When I implemented this in 9a4c24c, there were not many rows in
entity_documentproperties in the database here. Now that there are,
computing the document_document -> entity_documentproperties ->
entity_entity join and then filtering is really, really slow. Postgres
seems to materialize the whole join and then scan it.
If we get a set of matching document IDs for the entity query in a
subquery, and then just filter with IN on that, things are much faster:
scan entity_entity; in a nested loop, get the document_ids via
entity_documentproperties; hash this set; and then scan
document_document.
Searching for a single character, this brings the query from ~1.1s to
~400ms. Searching for a full word, ~800ms to 120ms
This condition is getting really ugly -- I am sorry!
References #2935
This should have been included with a8dcbbb, which changed the
related_name to access DocumentProperties from Document. (There's no
actual change to the database.)
Without this fix, a condition like:
{key: 'id', operator: '==', value: 'A/B'}
gets mapped to:
public_id__exact=('A/B'.lower())
which is wrong.
I introduced this bug in b3df5b8. I didn't catch it because I was
mostly interested in the 'layer' key -- but layer names are
conventionally lowercase anyway so lowercasing them had no effect.
Fetching documents for each entity in turn is expensive. (I have tried
using ArrayAgg to fetch them in the same query as the Entity — no
improvement. It's possible that being able to join to entity_entity,
and then use ArrayAgg, would be better.)
Even once you've fetched them all, if the same entity appears many
times in an item, then get(..., keys=['layers']) duplicates the whole
JSON for the entity many times: expensive to serialize, expensive to
send over the wire.
Pandora's own web interface only depends on the 'id' key of 'entity' in
each annotation, and refetches the rest of the entity to show the pop-up
dialog when you press E. So by just not bothering to fetch and send any
other keys, get(..., keys=['layers']) on an item with many entity
annotations is substantially faster.
(I experimented with splitting the full entities off to one side, so,
you'd have:
{
"layers": {
somelayer: [...,
{..., "entity": {"id": ABC}},
], ...
},
"entities": {
ABC: {...},
...
}
}
This is quicker than the status quo, but obviously not as fast as not
fetching & sending the rest at all!)
Clip.public_id uses self.item.public_id.
Clip.json() uses self.item.sort, so we should select_related on that
rather than the clip's own sort field. (They are identical objects. Is
Clip.sort ever used directly?)
With this change, findClips() issues one query to fetch clips plus one
query per flavour of annotation; before, it issued two extra queries per
clip.
Requiring layer to have the right case is consistent with
addAnnotation(), and means the _layer[_like] index can be used. In my
testing, if itemsQuery specifies a single item, then postgres doesn't
bother with the layer index anyway; but if not, it makes a pretty big
(~3×) difference.
Matching public_id and item__public_id case-sensitively also seems
reasonable (it's consistent with get() and getAnnotation()).
(Is lower() redundant for the case-insensitive comparisons? ie. is
UPPER(x.lower()) == UPPER(x)? I'm not sure, it's cheap, let's leave it.)
This is about 30% faster, presumably because it avoids allocation and/or
closing over variables is slow(?). It's not hugely significant (I
misread a line_profile report) but why not.
