Expose PG-compatible status/transaction_status/async_exec on the JDBC connection (+ CI Postgres 14)

[skunkworker]

Summary

Under JRuby the PostgreSQL raw connection is a PostgreSQLJdbcConnection, not a PG::Connection. ActiveRecord 7.2's PostgreSQL reconnection test suite calls PG::Connection methods and ::PG::* constants directly on that raw connection, so they blew up with NoMethodError / uninitialized constant PG and aborted those tests.

This PR makes the JDBC connection answer the small libpq-compatible API those paths invoke, and bumps the CI Postgres image to a supported release.

Changes

• status, transaction_status, async_exec on PostgreSQLRubyJdbcConnection (Java).
• ::PG::PQTRANS_* / ::PG::CONNECTION_* constants (lib/arjdbc/postgresql/pg_compat.rb), defined only when the real pg gem is absent.
• CI: PostgreSQL 11 → 14 (separate commit) — PG 11 is EOL and lacks pg_current_xact_id() (PG 13+), which the AR 7.2 suite uses.

Why the PG module shim is necessary

The real pg gem is a C extension and is unavailable under JRuby — the entire reason AR-JDBC exists — so the top-level ::PG namespace and its libpq constants are simply not defined.

AR-JDBC's own production code never references ::PG. Verified by runtime method resolution:

ActiveRecord::Base.connection.method(:retryable_query_error?).owner
#=> ActiveRecord::ConnectionAdapters::AbstractAdapter   # abstract_adapter.rb:1079

It is AR's native PostgreSQL adapter that compares transaction_status to ::PG::PQTRANS_INERROR, but AR-JDBC is class PostgreSQLAdapter < AbstractAdapter (it does not subclass AR's native PG adapter), so that method is never in the ancestry. The other native ::PG sites are bypassed the same way:

• #retryable_query_error? resolves to AbstractAdapter (no ::PG), not postgresql_adapter.rb#L826-L831.
• #reset! is overridden by AR-JDBC, so postgresql_adapter.rb#L367-L372 is never reached.
• #cancel_any_running_query / IDLE_TRANSACTION_STATUSES live in AR's native postgresql/database_statements.rb, which AR-JDBC never loads (it ships its own): database_statements.rb#L162-L166.

So the shim's only consumer is the Rails test suite. Rails' remote_disconnect helper hard-codes ::PG::* on the raw connection, in files this gem cannot edit:

• test/cases/adapter_test.rb#L913-L921 — .status == ::PG::CONNECTION_BAD (L917), .transaction_status == ::PG::PQTRANS_INTRANS (L918), .async_exec("begin") (L919)
• test/cases/adapter_test.rb#L885-L893 — .transaction_status == ::PG::PQTRANS_INTRANS (L891)

Because those constants are resolved on the caller's side (== ::PG::PQTRANS_INTRANS), nothing our methods return can satisfy them — a top-level ::PG must exist or the tests raise uninitialized constant PG.

Why shim rather than exclude these tests? They exercise behavior AR-JDBC genuinely supports and cares about — reconnect/retry after a dropped backend, the same scenario as the connection-loss work in the sibling PRs. The ::PG references are only test scaffolding used to induce the disconnect, not the behavior under test. With the shim in place these tests run and drop from 26 errors → 3 errors / 4 failures; the remaining 7 are real reconnect-recovery and transaction-state gaps. Excluding the tests would throw away that coverage and mask those genuine gaps. Exclusions are for things that cannot work under JRuby; these can. Hence the shim — an ~8-line, version-stable libpq constants module guarded by unless defined?(PG) — is the correct minimal solution.

Why the constant values / state mapping are correct

• libpq ConnStatusType (CONNECTION_OK = 0, CONNECTION_BAD = 1) and PGTransactionStatusType (PQTRANS_IDLE = 0, ACTIVE = 1, INTRANS = 2, INERROR = 3, UNKNOWN = 4) — PostgreSQL source:
  libpq-fe.h#L56-L59 and libpq-fe.h#L114-L120
• transaction_status reads the driver's protocol-level state via BaseConnection#getTransactionState() and maps IDLE/OPEN/FAILED → PQTRANS_IDLE/INTRANS/INERROR. Driver enum:
  pgjdbc TransactionState.java#L8-L11
  (PGConnection exposes no transaction state, hence the unwrap to BaseConnection.)

Verification

Local run of test/cases/adapter_test.rb (jruby-9.4.14.0, AR 7-2-stable, --seed=12345):

	runs	failures	errors
before	67	1	26 (33× transaction_status, 13× status NoMethodError)
after	67	4	3

The remaining 3 errors / 4 failures are deeper reconnect-recovery and transaction-state-tracking gaps, tracked separately. This PR does not target the too many clients connection-leak cascade — that is a distinct issue.

🤖 Generated with Claude Code