StarRocks/Doris expose external storage through a catalog. dbx had no
catalog concept, so connecting to e.g. `paimon_catalog.clip` failed: the
URL path's database was sent as the MySQL handshake schema and rejected
before any `SET catalog` could run, and table listing/preview ran against
the default catalog.
Add an opt-in `catalog=<name>` URL parameter for mysql-family connections:
- `mysql_connection_catalog`: reads the `catalog` param (percent-decoded).
- `mysql_setup_queries`: emits `SET catalog = <name>` when present. Pushed
last so mysql_async's back-to-front setup execution (Vec::pop) runs it
before `USE <database>`, which would otherwise fail (the database lives
in the external catalog). The pool re-runs setup after each reset, so the
catalog stays current.
- `mysql_async_url`: when a catalog is configured, strip the database path
from the URL handed to mysql_async so the handshake does not send the
external-catalog database as the schema. The original URL (with path) is
still used for setup, so `USE <database>` is emitted correctly.
- Add `catalog` to `is_dbx_handled_mysql_url_param` so mysql_async ignores
the unknown param.
This is opt-in: only connections that set `catalog=` are affected; all
existing mysql/StarRocks/Doris connections behave unchanged. Configure a
Paimon connection as `database=clip`, `url_params=catalog=paimon_catalog`.
Verified against a StarRocks + Paimon catalog via the dbx MCP bridge:
SHOW DATABASES, SHOW TABLES, DESCRIBE, and unqualified `SELECT * FROM <t>`
all resolve in the selected catalog.
Some SSH proxies and jump-hosts authenticate with the "none" method,
requiring no password or key. dbx only tried publickey/password/agent,
so connecting through such a host failed with "SSH public key
authentication failed" — and the frontend forced users to enter a
credential just to pass validation.
- Backend: probe authenticate_none first in connect_and_authenticate;
on rejection fall back to the configured credential method, matching
standard openssh behavior.
- Frontend: drop the "password/key/agent required" check in
validateTransportLayers so passwordless hops can be saved and tested.
Refs #1628
Root cause:
- AgentDriverClient::kill() is called during Drop to clean up child processes
- Old implementation called child.wait() and ignored errors with let _
- When wait() fails (e.g., process already killed by signal), child is not reaped and becomes zombie
Fix:
- Use try_wait() non-blocking polling, up to 100ms timeout
- Avoid long blocking in Drop during async cleanup
- Still perform final blocking wait() as fallback after timeout
- Log all wait errors for debugging
Impact:
- All JDBC agents (Oracle/DB2/DM8/XuguDB etc.)
- Prevents zombie process accumulation during long-running sessions
Disable COM_RESET_CONNECTION for single-connection session pools
(max_connections == 1) so that MySQL session state such as TEMPORARY
TABLEs is not cleared when a connection is returned to the pool.
Closes#1509
t8y2