diff options
| author | Pjotr <wrk@bosa.lan> | 2026-11-09 00:48:27 +0000 |
|---|---|---|
| committer | Pjotr <wrk@bosa.lan> | 2026-11-09 00:48:27 +0000 |
| commit | 731e085c42e22104f4cba9975a33f5d22b058caf (patch) | |
| tree | 34419f60994c54eebba4c246419fe7dac1b7d885 /src/error.rs | |
| parent | ab5fef05237309667dd2ca7f835a5fbf9c71a12a (diff) | |
| download | genenetwork4-731e085c42e22104f4cba9975a33f5d22b058caf.tar.gz | |
errors: log every 5xx, and remember the last database failure
A user reported that the token was fine and the service was up, but "everything backed by the relational database is failing" -- while `/search` (Xapian) and `/version` answered. By the time I looked, every data endpoint answered 200 and `/version` said `database: reachable`, and there was **no record anywhere** of what they had hit: the service logged nothing at all, so a caller's report could not be matched to a cause, and `/version`'s `database` is a *now* probe that says nothing about the minute before. Two things now make that diagnosable: * every 5xx is written to stderr as one JSON line -- status, code, operation_id, **request_id**, took_ms and a truncated detail -- so "500 at 12:31, request id ..." can be looked up. 4xx stay unlogged: they are the caller's business, and a 401/422 stream would drown this; * the last database failure is remembered (`db::record_failure`, called from `ApiError::from_db`, the one place every database error passes through) and `/version` reports it in `notes` for ten minutes, with its kind (`unavailable`, `connection_lost`, `query`) and the driver's message. That is the gap that misled the user: a healthy probe beside failing endpoints looks like an auth problem from outside. `is_connection_loss` gained a `_detail` form so the caller that already has the rendered message does not render it twice. * src/error.rs: the log line, and recording failures by kind; * src/db.rs: `FailureKind`, `Failure`, `record_failure`, `recent_failure`, and two unit tests (a failure is remembered and ages out; a driver message is classified correctly); * src/lib.rs: the `/version` note, absent unless something failed recently. Checked against a deliberately broken `SQL_URI` on a scratch instance: the data endpoint answers `503 db_unavailable`, the log carries the request id, and `/version` shows the note while still answering 200.
Diffstat (limited to 'src/error.rs')
| -rw-r--r-- | src/error.rs | 59 |
1 files changed, 50 insertions, 9 deletions
diff --git a/src/error.rs b/src/error.rs index a9b01bd..b84cf9b 100644 --- a/src/error.rs +++ b/src/error.rs @@ -121,27 +121,40 @@ impl ApiError { /// the inline `map_err` sites agree. pub fn from_db(error: crate::db::DbError, operation_id: &str) -> ApiError { match error { - crate::db::DbError::Unavailable(detail) => ApiError::new( - StatusCode::SERVICE_UNAVAILABLE, - "db_unavailable", - "Database unavailable", - detail, - operation_id, - ), + crate::db::DbError::Unavailable(detail) => { + crate::db::record_failure(crate::db::FailureKind::Unavailable, detail.clone()); + ApiError::new( + StatusCode::SERVICE_UNAVAILABLE, + "db_unavailable", + "Database unavailable", + detail, + operation_id, + ) + } crate::db::DbError::Query(error) => { // A connection that died under the query is not the // query's fault. Forget the pool so the next request // cannot draw the same dead sockets -- this is the one // place every database error passes through, so the // recovery does not depend on each caller remembering it. + let detail = error.to_string(); + let lost = crate::db::is_connection_loss_detail(&detail); + crate::db::record_failure( + if lost { + crate::db::FailureKind::ConnectionLost + } else { + crate::db::FailureKind::Query + }, + detail.clone(), + ); let api_error = ApiError::new( StatusCode::INTERNAL_SERVER_ERROR, "database_error", "Database error", - error.to_string(), + detail, operation_id, ); - if crate::db::is_connection_loss(&crate::db::DbError::Query(error)) { + if lost { crate::db::forget_pool(); } api_error @@ -153,10 +166,38 @@ impl ApiError { impl IntoResponse for ApiError { fn into_response(self) -> Response { let status = StatusCode::from_u16(self.status).unwrap_or(StatusCode::INTERNAL_SERVER_ERROR); + // Server-side failures get one line on stderr each, as JSON, so a + // caller's report ("500 at 12:31, request id ...") can be matched to + // what actually happened. 4xx are not logged: they are the caller's + // business (and a busy 401/422 stream would drown this out), but a + // 5xx is ours, and until this existed the service kept no record of + // one at all. + if status.is_server_error() { + eprintln!( + "{}", + serde_json::json!({ + "event": "error", + "status": status.as_u16(), + "code": self.code, + "operation_id": self.operation_id, + "request_id": self.request_id, + "took_ms": self.took_ms, + "detail": truncate(&self.detail, 500), + }) + ); + } (status, Json(self)).into_response() } } +/// `text` cut to `limit` characters (not bytes), for one log line. +fn truncate(text: &str, limit: usize) -> String { + if text.chars().count() <= limit { + return text.to_string(); + } + text.chars().take(limit).collect::<String>() + "…" +} + /// The 401 every token-gated endpoint can return, as one shared /// `IntoResponses` type (an `ApiError` body under the hood) instead of /// writing out `(status = 401, description = "...", body = ApiError)` |
