testsuite: don't run transport-agnostic archiver tests over rest:// (#9324)

[ThomasWaldmann]

What

Stop running the remote (rest://) archiver variant for tests whose behavior does not depend on the repository transport.

Why

Profiling the testsuite (-n auto, no coverage) showed every one of the 40 slowest tests was a remote_archiver variant. The remote variant spawns a borgstore-server-rest subprocess and does synchronous HTTP-over-stdio round trips per object, so a remote run costs several times its local twin — e.g. test_prune_repository_example: ~2.8s local vs ~12.2s remote (4.4x).

Most of those tests exercise pure command logic that operates on metadata or only formats output (prune scheduling, diff/list/info rendering, rename, delete/undelete, recreate, key handling, debug, return codes, fuse mount, lock_cmds, repo-create/repo-delete, repo-space, tar). That logic is identical regardless of transport — none of it calls a store method whose implementation actually differs per backend — and the wire-transport-specific behavior is already covered by the create/extract/check/compact/transfer tests.

What changed

• 20 archiver test files switched from kinds="local,remote,binary" (or "local,remote") to kinds="local,binary" (or "local").
• generate_archiver_tests()'s docstring now states the rule directly: only keep remote where a backend genuinely branches its implementation per transport (not just "moves real archive data"), and that branch isn't already exercised by another remote-tested file.
• Added check_cmd_remote_test.py: check_cmd_test.py itself dropped remote (its tests only exercise generic store list/load/delete, same as prune/list/etc.), but Repository.check()'s repository-only mode relies on store.hash(), which the REST backend computes server-side (nothing downloaded) instead of the default download-then-hash. No existing test actually exercised a plain --repository-only check catching pack corruption via that hash comparison (the existing corruption tests all go through --verify-data, which decrypts client-side instead), so this closes a real coverage gap rather than just moving it.

remote is now kept only where the store I/O pattern itself genuinely differs by backend:

• create / extract: Repository.get()'s store.load(offset=, size=) (partial reads — HTTP Range requests for REST vs. seek/read elsewhere) and store.store()'s content-hash upload header.
• check (new check_cmd_remote_test.py): store.hash() computed server-side.
• compact: its own store-level defrag/pack-rewrite path.
• transfer: the only command that opens two live repository/store connections at once (source + destination).
• Plus the already-local,remote checks_test.py (shared setup/teardown for the check suite).

Commands whose remote testing turned out to be redundant on closer look (all confirmed by reading the actual command implementation, not just assumption):

• tar cmds (export-tar/import-tar): route through the exact same Archive/Repository.get()/put() machinery as extract/create, so their own remote run is redundant — same category as recreate.
• repo-create / repo-delete: store.create()/store.destroy() are trivial 1:1 RPC forwards in the REST backend (no content-dependent divergence); still incidentally exercised over the wire by every remaining remote test file's own repo-create setup step, and by checks_test.py's repo-delete.
• repo-space: grepped borg's own code for .quota( — it's never called. repo_space_cmd.py only does generic store_list/store_store/store_delete on config/space-reserve.*, despite quota() being a genuinely backend-specific method in borgstore that borg simply doesn't use.
• lock_cmds: Lock only does generic list/store/delete on locks/*; its one host-specific check (process_alive() across hosts) isn't exercised differently by this test infra regardless of "kind" anyway (client hostid is the same host either way).

Effect

Two independent measurements, both -n auto, no coverage, -k "not binary":

The 374 → 178 → 128 invocation counts and the original 242.7s → 154.2s wall-clock come from the PR author's machine (see the numbers earlier in this description). On top of that, this update was verified with a same-machine, apples-to-apples before/after (12-core machine, checked out the pre-PR commit into the working tree, ran the full archiver suite, then restored the branch and re-ran):

	before this PR (pre-#9837, same machine)	after this update
archiver suite wall-clock (-n auto)	219.7s (1 failed, 913 passed, 55 skipped)	96.0s (683 passed, 39 skipped)
remote_archiver invocations (archiver suite)	374	128 (−66%)

The one failure in the "before" run (test_fuse_allow_damaged_files, a FUSE-daemonization timeout) is an unrelated, known-flaky test (stale/slow FUSE mount teardown), not caused by this change — it did not reproduce in the "after" run. Test counts differ between the two runs because kind-pruning also removes the parametrized [remote_archiver] variants themselves (not just their execution time), so "after" collects fewer test IDs overall.

Notes for review

The categorization is per-file (the existing granularity). Previously-open borderline calls, now resolved:

• recreate / tar — both rewrite/stream chunk data; recreate was already dropped from remote (covered by create+extract); tar is now dropped too, for the same reason (redundant with create+extract's coverage of the actual wire-specific Repository.get()/put() behavior).
• check_cmd was the single biggest time sink and was originally kept whole on remote "since it is borg's integrity command." On closer inspection, only its store.hash()-based repository-only corruption check is actually backend-specific; the rest of check_cmd_test.py's 17 tests are transport-agnostic (metadata presence/absence, --verify-data decryption-based corruption which never touches store.hash()). Dropped remote from check_cmd_test.py and added one small dedicated check_cmd_remote_test.py test for the hash-comparison path instead — which also fixed a real gap, since no prior test exercised plain --repository-only corruption detection under any transport.
• The two remaining get_kind() == "remote" branches in the touched files (test_repo_list_from_borg1, test_migrate_lock_alive) were already pytest.skip("only works locally"), so nothing was left dead.

Refs #9324.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.17%. Comparing base (1206808) to head (dbadb32).
⚠️ Report is 17 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #9837      +/-   ##
==========================================
+ Coverage   85.14%   85.17%   +0.02%     
==========================================
  Files          93       93              
  Lines       15337    15346       +9     
  Branches     2306     2308       +2     
==========================================
+ Hits        13059    13071      +12     
+ Misses       1586     1581       -5     
- Partials      692      694       +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.