fix(statistics): use subquery instead of join to avoid cartesian product #558
Conversation
The way we filter in the statistics view, any added filter (that affects reports) adds a "dimension" to the cartesian product, exploding the total number of hours reported. Instead of using JOIN, we do EXISTS(SUBQUERY) now, which should avoid this issue. Might be a tiny bit slower, but let's try to make it correct first, then fast.
winged
force-pushed
the
fix_result_explosion_in_statistics
branch
from
084550f to
0fcff27
Compare
There was a problem hiding this comment.
I agree with that. The performance bump is relatively high with small numbers but still in an acceptable range in human time. It's very soon outperformed on larger numbers, so I think that's definitely the way to go.
I would suggest that the EXISTS variant is generally preferable where large number of relations potientially exist. Even if most queries are not eligible, those few that acutally trigger those high-dimension joins will affect every other query. So basically, starting from a certain size of a table that might explode the query, that pattern is preferable. And, given the relatively small trade off when below the threshold, it's probably safe to start from here and fine tune on demand.
|
Note this was a one-shot attempt at fixing our issues, created during a hack session with @trowik and @MitanOmar . However it doesn't solve the issue and I think a general refactoring of the whole statistics backend is in order. |
The way we filter in the statistics view, any added filter (that affects reports) adds a "dimension" to the cartesian product, exploding the total number of hours reported.
Instead of using JOIN, we do EXISTS(SUBQUERY) now, which should avoid this issue. Might be a tiny bit slower, but let's try to make it correct first, then fast.
Fixes #89