# Migrate the server from SQLite to PostgreSQL - DONE Done on 2026-08-30. The server runs on Postgres 18: sqlx switched to the `postgres` feature, 95 placeholders renumbered to `$N`, `REAL` columns widened to `DOUBLE PRECISION` (Postgres REAL is float4 and will not decode into f64), the `flagged` and `is_bot` flags made real BOOLEANs, the leaderboard's derived table given the alias Postgres requires, and `INSERT OR IGNORE` rewritten as `ON CONFLICT DO NOTHING`. Integration tests run against a real database, each in its own throwaway schema, since Postgres has no in-memory mode. ## Still open - **Timestamps are still TEXT.** They are RFC3339 and sort correctly as text, so this is not a correctness problem, but `TIMESTAMPTZ` would let the database do date arithmetic instead of the application. - **LISTEN/NOTIFY is unused.** Multiplayer rooms are still in-process, so the server cannot yet run as more than one instance. --- ## Original note ## Why The server is a multi-user network service. It handles auth, sessions, friendships, leaderboards, live multiplayer races, and rate limiting. SQLite allows only one writer at a time. Races, leaderboard writes, and rate-limit counters all write at the same time, so the single writer is the wrong shape for this workload. PostgreSQL gives three things the server needs: - MVCC. Concurrent writers do not block each other. - Real timestamp types. `created_at` and `expires_at` are `TEXT` today. - `LISTEN`/`NOTIFY`. This maps onto multiplayer room pub/sub, and lets the server run as more than one instance. Keep `mitmux` on SQLite. It is a single-user local tool, and its pure-Go SQLite driver is what keeps cross-compilation free of CGO. ## Scope Server only (`crates/server`). The TUI, WASM, and Steam builds do not touch the database. ## What makes this cheap - The migrations use no SQLite-only syntax. There is no `AUTOINCREMENT`, no `PRAGMA`, no `strftime`, and no `WITHOUT ROWID`. - Primary keys are `TEXT` UUIDs, not integer rowids. - All 33 call sites use the runtime `sqlx::query` function, not the `sqlx::query!` macro. There is no compile-time schema metadata, so no `cargo sqlx prepare` step and no offline query cache to regenerate. ## Steps 1. Change the sqlx feature in `crates/server/Cargo.toml` from `sqlite` to `postgres`. Keep `runtime-tokio-rustls`, `time`, and `uuid`. 2. Change the pool type from `SqlitePool` to `PgPool` in `state.rs`, and change the connection string handling in `main.rs`. 3. Convert every placeholder from `?` to `$1`, `$2`, and so on. Postgres numbers its placeholders. This is the largest mechanical change: 33 queries across `auth.rs`, `stats.rs`, `friends.rs`, `multiplayer.rs`, `cosmetics.rs`, `anticheat.rs`, `rate_limit.rs`, and `spotify.rs`. 4. Change `created_at TEXT` and `expires_at TEXT` to `TIMESTAMPTZ` in the migrations. Update the Rust structs to `time::OffsetDateTime`. 5. Replace any `INSERT OR REPLACE` or `INSERT OR IGNORE` with `INSERT ... ON CONFLICT`. Check `cosmetics.rs` and `rate_limit.rs` first. 6. Add a `docker-compose.yml` with a `postgres` service, so a contributor can start a database with one command. 7. Run the full test suite against the new backend. ## Verify - `cargo test --workspace` passes. - Two clients can finish a multiplayer race at the same time, and both results are written. - A session expires at the correct time. This confirms the `TIMESTAMPTZ` conversion. - Rate limiting still rejects a burst from one user. ## Follow-up Move multiplayer room pub/sub from in-process state to `LISTEN`/`NOTIFY`. Only after that can the server run more than one instance.