Updated the Probe method in the Persistence package with a query deadline.

This commit is contained in:
2026-08-13 00:11:59 +02:00
parent 8303598244
commit d7465d03cc
4 changed files with 138 additions and 15 deletions
+3 -3
View File
@@ -7,7 +7,7 @@ The [Fluent](https://github.com/hummingbird-project/hummingbird-fluent)-based da
| Backend selection | `Driver` (`postgres` or `inMemory`), `Configuration` (the PostgreSQL connection parameters), `TLS` (the connection's TLS posture) |
| Service | `Service`, which builds the `Fluent` service configured for the chosen driver |
| Migrations | `PrepareDB`, the single registrar declaring every migration, in order |
| Readiness | `Probe`, which reports whether the default database answers a `SELECT 1` |
| Readiness | `Probe`, which reports whether the default database answers a `SELECT 1` within a deadline |
| Scaffolding (internal) | `ExampleRecord`, `CreateExampleRecord`, and `ExampleRepository` — the model → migration → repository pattern, to be replaced by the first real domain model |
## Design rules
@@ -15,7 +15,7 @@ The [Fluent](https://github.com/hummingbird-project/hummingbird-fluent)-based da
- **One default database.** `Service` registers the selected backend as the *default* database, so repositories resolve it with a plain `fluent.db()` and stay agnostic of which driver is in use.
- **Migrations are declared once, and append-only.** `PrepareDB` registers every migration in the order it must run; alter the schema by adding a migration, never by editing one that has already run. Registering does not apply them: the in-memory backend migrates on startup, while a shared PostgreSQL database is migrated out of band (the executable's migrate-and-exit mode), so booting instances never race.
- **Models never cross a concurrency boundary.** FluentKit models are mutable reference types, so they stay internal to the package and repositories return `Sendable` value-type snapshots (e.g. `Example`) instead.
- **Readiness never throws.** `Probe` runs a `SELECT 1` — the cheapest statement both backends understand, independent of any schema — and maps every failure to `false`, so callers translate it straight into a readiness response.
- **Readiness never throws, and never hangs.** `Probe` runs a `SELECT 1` — the cheapest statement both backends understand, independent of any schema — and maps every failure to `false`, so callers translate it straight into a readiness response. The query is raced against a deadline (2 seconds unless the probe is built with another), so a database that hangs rather than refuses yields a prompt "not ready" instead of a hanging readiness endpoint.
- **A single connection for the in-memory store.** The SQLite backend is capped at one connection per event loop, so every query reaches the same in-memory database instead of each pooled connection getting a private one.
- **Method structs.** `Service`, `PrepareDB`, and `Probe` hold their lifetime-fixed configuration in `init` and take only per-call inputs in `callAsFunction`.
@@ -36,7 +36,7 @@ Sources/
Tests/
├── Cases/ the test suites, mirroring the Sources/ layout
└── Utils/ the NotSQL* fakes backing the probe's non-SQL-database case, the
plaintext-only fake PostgreSQL server, and the suite Tag constants
plaintext-only and silent fake PostgreSQL servers, and the suite Tag constants
```
## Testing
@@ -12,14 +12,21 @@ public struct Probe: Sendable {
/// The `Fluent` service whose default database is probed.
private let fluent: Fluent
/// The longest the probe waits for the database's answer before reporting it as not reachable.
private let timeout: Duration
// MARK: Initializers
/// Creates a probe for the default database of the given `Fluent` service.
/// - Parameter fluent: the `Fluent` service whose default database is probed.
/// - Parameters:
/// - fluent: the `Fluent` service whose default database is probed.
/// - timeout: the longest the probe waits for the database's answer before reporting it as not reachable.
public init(
fluent: Fluent
fluent: Fluent,
timeout: Duration = .seconds(2)
) {
self.fluent = fluent
self.timeout = timeout
}
// MARK: Methods
@@ -28,21 +35,43 @@ public struct Probe: Sendable {
///
/// Runs a trivial `SELECT 1` against the default database the cheapest statement both the PostgreSQL and SQLite backends understand so
/// a readiness check does not depend on any particular schema or model. Any failure (connection refused, authentication error, pool exhausted) is
/// reported as not reachable rather than thrown, so callers can map it straight onto a readiness response. A default database that is not an SQL
/// database is likewise reported as not reachable.
