hyperhive/hive-c0re/src/stores/db.rs

376 lines
14 KiB
Rust

//! Shared sqlite connection setup and versioned schema migrations for
//! hive-c0re's host-side stores.
//!
//! All stores share one sqlite file (`db/broker.sqlite`) but each owns
//! its own `Mutex<Connection>`. [`open`] handles the connection open
//! dance; [`apply_versioned_migrations`] handles schema evolution.
//!
//! ## Migration strategy
//!
//! Each store calls [`apply_versioned_migrations`] from its `open`. It
//! tracks the applied count in a `schema_versions` table (one row per
//! `subsystem` key). Only migrations past the stored version run.
//!
//! Legacy databases (written before versioning) need no special marker:
//! each column-adding migration carries the `(table, column)` it
//! introduces, so re-walking a pre-versioning DB from version 0 simply
//! skips the migrations whose column already exists — converging a fully-
//! or partially-migrated legacy DB without ever re-running an `ADD COLUMN`
//! against an existing column. See each store's `MIGRATIONS` constant.
use std::path::Path;
use std::time::Duration;
use anyhow::{Context, Result};
use rusqlite::{Connection, OptionalExtension, params};
/// How long a write waits on another connection's lock before erroring.
/// Generous relative to the stores' tiny transactions — a timeout here
/// means something is genuinely wedged, not ordinary contention.
const BUSY_TIMEOUT: Duration = Duration::from_secs(5);
const SCHEMA_VERSIONS_DDL: &str = "
CREATE TABLE IF NOT EXISTS schema_versions (
store TEXT PRIMARY KEY,
version INTEGER NOT NULL DEFAULT 0
)";
/// One numbered schema migration.
///
/// `sql` runs once, at the index whose value is the current stored version.
/// `adds_column`, when set, is the `(table, column)` this migration
/// introduces: if that column already exists — a pre-versioning DB migrated
/// by the old try-and-ignore path, or a partially-migrated one — `sql` is
/// skipped and only the version counter advances. This makes column-adding
/// migrations idempotent *without* `SQLite`'s unsupported `ALTER TABLE … ADD
/// COLUMN IF NOT EXISTS`, so re-walking a legacy DB from version 0 never
/// re-runs an `ADD COLUMN` against a column that is already there (which
/// would fail with "duplicate column name" and wedge startup permanently).
///
/// Multi-statement migrations wrap their `ALTER` + backfill / index rebuild
/// in an explicit `BEGIN; … COMMIT;` batch, so "the added column exists" is a
/// faithful proxy for "this whole migration committed".
pub struct Migration {
/// SQL for this version — a single statement or a `BEGIN; … COMMIT;` batch.
pub sql: &'static str,
/// `(table, column)` this migration adds, if any. Present → the migration
/// is skipped when the column already exists (idempotent replay).
pub adds_column: Option<(&'static str, &'static str)>,
}
/// Apply numbered schema migrations for `subsystem`, tracking progress in
/// the shared `schema_versions` table.
///
/// - `subsystem` — unique key for this store in `schema_versions`
/// (e.g. `"approvals"`, `"broker"`).
/// - `migrations` — ordered list of [`Migration`]s; migration `i` runs when
/// the stored version is `i`. A pre-versioning DB (no `schema_versions`
/// row) starts at version 0 and re-walks every migration, but each
/// column-adding migration is skipped when its `adds_column` is already
/// present — so a fully- or partially-migrated legacy DB converges without
/// re-running an `ADD COLUMN` against an existing column.
///
/// # Errors
///
/// Returns an error if any migration statement fails or the version update
/// fails. Each migration + its version bump is a single `execute_batch`
/// (multi-statement migrations wrap themselves in `BEGIN; … COMMIT;`), so a
/// failure leaves the DB at the last successfully committed version.
pub fn apply_versioned_migrations(
conn: &Connection,
subsystem: &str,
migrations: &[Migration],
) -> Result<()> {
// Ensure the version-tracking table exists (idempotent).
conn.execute_batch(SCHEMA_VERSIONS_DDL)
.context("create schema_versions table")?;
// Current version for this subsystem — 0 (and a fresh row) when absent.
