fix(transfer): rebuild dropped pools between transfer statements
This commit is contained in:
3 files changed
+221
-13
No files matched your search
@@ -27,7 +27,8 @@ use crate::models::connection::{ConnectionConfig, DatabaseType};
|
||||
use crate::object_source_sql::{build_executable_object_source_statements, EditableObjectSourceSqlInput};
|
||||
use crate::query::{
|
||||
agent_execute_query_params, is_dbx_query_timeout_error, pool_error_action, query_timeout_duration,
|
||||
wait_for_query_opt, PoolErrorAction, QueryExecutionOptions, StreamProgressClock, AGENT_PROTOCOL_MAX_ROWS,
|
||||
should_discard_pool_after_query_timeout, wait_for_query_opt, PoolErrorAction, QueryExecutionOptions,
|
||||
StreamProgressClock, AGENT_PROTOCOL_MAX_ROWS,
|
||||
};
|
||||
use crate::sql::{split_sql_statements, split_sql_statements_for_database};
|
||||
use crate::sql_dialect::{
|
||||
@@ -1920,9 +1921,9 @@ async fn execute_transfer_write_statement(
|
||||
crate::query::check_read_only_for_connection(state, target_pool_key, sql).await?;
|
||||
let (_, _, _, query_timeout_secs) = transfer_pool_context(state, target_pool_key).await;
|
||||
let query_timeout = query_timeout_duration(query_timeout_secs);
|
||||
let pool_handle = state.pool_handle(target_pool_key).await;
|
||||
let client = match pool_handle.as_ref() {
|
||||
Some(PoolKind::SqlServer(client)) => client.clone(),
|
||||
let pool = ensure_transfer_statement_pool(state, target_pool_key).await?;
|
||||
let client = match &pool {
|
||||
PoolKind::SqlServer(client) => client.clone(),
|
||||
_ => return Err("SQL Server connection not found".to_string()),
|
||||
};
|
||||
let mut client = client.lock().await;
|
||||
@@ -6378,6 +6379,52 @@ fn should_prepare_fresh_agent_transfer_session(db_type: Option<DatabaseType>, er
|
||||
)
|
||||
}
|
||||
|
||||
/// Whether a discarded transfer pool has to be replaced before the next statement runs.
|
||||
///
|
||||
/// Every error that reaches the `Discard` arm tears the pool down: the failed statement
|
||||
/// is never replayed (its outcome on the server is unknown), and a driver whose query
|
||||
/// timed out may still own a checked-out connection. The pool key, though, is shared by
|
||||
/// every table of the transfer, so leaving it missing turns one transient failure into a
|
||||
/// *misleading* "Connection not found" on the next table instead of that table running
|
||||
/// (or failing) on its own. Prepare a replacement for native drivers; agent/JDBC
|
||||
/// sessions keep their stricter rule, where recovery stays gated on the structured
|
||||
/// quarantine decision.
|
||||
fn should_reconnect_discarded_transfer_pool(db_type: Option<DatabaseType>) -> bool {
|
||||
db_type.is_some_and(|db_type| !crate::database_capabilities::is_agent_type(&db_type))
|
||||
}
|
||||
|
||||
/// Resolve the pool a transfer statement runs on, re-creating it when it is gone.
|
||||
///
|
||||
/// A bulk transfer lives for tens of minutes and shares one pool key across every table,
|
||||
/// so a pool that disappears between two statements used to abort the run with a
|
||||
/// misleading `Connection not found` that says nothing about the database being
|
||||
/// unreachable. The pool can legitimately be gone: `execute_on_pool_once` drops the pool
|
||||
/// of a driver whose query timed out, the connection keepalive tears down a pool whose
|
||||
/// ping failed or timed out, and every other transfer statement reuses the same key.
|
||||
/// Rebuilding it here is safe -- the statement has not run yet, so nothing is replayed --
|
||||
/// and keeps the transfer working on a fresh connection instead of failing the table.
