DataRobot JDBC SQL

DataRobot opens a JDBC connection from its own network with the URL and driver properties you send. It runs one statement and returns the rows or a status. Nothing is registered in DataRobot.

If the database is already a DataRobot data connection, use datarobot-connectivity-datastore-sql: it uses the stored credential, so the agent never handles secrets.

Prerequisites

• Auth: set both DATAROBOT_ENDPOINT (ends in /api/v2) and DATAROBOT_API_TOKEN, or neither, in which case both come from ~/.config/datarobot/drconfig.yaml. An --env-file can also supply them. If auth fails, invoke datarobot-setup.
• Python 3.9+, standard library only. Run the script by its path in this skill's folder: python3 <skill-dir>/scripts/jdbc_sql.py. The .env stays in the user's project; pass its absolute path.
• Network: DataRobot itself opens the connection, so the database must be reachable from DataRobot: its cloud on SaaS, or the customer's cluster on self-managed installs. A database reachable only from the user's machine or private network, localhost included, will not work.

Secrets stay out of the chat

• Never ask the user to paste a password, token, or key into the chat, and never echo one. Have the user put the values in a gitignored .env, editing it in their own editor:

  JDBC_URL=jdbc:postgresql://<host>:5432/<db>
  DB_USER=<user>
  DB_PASSWORD=<password>
  

• If the user already pasted a secret, don't repeat it. Ask the user to put it in .env, or write it there with a file edit; a shell command can leave it in logs. Recommend rotating it unless it is a published public credential.

• Pass secrets as driver properties read from the environment: --env-file <path>/.env --param-env user=DB_USER --param-env password=DB_PASSWORD. Non-secret values can use --param key=value. Prefer properties over credentials embedded in the URL.

• --env-file loads KEY=VALUE lines. It skips comments, strips quotes, never overrides variables that are already set, and can include DataRobot auth. Without --jdbc-url, the script reads JDBC_URL (or Talk to My Data's JDBC_URI) together with the JSON object in JDBC_CONNECTION_PARAMETERS. With an explicit URL it uses only the properties you pass.

• Masking: all output masks secrets. A plan names each secret's source, e.g. *** (env DB_PASSWORD). The confirm token covers the source, not the value: rotating a password in .env needs no new approval, but switching to another variable does.

Script

jdbc_sql.py <command> -h lists every option. Every command except routes takes the connection flags --jdbc-url or --jdbc-url-env VAR, --param, --param-env, --params-json-env, and --env-file.

Command	Purpose
check	Connects and logs in (SELECT 1, or SELECT 1 FROM DUMMY on SAP). Does not check read access to tables
tables [--catalog C --schema S --like TEXT]	Tables and views, sorted (default --max-rows 1000). --like matches part of the name, any case
columns --table T [--catalog C --schema S]	Column names, types, and nullability, in table order
query --sql "…" | --sql-file F [--bind '[42, "a"]']	One read-only query (default --max-rows 100, max 10000)
plan --sql "…" | --sql-file F [--bind …]	DML/DDL plan and confirm token. Makes no API call
execute <same arguments> --confirm TOKEN	Runs the approved plan
routes	Which JDBC routes this DataRobot and API token can use, with bind support. Sends no SQL

Output options:

• --format table|csv|json: CSV sends the # notes to stderr.
• --show N: rows printed, default 20 (100 for tables and columns).
• --width N: cell width.
• --out FILE: saves every fetched row (JSON for .json, else CSV); stdout then shows only a preview.

Exit codes: 0 ok · 1 usage · 2 API error, or a write whose outcome is unknown · 3 the confirm token no longer matches · 4 the database reported a write failure · 5 route unavailable (routes shows which).

Workflow

- [ ] 1. URL and secrets in env/.env (never in chat); URL format from references/jdbc-urls.md
- [ ] 2. `check`: fix connectivity before anything else
- [ ] 3. Discover: `tables --schema <s> --like <text>`, `columns --table <t>`
- [ ] 4. Read: `query` with WHERE/LIMIT/aggregation
- [ ] 5. Write, only if the user asked: checks → `plan` → approval → `execute --confirm` → verify

Empty discovery: catalog views list only what the login can access. tables explains an empty result. On PostgreSQL, tables and columns show objects the login can't read, marked readable=False. A 403 permission denied means missing grants, not a wrong password.

