perf: use standard allocator, limit sqlite mmap to 4GB

This is an experimental change to see if we can reduce memory usage
with large SQLite databases.  If successful, we'll do this again and
further reduce the database mmap size.

This will cause greater use of the page cache, but that is more easily
reclaimed by the kernel, and should reduce memory pressure, as well as
making it clearer how much memory the application is actually using
for connections, subscriptions, etc.

Cargo.toml

@@ -54,8 +54,6 @@ bech32 = "0.9.1"
 url = "2.3.1"
 qrcode = { version = "0.12.0", default-features = false, features = ["svg"] }
 nostr = { version = "0.18.0", default-features = false, features = ["base", "nip04", "nip19"] }
-[target.'cfg(not(target_env = "msvc"))'.dependencies]
-tikv-jemallocator = "0.5"
 
 [dev-dependencies]
 anyhow = "1"

src/main.rs

@@ -9,13 +9,6 @@ use std::sync::mpsc::{Receiver as MpscReceiver, Sender as MpscSender};
 use std::thread;
 use tracing::info;
 
-#[cfg(not(target_env = "msvc"))]
-use tikv_jemallocator::Jemalloc;
-
-#[cfg(not(target_env = "msvc"))]
-#[global_allocator]
-static GLOBAL: Jemalloc = Jemalloc;
-
 /// Start running a Nostr relay server.
 fn main() {
     let args = CLIArgs::parse();

src/repo/sqlite_migration.rs

@@ -19,7 +19,7 @@ PRAGMA foreign_keys = ON;
 PRAGMA journal_size_limit = 32768;
 PRAGMA temp_store = 2; -- use memory, not temp files
 PRAGMA main.cache_size = 20000; -- 80MB max cache size per conn
-pragma mmap_size = 17179869184; -- cap mmap at 16GB
+pragma mmap_size = 4294967296; -- cap mmap at 4GB
 "##;
 
 /// Latest database version