositionDocids<'t, 'u, 'i> {
// We remove all the entries that are no more required in this word prefix position
// docids database.
let mut iter =
self.index.word_prefix_position_docids.iter_mut(self.wtxn)?.lazily_decode_data();
while let Some(((prefix, _), _)) = iter.next().transpose()? {
if del_prefix_fst_words.contains(prefix.as_bytes()) {
unsafe { iter.del_current()? };
// We also avoid iterating over the whole `word_prefix_position_docids` database if we know in
// advance that the `if del_prefix_fst_words.contains(prefix.as_bytes()) {` condition below
// will always be false (i.e. if `del_prefix_fst_words` is empty).
if !del_prefix_fst_words.is_empty() {
let mut iter =
self.index.word_prefix_position_docids.iter_mut(self.wtxn)?.lazily_decode_data();
while let Some(((prefix, _), _)) = iter.next().transpose()? {
if del_prefix_fst_words.contains(prefix.as_bytes()) {
unsafe { iter.del_current()? };
}
}
drop(iter);
}
drop(iter);
// We finally write all the word prefix position docids into the LMDB database.
sorter_into_lmdb_database(
self.wtxn,