Using sqlalchemy's subqueryload to fetch tags works like this:
1. Get basic info about posts with query X
2. Copy query X
3. SELECT all tags WHERE post_id IN (SELECT post_ids FROM query X)
4. Associate the resulting tags with the posts
When original query contains .order_by(func.random()), it looks like
this:
1. SELECT post.* FROM post ORDER BY random() LIMIT 10
2. Copy "ORDER BY random() LIMIT 10"
3. SELECT tag.* FROM tag WHERE tag.post_id IN (
SELECT id FROM post ORDER BY random() LIMIT 10)
4. Disaster! Each post now has completely arbitrary tags!
To circumvent this, we replace eager loading with lazy loading. This
generates one extra query for each result row, but it has no chance of
producing such anomalies. This behavior is activated only for
queries containing "sort:random" and derivatives so it shouldn't hit
performance too much.
Rather than flushing the post right away only to find out that there
were validation errors, try to postpone flushing for as long as
possible.
The previous behavior has led to too eager spending of post IDs - each
flush calls nextval(post_id_seq), and postgres sequences are not
affected by transaction rollbacks, so each erroneous post creation
discarded a post ID, which has led to gaps in post IDs.
This changes the checksums to ones that are compatible with 1.x, which
relieves the migration script from recalculating the checksums for all
the posts.
Ignored only the rules about continuing / hanging indentation.
Also, added __init__.py to tests so that pylint discovers them. (I don't
buy pytest's BS about installing your package.)
For quite some time, I hated Falcon's class maps approach that caused
more chaos than good for Szurubooru. I've taken a look at the other
frameworks (hug, flask, etc) again, but they all looked too
bloated/over-engineered. I decided to just talk to WSGI myself.
Regex-based routing may not be the fastest in the world, but I'm fine
with response time of 10 ms for cached /posts.