|
||||
async fn ensure_transfer_statement_pool(state: &AppState, pool_key: &str) -> Result<PoolKind, String> {
|
||||
if let Some(pool) = state.pool_handle(pool_key).await {
|
||||
return Ok(pool);
|
||||
}
|
||||
let (connection_id, database, _, _) = transfer_pool_context(state, pool_key).await;
|
||||
let Some(connection_id) = connection_id else {
|
||||
return Err("Connection not found".to_string());
|
||||
};
|
||||
let catalog = catalog_from_pool_key(pool_key).map(str::to_string);
|
||||
let client_session_id = client_session_id_from_pool_key(pool_key).map(str::to_string);
|
||||
state
|
||||
.get_or_create_pool_for_session_with_catalog(
|
||||
&connection_id,
|
||||
database.as_deref(),
|
||||
catalog.as_deref(),
|
||||
client_session_id.as_deref(),
|
||||
)
|
||||
.await
|
||||
.map_err(|error| format!("Connection not found: {error}"))?;
|
||||
state.pool_handle(pool_key).await.ok_or_else(|| "Connection not found".to_string())
|
||||
}
|
||||
|
||||
async fn transfer_pool_context(
|
||||
state: &AppState,
|
||||
pool_key: &str,
|
||||
@@ -6437,7 +6484,8 @@ async fn execute_on_pool_with_options(
|
||||
// the failed statement above, only prepares the pool for
|
||||
// whichever statement runs next.
|
||||
let prepare_agent_session = should_prepare_fresh_agent_transfer_session(db_type, error);
|
||||
if pool_error_action(db_type, error) == PoolErrorAction::ReconnectAndRetry || prepare_agent_session {
|
||||
let reconnect_pool = should_reconnect_discarded_transfer_pool(db_type);
|
||||
if reconnect_pool || prepare_agent_session {
|
||||
if let Some(connection_id) = connection_id.as_deref() {
|
||||
let catalog = catalog_from_pool_key(¤t_pool_key).map(str::to_string);
|
||||
let recovery = if prepare_agent_session {
|
||||
@@ -6501,13 +6549,12 @@ async fn execute_on_pool_once(
|
||||
sql: &str,
|
||||
max_rows: Option<usize>,
|
||||
) -> Result<db::QueryResult, String> {
|
||||
let (_connection_id, _database, _db_type, query_timeout_secs) = transfer_pool_context(state, pool_key).await;
|
||||
let (_connection_id, _database, db_type, query_timeout_secs) = transfer_pool_context(state, pool_key).await;
|
||||
let query_timeout = query_timeout_duration(query_timeout_secs);
|
||||
|
||||
// Read-only check: block transfer operations in readonly mode.
|
||||
crate::query::check_read_only_for_connection(state, pool_key, sql).await?;
|
||||
let pool_handle = state.pool_handle(pool_key).await;
|
||||
let pool = pool_handle.as_ref().ok_or("Connection not found")?;
|
||||
let pool = ensure_transfer_statement_pool(state, pool_key).await?;
|
||||
|
||||
// Transfer reads run under the per-connection operation budget. Drivers that
|
||||
// expose an incremental result stream (MySQL, PostgreSQL, SQLite, SQL Server)
|
||||
@@ -6517,7 +6564,7 @@ async fn execute_on_pool_once(
|
||||
// whose protocol returns the whole result in one shot — ClickHouse, InfluxDB,
|
||||
// the Agent/JDBC path, external drivers and the DuckDB sidecar worker — expose
|
||||
// no incremental progress, so they keep the plain wall-clock timeout.
|
||||
let result = match pool {
|
||||
let result = match &pool {
|
||||
PoolKind::Mysql(p, mode) => {
|
||||
let p = p.clone();
|
||||
let bare = *mode == crate::connection::MysqlMode::Bare;
|
||||
@@ -6627,11 +6674,17 @@ async fn execute_on_pool_once(
|
||||
}
|
||||
_ => Err("Unsupported database type for transfer".to_string()),
|
||||
};
|
||||
drop(pool_handle);
|
||||
if result.as_ref().is_err_and(|error| is_transfer_query_timeout(error)) {
|
||||
drop(pool);
|
||||
if result.as_ref().is_err_and(|error| is_transfer_query_timeout(error))
|
||||
&& should_discard_pool_after_query_timeout(db_type)
|
||||
{
|
||||
// A timed-out native driver future may still own a checked-out
|
||||
// connection. Discard the pool so a late server response cannot be
|
||||
// reused by the next transfer statement.
|
||||
// reused by the next transfer statement. Drivers that keep their pool on
|
||||
// a timeout (`pool_error_action` -> `Keep`, e.g. the embedded SQLite
|
||||
// worker) must not lose it here either: the pool key is shared by every
|
||||
// table in the transfer, so dropping it would make all later tables fail
|
||||
// with "Connection not found".
