When connected to a Greenplum cluster, schema introspection now surfaces
the two attributes that dominate query performance there: the data
distribution policy (DISTRIBUTED BY / RANDOMLY) and the storage kind
(heap / AO row / AO column). Without these, the agent writes
PG-optimal SQL that triggers Redistribute Motion in GP and runs orders
of magnitude slower than necessary.
- track Greenplum major version (6 vs 7) at connect time, since GP6
uses pg_class.relstorage and GP7 dropped it in favor of pg_am
- new fetch_gp_table_extras helper queries gp_distribution_policy and
the version-appropriate storage catalog, returns a per-table map
- format_table_block prints ` -- GP: DISTRIBUTED BY (...) | STORAGE: ...`
under the TABLE header when extras are available
- build_overview_postgres appends a GREENPLUM NOTES block with the
distribution-aware-join rules and a skew-detection one-liner; the
agent sees this in every system prompt on a GP connection
- build_schema_context (legacy generate_sql / explain_sql / fix_sql_error
path) and the chat get_columns tool both feed extras into format_table_block
P0 of the GP support arc — partitioning, EXPLAIN motion-aware parsing,
external tables, resource queues, and a skew-check tool are deliberately
deferred to follow-up commits.
Aleksey Shakhmatov
5c5d256cee