- basic architectural understanding required for understanding postgres performance in nontrivial loads
- workload, OS, SSD vs. rotational, etc are playing a big role
- shared memory, process based model
  - basic graph
- shared_buffers is the biggest, sizewise, part of shared memory
- backends are connected to shared memory, including shared_buffers
  - basic graph
- executor wants to read something -> shared buffer hit/miss -> loaded from disk
- executor/utility wants to write something -> shared buffer hit/miss -> loaded from disk
                                                                      -> *not* immediately written back
- no free buffer -> clock sweep
- write stuff back randomly -> corrupted on disk state
- WAL logging
- checkpoints
  - spread checkpoints
  - full page writes
  - checkpoint_segments, min_wal_size, max_wal_size
  - checkpoint_timeout
  - checkpoint_completion_target
- WAL writer
  - tries to flush ahead of other backends
- background writer
  - bgwriter_lru_maxpages, bgwriter_delay, bgwriter_lru_multiplier
  - limited effectiveness, not aggressive enough

-- overview done --
- show flushing problems, including dirty bytes, writeback bytes
  - dirty_writeback/expire_centisecs, dirty_bytes/ratio, dirty_background_bytes/ratio
- how to size shared buffers
  - large if entire working set fits
  - small if very frequent relation drops/reindexes
  - small if relation extension is very frequent
- double buffering

-- improvements
- checkpoint sorting & flushing (9.6)
- relation extension scalability (9.6)
- buffer replacement scalability (9.5)
- new bgwriter algorith (9.6, 9.7?)
- different buffer mapping algorithm (radix tree) (9.7?)
- better cache replacement algorithm
- write combining during writeout
- direct IO (?)