let stored: Option<i64> = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = ?1",
[subsystem],
|row| row.get(0),
)
.optional()
.context("read schema_versions")?;
let version = if let Some(v) = stored {
usize::try_from(v.max(0)).unwrap_or(0)
} else {
conn.execute(
"INSERT INTO schema_versions (store, version) VALUES (?1, 0)",
[subsystem],
)
.context("insert schema_versions row")?;
0
};
if version >= migrations.len() {
return Ok(());
}
for (i, mig) in migrations.iter().enumerate().skip(version) {
// Skip a column-adding migration whose column is already present:
// a legacy DB migrated by the old path, or a resumed partial one.
let already_applied = match mig.adds_column {
Some((table, column)) => column_exists(conn, table, column)?,
None => false,
};
if !already_applied {
conn.execute_batch(mig.sql)
.with_context(|| format!("{subsystem} migration v{} failed: {}", i + 1, mig.sql))?;
}
let next = i64::try_from(i + 1).unwrap_or(i64::MAX);
conn.execute(
"UPDATE schema_versions SET version = ?1 WHERE store = ?2",
params![next, subsystem],
)
.with_context(|| format!("{subsystem} update schema_versions to v{}", i + 1))?;
}
Ok(())
}
/// Whether `column` exists on `table` in this connection's schema. Both
/// arguments are bound parameters — `pragma_table_info(?1)` accepts the table
/// name as a bound value — so this is not a SQL-injection surface even if a
/// caller ever passes a non-static name.
fn column_exists(conn: &Connection, table: &str, column: &str) -> Result<bool> {
conn.prepare("SELECT 1 FROM pragma_table_info(?1) WHERE name = ?2")
.context("prepare column probe")?
.exists(params![table, column])
.context("check column existence")
}
/// Open a connection to the sqlite file at `path`, creating the parent
/// directory if needed. `subsystem` labels error contexts (`"broker"`,
/// `"approvals"`, …).
pub fn open(path: &Path, subsystem: &str) -> Result<Connection> {
if let Some(parent) = path.parent() {
std::fs::create_dir_all(parent)
.with_context(|| format!("create {subsystem} db parent {}", parent.display()))?;
}
let conn = Connection::open(path)
.with_context(|| format!("open {subsystem} db {}", path.display()))?;
conn.busy_timeout(BUSY_TIMEOUT)
.with_context(|| format!("set {subsystem} busy_timeout"))?;
Ok(conn)
}
#[cfg(test)]
mod tests {
use super::*;
use std::sync::atomic::{AtomicU64, Ordering};
static TEST_COUNTER: AtomicU64 = AtomicU64::new(0);
fn open_tmp() -> Connection {
let n = TEST_COUNTER.fetch_add(1, Ordering::Relaxed);
let path = format!("/tmp/hive-db-test-{}-{}.sqlite", std::process::id(), n);
Connection::open(&path).expect("open tmp db")
}
const TEST_SCHEMA: &str = "CREATE TABLE IF NOT EXISTS things (id INTEGER PRIMARY KEY)";
const TEST_MIGRATIONS: &[Migration] = &[
Migration {
sql: "ALTER TABLE things ADD COLUMN alpha TEXT",
adds_column: Some(("things", "alpha")),
},
Migration {
sql: "ALTER TABLE things ADD COLUMN beta INTEGER NOT NULL DEFAULT 0",
adds_column: Some(("things", "beta")),
},
Migration {
sql: "ALTER TABLE things ADD COLUMN gamma TEXT",
adds_column: Some(("things", "gamma")),
},
];
#[test]
fn fresh_install_runs_all_migrations() {
let conn = open_tmp();
conn.execute_batch(TEST_SCHEMA).unwrap();
apply_versioned_migrations(&conn, "things", TEST_MIGRATIONS).unwrap();
// All three columns should exist.
for col in ["alpha", "beta", "gamma"] {
let exists: bool = conn
.prepare(&format!(
"SELECT 1 FROM pragma_table_info('things') WHERE name = '{col}'"
))
.unwrap()
.exists([])
.unwrap();
assert!(exists, "column {col} missing after fresh migration");
}
// Version should be 3.
let v: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'things'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(v, 3);
}
#[test]
fn legacy_install_skips_all_migrations() {
let conn = open_tmp();
// Simulate a fully-migrated legacy DB: all columns exist,
// but schema_versions does not yet.
