Fresh-backend mode

scripts/pgtest.sh fresh [TEST...] runs the regression tests of test/sql with a new backend for every top-level statement, to find the bugs of the first statement of a backend: a table or a shared structure that is attached lazily, a cache that is empty, a memory context that does not exist yet. A test that passes in one long session and fails here (or the other way round) has a statement that depends on what the statements before it left in its backend. Both wrong answers and the cold-backend crash of a table slot made in a short-lived context are of this kind.

How: scripts/fresh-psql.py stands first in PATH as psql for pg_regress only. It reads the script that pg_regress feeds on stdin and puts a silent \connect before every statement that is not inside a transaction block, replaying the SET/RESET of the session first. The output is psql’s, so the expected files apply unchanged; the first difference is what the cold backend does differently.

From the first statement that binds the session to its backend (PREPARE, DECLARE, LISTEN, LOAD, a temporary table, a session advisory lock, set_config, setseed, \if, \i) the rest of the file runs in the backend it has, because a new connection would lose what the file uses afterwards. FRESH_STATS=FILE appends a line new backends<TAB>statements per run (pgtest.sh fresh writes it next to the results).

A setting of the extension (sift.*) is unknown to a backend that has not loaded the library, as for any extension that is not preloaded: a test that sets one out of range or shows one fails in the default mode. FRESH_LOAD=pg_sift loads the library first in every new backend, as session_preload_libraries would, and leaves the cold first statement to the lazily attached state only (FRESH_LOAD=pg_sift scripts/pgtest.sh fresh TEST...).

Limits that keep it safe on a host: the shim refuses to run without FRESH_PSQL (the real psql, set by pgtest.sh), inside itself (FRESH_ACTIVE) or as its own psql, so it cannot start itself without end; pgtest.sh fresh does not nest and takes one of SIFT_FRESH_RUNS locks ($XDG_RUNTIME_DIR/sift-fresh.N.lock; default 1, at most 4), so no more fresh runs than that are at work on a host, each on a cluster of its own, and the tests of a run follow one another. Calls of psql with a command of their own (-c, -f) are not made fresh.

Cost: a connection per statement, about 2 ms with the unix socket, so a file of a few hundred statements takes a second or two longer. Run it over the files that touch src/agg/exec, src/col/am and src/table (the focus of scripts/ci.sh affected), and over the whole suite when the lazily attached state changes. Self-test: test/t/fresh_backend.sh; the crash class of a slot in the wrong context is the slot_owner test of the assert lane.