/// - Returns: `true` when the database answers the probe, `false` otherwise.
/// reported as not reachable rather than thrown, and an answer that does not arrive within the timeout is likewise reported as not reachable so
/// a database that hangs yields a prompt "not ready" instead of a hanging readiness endpoint. A default database that is not an SQL database is
/// also reported as not reachable.
/// - Returns: `true` when the database answers the probe in time, `false` otherwise.
public func callAsFunction() async -> Bool {
guard let database = fluent.db() as? any SQLDatabase else {
return false
}
do {
try await database.raw("SELECT 1").run()
return true
} catch {
return false
// The query is raced against the deadline from unstructured tasks whose first answer wins: a structured group
// would await the query child before returning, so a database that hangs mid-query the very failure the
// deadline exists for would hang the probe with it. The loser is cancelled and abandoned; a late answer lands
// in a finished stream and is dropped.
let (answers, continuation) = AsyncStream.makeStream(of: Bool.self)
let query = Task {
do {
try await database
.raw("SELECT 1")
.run()
continuation.yield(true)
} catch {
continuation.yield(false)
}
}
let deadline = Task {
try? await Task.sleep(for: timeout)
continuation.yield(false)
}
var answer = answers.makeAsyncIterator()
let isReachable = await answer.next() ?? false
query.cancel()
deadline.cancel()
return isReachable
}
}
@@ -57,6 +57,46 @@ struct ProbeTests {
#expect(!isReachable)
}
@Test
func `reports a hanging database as unreachable within its timeout`() async throws {
// The silent server accepts the TCP connection and never answers, so the probe's query can only
// ever be resolved by its deadline without one, it would wait out the driver's own connect
// timeout (10 seconds) instead.
let server = try await SilentPostgresServer.start()
let service = try Service(
driver: .postgres(
.init(
host: "127.0.0.1",
port: server.port,
name: "hanging",
username: "nobody",
password: "nothing",
tls: .off,
maxConnectionsPerEventLoop: 1
)
),
logger: Logger(label: "test")
)
let fluent = service()
let probe = Probe(
fluent: fluent,
timeout: .milliseconds(100)
)
let clock = ContinuousClock()
let start = clock.now
let isReachable = await probe()
let elapsed = clock.now - start
try await fluent.shutdown()
try await server.stop()
#expect(!isReachable)
// Well past the 100-millisecond deadline to absorb scheduling noise, yet far below the driver's
// 10-second connect timeout only the deadline can answer this fast.
#expect(elapsed < .seconds(5))
}
@Test
func `reports a default database that is not an SQL database`() async throws {
let fluent = Fluent(logger: Logger(label: "test"))
@@ -0,0 +1,54 @@
import NIOCore
import NIOPosix
/// A fake server accepting connections and never answering.
///
/// A client connecting to it completes the TCP handshake and then waits forever for the first protocol byte the shape of a database that hangs rather
/// than refuses. This is what the probe's deadline test connects to, proving the probe answers within its timeout instead of hanging alongside the server.
final class SilentPostgresServer {
// MARK: Properties
/// The port the server listens on, assigned by the system at bind time.
let port: Int
/// The listening channel the server accepts connections through.
private let channel: Channel
// MARK: Initializers
private init(
channel: Channel,
port: Int
) {
self.channel = channel
self.port = port
}
// MARK: Functions
/// Starts a server on the loopback interface, on a system-assigned port.
/// - Returns: the running server, ready to be connected to at ``port``.
static func start() async throws -> SilentPostgresServer {
let channel = try await ServerBootstrap(group: MultiThreadedEventLoopGroup.singleton)
.bind(host: "127.0.0.1", port: 0)
.get()
guard let port = channel.localAddress?.port else {
throw ChannelError.unknownLocalAddress
}
return .init(
channel: channel,
port: port
)
}
/// Stops the server, closing its listening channel.
func stop() async throws {
try await channel
.close()
.get()
}
}