One tool reads visit counts out of the audit log and another ranks tables by health score — questions you would otherwise write and then maintain SQL for. The rest is shaped to keep large output away from the model: a query result can be redirected to a file, the table overview caps its sample at three rows, and both overview tools answer from cache unless refreshed. The dump-and-profile tool is the deliberate exception, and its own description warns the output is very large.
StarRocks' own server over a cluster: SQL execution, query profiling, audit-log-derived table rankings and a chart path, plus starrocks:// and proc:// resources for schemas and internal system state.
- Result-returning queries and DDL or DML statements kept in separate tools, with a large result written to a file as CSV, TSV, JSON or JSONL instead of returned inline
- Query analysis from the cluster's own profile — by query id or by SQL — plus a full dump and profile collection for the cases that need offline processing
- Rankings the cluster computes about itself: hottest tables by audit-log visit count within a time window, and worst tables by health score
- Orientation before querying — a table overview with columns, up to 3 sample rows and a row count, and a database summary with table schemas and sizes, both served from cache until a refresh is asked for
- A chart path that runs SQL and a Plotly expression together, returned as an image or written out as an interactive HTML file
- A per-session default database, so the db argument stops being repeated on every call
- Internal cluster state as readable resources: frontends, backends, partitions, transactions, jobs and catalogs under proc://
Python 3.11 or newer, uv, and a reachable StarRocks FE service — port 9030, the MySQL protocol port rather than the HTTP one. Connection details come either as STARROCKS_HOST, STARROCKS_PORT, STARROCKS_USER, STARROCKS_PASSWORD and STARROCKS_DB, or as one STARROCKS_URL that overrides them; on macOS the password can come from Keychain instead. Grant the account only what you intend, since the write tool runs DDL and DML. TLS is not verified unless you set a CA and turn verification on. Files that tools write land on the machine running the server, which is not the client machine for an HTTP deployment. Apache-2.0.
Build from source — clone the repository and build it, then point your client at the binary