|
||||
state.remove_pool_by_key(pool_key).await;
|
||||
}
|
||||
result
|
||||
@@ -18385,4 +18438,71 @@ CREATE INDEX items_name_idx ON public.items (id);"#;
|
||||
|
||||
assert_eq!(receiver.len(), TRANSFER_PROGRESS_CHANNEL_CAPACITY);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn discarded_transfer_pool_is_replaced_for_native_drivers() {
|
||||
// Discarding a pool never replays the failed statement, but the rest of a
|
||||
// multi-table transfer keeps reusing the same pool key, so the pool has to
|
||||
// be replaced for later tables. Only agent/JDBC sessions are excluded: their
|
||||
// recovery stays gated on the structured quarantine decision.
|
||||
for db_type in [
|
||||
DatabaseType::Mysql,
|
||||
DatabaseType::Postgres,
|
||||
DatabaseType::SqlServer,
|
||||
DatabaseType::ClickHouse,
|
||||
DatabaseType::Gaussdb,
|
||||
DatabaseType::Sqlite,
|
||||
] {
|
||||
assert!(
|
||||
should_reconnect_discarded_transfer_pool(Some(db_type)),
|
||||
"{db_type:?} must prepare a replacement pool for later tables"
|
||||
);
|
||||
}
|
||||
for db_type in [DatabaseType::Oracle, DatabaseType::Dameng] {
|
||||
assert!(
|
||||
!should_reconnect_discarded_transfer_pool(Some(db_type)),
|
||||
"{db_type:?} keeps the stricter agent session rule"
|
||||
);
|
||||
}
|
||||
assert!(!should_reconnect_discarded_transfer_pool(None));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn every_discard_decision_leaves_a_usable_pool_behind() {
|
||||
// Connection drops on a write are downgraded to Discard (the write is never
|
||||
// replayed) and query timeouts tear the pool down too. Whatever produced the
|
||||
// Discard, a native driver must still get a pool for the next table instead
|
||||
// of the misleading "Connection not found" seen in the wild.
|
||||
for (db_type, error) in [
|
||||
(DatabaseType::Postgres, "connection reset by peer"),
|
||||
(DatabaseType::Mysql, "Connection timed out"),
|
||||
(DatabaseType::Postgres, "Query timed out after 30 seconds"),
|
||||
(DatabaseType::Sqlite, "PostgreSQL schema.reset cleanup failed: schema.reset timed out after 3 seconds"),
|
||||
] {
|
||||
assert_eq!(
|
||||
transfer_pool_error_action(TransferExecutionSafety::WriteNoReplay, Some(db_type), error),
|
||||
PoolErrorAction::Discard,
|
||||
"{db_type:?} must discard its pool for: {error}"
|
||||
);
|
||||
assert!(
|
||||
should_reconnect_discarded_transfer_pool(Some(db_type)),
|
||||
"{db_type:?} must be handed a replacement pool after discarding"
|
||||
);
|
||||
}
|
||||
// Errors that keep the pool never reach the reconnect branch.
|
||||
assert_eq!(
|
||||
transfer_pool_error_action(
|
||||
TransferExecutionSafety::WriteNoReplay,
|
||||
Some(DatabaseType::Postgres),
|
||||
"duplicate key value violates unique constraint \"items_pkey\""
|
||||
),
|
||||
PoolErrorAction::Keep
|
||||
);
|
||||
// Agent sessions only replace the session on a structured quarantine hint.
|
||||
assert!(!should_prepare_fresh_agent_transfer_session(
|
||||
Some(DatabaseType::Oracle),
|
||||
"Query timed out after 30 seconds"
|
||||
));
|
||||
assert!(!should_reconnect_discarded_transfer_pool(Some(DatabaseType::Oracle)));
|
||||
}
|
||||
}
|
||||
@@ -1429,7 +1429,7 @@ fn options_for_sequential_statements(
|
||||
statement_options
|
||||
}
|
||||
|
||||
fn should_discard_pool_after_query_timeout(db_type: Option<DatabaseType>) -> bool {
|
||||
pub(crate) fn should_discard_pool_after_query_timeout(db_type: Option<DatabaseType>) -> bool {
|
||||
let Some(db_type) = db_type else {
|
||||
return false;
|
||||
};
|
||||
|
||||
Reference in new issue
Block a user