config: move postgres tuning into compose, out of ALTER SYSTEM

[moogoo78]

Summary

Postgres tuning on production lived in two files inside the pgdata Docker volume, neither in version control:

setting	value	source
shared_buffers	128MB	pgdata/postgresql.conf — initdb default, nobody set it
statement_timeout	30s	pgdata/postgresql.auto.conf — ALTER SYSTEM
log_autovacuum_min_duration	0	pgdata/postgresql.auto.conf — ALTER SYSTEM

A fresh environment built from this repo would not have reproduced production. Command-line -c outranks both postgresql.conf and postgresql.auto.conf, so declaring them in compose makes the repo the single source of truth.

Why the memory settings are env-driven

compose.prod.yml is shared by production and staging, which are not the same size. A hardcoded shared_buffers=2GB could stop Postgres starting on a smaller box.

Defaults are the stock PostgreSQL values, so any environment that sets nothing behaves exactly as it does today — no surprise for staging. Production sets the tuned values in its .env:

POSTGRES_SHARED_BUFFERS=2GB
POSTGRES_EFFECTIVE_CACHE_SIZE=6GB
POSTGRES_MAINTENANCE_WORK_MEM=256MB

statement_timeout and log_autovacuum_min_duration are not size-dependent, so they're pinned for every deployed environment.

Why 2GB

Production is 8 GiB with a 3.4 GB database (named_area alone is 1.8 GB of PostGIS geometry). At 128MB, Postgres was caching almost nothing itself and leaning entirely on the OS page cache. Rule of thumb applied: shared_buffers ~25% of RAM, effective_cache_size ~75%.

Verification

Rendered config substitutes correctly both ways:

• with env vars → shared_buffers=2GB, effective_cache_size=6GB, maintenance_work_mem=256MB
• without → 128MB / 4GB / 64MB (stock defaults)

Booted locally with the production values; all five report the highest-precedence source:

 effective_cache_size        | 786432 | 8kB | command line
 log_autovacuum_min_duration | 0      | ms  | command line
 maintenance_work_mem        | 262144 | kB  | command line
 shared_buffers              | 262144 | 8kB | command line
 statement_timeout           | 30000  | ms  | command line

Follow-up after merge

1. Add the three vars to production's .env (per-host config, like the existing secrets)
2. Restart Postgres so they take effect — brief downtime
3. Remove the two ALTER SYSTEM entries from postgresql.auto.conf so there is one source of truth rather than two silently competing

Also left alone deliberately: the inert docker/postgresql.conf mount. Uncommenting its config_file line would replace the image's entire default config and drop listen_addresses, breaking networking. Its pg_stat_statements settings could be moved onto this command line as a separate change.

🤖 Generated with Claude Code