[tamnd]

Summary

A join that makes more tuples than 1 GiB holds now gives the error out of memory (53200). Before, the server used all the memory of the machine and stopped. The executor keeps all the tuples of a join in memory, so this limit keeps the server up until the streaming executor comes.

In the regression suite, the cross join of tenk1 with itself in groupingsets makes 100 million tuples. The server stopped there, and about 80 tests after it failed because they could not connect.

Changes

• join and the loop of LATERAL in rupg-exec/src/scan.rs stop with an error when the join makes more tuples than join_limit lets it. The limit is 1 GiB divided by the size of a tuple. The tests use 4 MiB.
• The error has the detail "The join makes more than N rows, and rupg keeps all the rows of a join in memory."

Known differences

PostgreSQL streams the rows of a join and gives the result. rupg gives an error for a join with more than about 27 million tuples of two relations.

Tests

A test in rupg-exec runs a cross join of 90,000 tuples, which passes, and one of 160,000 tuples, which gives 53200. fmt, clippy, the workspace tests and the checks of xtask pass on server1.