palloc quantises sizes to powers of two, keeps eleven freelists, carves dying blocks into reusable chunks, and recycles memory contexts so a per-tuple context costs zero mallocs.
Committing doesn’t fsync: backends copy WAL into shared memory and piggyback on one fsync, using LWLockAcquireOrWait so followers can observe someone else’s flush without ever …
Every byte tmux sends to a terminal goes through one function that appends to a per-client buffer, and that buffer is flushed with a single write() per event loop iteration.
The page checksum runs 32 FNV-1a hashes in parallel over a 32-column view of the page to hide multiply latency — and 32 is baked into the on-disk format.
B-tree leaves store one posting list tuple per distinct key instead of one entry per row, and an over-long list is split in place inside the insert’s critical section.