Twins that start their own fixture servers collide under pytest -n: pick_port gives every xdist worker the same port #1742

Three reviews found the same race independently:

• test: iceberg_objstore gains its pytest twin, and an endpoint the allow-list never refused no longer passes (#1341) #1635 (iceberg_objstore, my review): red under -n 2;
• test: objstore_listing gains its pytest twin, and each listing refusal is read by its own guard (#1341) #1714 (objstore_listing): red under -n 2 in 3 runs of 3; green under -n 0 and --dist loadfile;
• test: objstore_sink_write gains its pytest twin, and multipart is read as S3 does it (#1341) #1721 (objstore_sink_write): both workers take pick_port(30); red in 3 runs of 3 under -n 2, 36/36 serially.

CI runs the cluster files serially, so CI is green. But the guard job runs with -n 4, a developer running -n gets false reds, and #1740 notes that moving the cluster job to xdist (to cut its 17-minute run) waits on this.

Why: a module fixture that starts an S3/HTTP fixture server picks its port with pick_port(base), which does not depend on the xdist worker. Two workers probe the same free port at the same moment and both bind it (or one fails with Address already in use).

Next step: a guard that runs one such twin under -n 2 and requires green, red first; then make the port depend on the worker (as pgc_cluster does for its own clusters), or bind port 0 and read back the port.

🤖 Generated with Claude Code