benchmarked on a synthetic db (10k users, 500k follows, 500k stars, 50k repos), the per-source queries behind the following timeline were slow because of missing indexes and a poor query shape: - stars filtered by did had no usable index; sqlite picked idx_stars_subject_type which matches nearly every row (~312ms) - follower/following counts group by subject_did, which had no index, forcing a full scan (~229ms) - follows ordered by followed_at with no index sorted every matched row in a temp b-tree (~157ms) fixes: - new indexes: stars(did, subject_type, created), follows(subject_did), follows(followed_at), repos(created); the latter two also let the global timeline read newest-first straight off an index instead of sorting - push the following-set membership check into sql with a subquery (orm.FilterInSubquery) instead of materializing all followed dids in go and binding them as hundreds of placeholders in three separate queries all benchmarked queries now run in ~1ms. Signed-off-by: Anirudh Oppiliappan <anirudh@tangled.org>