conn.execute_batch(
"CREATE TABLE IF NOT EXISTS things (id INTEGER PRIMARY KEY,
alpha TEXT, beta INTEGER NOT NULL DEFAULT 0, gamma TEXT)",
)
.unwrap();
apply_versioned_migrations(&conn, "things", TEST_MIGRATIONS).unwrap();
// Version should be set to migrations.len() (3), not 0.
let v: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'things'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(v, 3, "legacy install should skip to latest version");
}
#[test]
fn partial_migration_resumes_from_version() {
let conn = open_tmp();
// Simulate a DB that has only the first migration applied.
conn.execute_batch(
"CREATE TABLE IF NOT EXISTS things (id INTEGER PRIMARY KEY, alpha TEXT);
CREATE TABLE IF NOT EXISTS schema_versions (store TEXT PRIMARY KEY, version INTEGER NOT NULL DEFAULT 0);
INSERT INTO schema_versions (store, version) VALUES ('things', 1)",
)
.unwrap();
apply_versioned_migrations(&conn, "things", TEST_MIGRATIONS).unwrap();
// Only beta and gamma should have been added (alpha already existed).
for col in ["alpha", "beta", "gamma"] {
let exists: bool = conn
.prepare(&format!(
"SELECT 1 FROM pragma_table_info('things') WHERE name = '{col}'"
))
.unwrap()
.exists([])
.unwrap();
assert!(
exists,
"column {col} missing after partial migration resume"
);
}
let v: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'things'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(v, 3);
}
#[test]
fn already_at_latest_is_noop() {
let conn = open_tmp();
conn.execute_batch(TEST_SCHEMA).unwrap();
apply_versioned_migrations(&conn, "things", TEST_MIGRATIONS).unwrap();
// Second call is a no-op.
apply_versioned_migrations(&conn, "things", TEST_MIGRATIONS).unwrap();
let v: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'things'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(v, 3);
}
#[test]
fn multiple_stores_in_same_db() {
let conn = open_tmp();
conn.execute_batch(
"CREATE TABLE IF NOT EXISTS a (id INTEGER PRIMARY KEY);
CREATE TABLE IF NOT EXISTS b (id INTEGER PRIMARY KEY)",
)
.unwrap();
apply_versioned_migrations(
&conn,
"store_a",
&[Migration {
sql: "ALTER TABLE a ADD COLUMN col1 TEXT",
adds_column: Some(("a", "col1")),
}],
)
.unwrap();
apply_versioned_migrations(
&conn,
"store_b",
&[Migration {
sql: "ALTER TABLE b ADD COLUMN colx INTEGER NOT NULL DEFAULT 0",
adds_column: Some(("b", "colx")),
}],
)
.unwrap();
// Each store tracks independently.
let va: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'store_a'",
[],
|r| r.get(0),
)
.unwrap();
let vb: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'store_b'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(va, 1);
assert_eq!(vb, 1);
}
/// A partially-migrated legacy database: some columns from the old
/// try-and-ignore path exist, with no `schema_versions` row yet. The
/// per-migration `adds_column` guard must skip the already-present
/// columns and add only the missing ones — no "duplicate column name".
#[test]
fn partial_legacy_completes_without_error() {
let conn = open_tmp();
// Simulate a DB that has `alpha` (v1) but not `beta` (v2) or
// `gamma` (v3). No schema_versions row exists → start from v0 and
// guard-skip alpha while adding beta + gamma.
conn.execute_batch(
"CREATE TABLE IF NOT EXISTS things (id INTEGER PRIMARY KEY, alpha TEXT)",
)
.unwrap();
apply_versioned_migrations(&conn, "things", TEST_MIGRATIONS).unwrap();
// All three columns must now exist.
for col in ["alpha", "beta", "gamma"] {
let exists: bool = conn
.prepare(&format!(
"SELECT 1 FROM pragma_table_info('things') WHERE name = '{col}'"
))
.unwrap()
.exists([])
.unwrap();
assert!(
exists,
"column {col} missing after partial-legacy migration"
);
}
// Version must be at the latest.
let v: i64 = conn
.query_row(
"SELECT version FROM schema_versions WHERE store = 'things'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(
v, 3,
"must reach latest version after completing partial legacy"
);
}
}