Writes:

1. Check connectivity. check must pass for the target before you plan. If it fails, stop and resolve it with the user.
2. Check the target and impact. Confirm the target exists and is a table, not a view (tables, columns). If it doesn't, stop and ask the user. Measure the impact with a read: SELECT COUNT(*) FROM t WHERE <same condition> for UPDATE or DELETE; required columns and existing keys for INSERT.
3. Plan. plan … --sql "<sql>" prints WRITE PLAN (nothing has run) and a token. Show the plan to the user and ask for an explicit yes.
4. Execute. Run execute with the same arguments plus --confirm <token>. The token covers the URL, the properties (for a secret, its source), SQL and binds. Any change means a new plan and a new approval (exit 3).
5. Verify. No row count comes back, so verify DML with query and DDL with tables/columns.

How DataRobot runs the SQL

• Reads run inside a subquery with a row limit, e.g. select * from (<sql>) as "_table" fetch next N rows only on PostgreSQL.
  • Send one query. The script strips a trailing ; and trailing comments.
  • Your own LIMIT is fine and saves the database work.
  • SHOW, DESCRIBE and EXPLAIN cannot run: use tables/columns or catalog queries.
  • SQL Server: column names must be unique, ORDER BY needs TOP, and CTEs fail.
  • BigQuery: quote names with backticks, as in `project.dataset.table`; tables and columns need --schema <dataset>.
  • Error positions count from the start of the wrapped SQL.
• Writes are one JDBC execute() call with the driver's autocommit. There is no transaction across statements, and most drivers reject several statements in one call, so send one statement per execute. The server cancels the call after 30 s.
• Values come back as strings or NULL; the # types: line comes from resultSchema. A # warning: hit maxRows line means rows were probably cut.

Rules

• Never pass --confirm until the user has approved that exact plan.
• Never retry a write after exit 2 until a read shows its effect: the outcome is unknown.
• Writes run as the database user in the properties: suggest a least-privileged user.
• query runs only reads. execute runs only SQL the user approved verbatim in a plan; an app must never send generated SQL to /jdbcExecution/ by itself.
• Keep results small; save large results with --out FILE.

Common errors

Error	Fix
400 The connection attempt failed.	DataRobot cannot reach the host: check the host, the port, and that the database firewall allows DataRobot
400 No suitable JDBC driver…	Unsupported URL prefix; see references/jdbc-urls.md
403 … Details: … naming a privilege (permission denied, insufficient privileges, command denied)	The login works but lacks grants (USAGE on the schema, SELECT on the table)
403 … Details: … (authentication)	The database rejected the user or the password
400 SQL [select * from (…)…]; …	A SQL error in the wrapped query; check the dialect notes
408 Request timed out…	query and discovery stop at 300 s: narrow the query, or let a cold warehouse start. execute stops at 30 s, login included: warm the warehouse with a query first. A timed-out write may still have run
422 … is not allowed key	Use camelCase fields. JDBC routes have no credentialId

Full list: references/api-reference.md.

Wiring an app (Talk to My Data)

To move an app's SQL onto these routes, follow references/app-integration.md: a minimal client, type restoration, timeouts and truncation, and the Talk to My Data migration checklist.

In Talk to My Data, the JDBC-URL source maps to /jdbcDataPreview/. Its DataRobot data-connection sources map to datarobot-connectivity-datastore-sql.

Using DataRobot SDK

The script calls REST directly. That avoids SDK client retries, which could run a write twice, and keeps driver property names unchanged. In the datarobot SDK:

• JdbcPreview.execute_update(...) calls /jdbcExecution/.
• JdbcPreview.preview(...) in 3.19+ calls /jdbcPreviewQuery/, which some accounts and app API keys cannot use (403). For reads, prefer a direct POST {DATAROBOT_ENDPOINT}/jdbcDataPreview/.

References

• references/api-reference.md: request and response contract, limits, errors
• references/jdbc-urls.md: URL templates, auth properties, dialect limits, catalog SQL
• references/app-integration.md: Python clients, types, and the Talk to My Data checklist

Related skills

• datarobot-connectivity-datastore-sql: SQL through a DataRobot data connection with stored credentials
• datarobot-setup: authentication and SDK setup