Problem
Sequential multi-file runs slow down quadratically with the number of test files. The framework sources every test file into the main shell and never unsets the test functions, so the shell grows as the run progresses — and since every test executes in a $() subshell, every fork copies an ever-fatter process. The cost is fork overhead (sys time), invisible to any per-file profile.
Measured on bashunit's own unit suite (63 files, 1072 tests, macOS, Bash 3.2):
| Files run |
Wall |
Marginal cost/file |
| first 10 |
3.0s |
0.30s |
| first 30 |
13.9s |
0.55s |
| first 45 |
26.3s |
0.82s |
| all 63 |
~64s |
~2s |
Smoking gun: the last 18 files take 8.7s standalone but ~38s as the tail of the full run — a 4.4x penalty purely from accumulated shell state. Sys time explodes correspondingly (10 files: 1.3s → 45 files: 8.6s).
Fix (prototype validated)
Unset the file's test functions after the file has been processed, next to the existing hook cleanup (bashunit::runner::clean_set_up_and_tear_down_after_script, which already unsets set_up/tear_down/set_up_before_script/tear_down_after_script — src/runner.sh:2069).
Prototype diff (sequential + parallel path, after run_tear_down_after_script in bashunit::runner::load_test_files, ~src/runner.sh:494):
local _fn
for _fn in $_cached_fns; do
unset -f "$_fn" 2>/dev/null || true
done
_cached_fns is the output of runner::functions_for_script (the file's test functions). unset -f is a builtin — zero forks, Bash 3.0-safe.
Prototype results (identical pass/skip/snapshot counts before/after: 1065/5/2, assertions 1447/5/2):
| Suite |
Before |
After |
./bashunit tests/unit/ |
~64s |
21.7s |
./bashunit --parallel tests/unit/ |
~26s |
6.8s |
CPU time dropped from ~61s to ~19.5s (sequential). Parallel benefits too: workers fork from the main shell, so they inherit the same bloat today.
Why it's safe
- Unset happens after
call_test_functions returns (sequential) or after the & worker has forked (parallel — the subshell got its copy at fork time), and after run_tear_down_after_script.
- Data-provider functions and file-local helper functions are not in
functions_for_script output and stay defined (same as today).
- Cross-file behavior today: a later file's same-named test function silently overwrites the earlier one (per-file duplicate check only) — unsetting produces the same observable results.
- Header count (
find_total_tests) sources files in its own subshell, unaffected.
Implementation notes
- TDD: RED test first — e.g. a runner test asserting a test function from file A is no longer defined when file B is being processed (or a timing-free proxy: after
load_test_files over two fixture files, declare -F no longer lists file A's tests). Study tests/unit/runner_test.sh patterns.
- Word-splitting the newline-separated
$_cached_fns with default IFS is intended; function names contain no glob chars, but keep the loop consistent with existing style in load_test_files.
- Watch the fork-budget acceptance tests (
tests/acceptance/bashunit_coldstart_forks_test.sh, bashunit_run_forks_test.sh) — unset -f adds no forks, so they must stay green as-is.
- Check interactions:
--rerun-failed, --retry, --random-order, --shard, providers, --parallel (aggregation), --stop-on-failure.
- CHANGELOG entry under
## Unreleased → ### Changed (perf, no behavior change).
Constraints
- Bash 3.0+ (
.claude/rules/bash-style.md); unset -f qualifies
- Console output byte-identical (snapshot/acceptance suites green)
make sa, make lint clean
Acceptance criteria
Verification commands
time ./bashunit tests/unit/
time ./bashunit --parallel tests/unit/
./bashunit tests/ && ./bashunit --parallel tests/ && ./bashunit --parallel --simple --strict tests/
make sa && make lint
Problem
Sequential multi-file runs slow down quadratically with the number of test files. The framework sources every test file into the main shell and never unsets the test functions, so the shell grows as the run progresses — and since every test executes in a
$()subshell, every fork copies an ever-fatter process. The cost is fork overhead (sys time), invisible to any per-file profile.Measured on bashunit's own unit suite (63 files, 1072 tests, macOS, Bash 3.2):
Smoking gun: the last 18 files take 8.7s standalone but ~38s as the tail of the full run — a 4.4x penalty purely from accumulated shell state. Sys time explodes correspondingly (10 files: 1.3s → 45 files: 8.6s).
Fix (prototype validated)
Unset the file's test functions after the file has been processed, next to the existing hook cleanup (
bashunit::runner::clean_set_up_and_tear_down_after_script, which already unsetsset_up/tear_down/set_up_before_script/tear_down_after_script— src/runner.sh:2069).Prototype diff (sequential + parallel path, after
run_tear_down_after_scriptinbashunit::runner::load_test_files, ~src/runner.sh:494):_cached_fnsis the output ofrunner::functions_for_script(the file's test functions).unset -fis a builtin — zero forks, Bash 3.0-safe.Prototype results (identical pass/skip/snapshot counts before/after: 1065/5/2, assertions 1447/5/2):
./bashunit tests/unit/./bashunit --parallel tests/unit/CPU time dropped from ~61s to ~19.5s (sequential). Parallel benefits too: workers fork from the main shell, so they inherit the same bloat today.
Why it's safe
call_test_functionsreturns (sequential) or after the&worker has forked (parallel — the subshell got its copy at fork time), and afterrun_tear_down_after_script.functions_for_scriptoutput and stay defined (same as today).find_total_tests) sources files in its own subshell, unaffected.Implementation notes
load_test_filesover two fixture files,declare -Fno longer lists file A's tests). Studytests/unit/runner_test.shpatterns.$_cached_fnswith default IFS is intended; function names contain no glob chars, but keep the loop consistent with existing style inload_test_files.tests/acceptance/bashunit_coldstart_forks_test.sh,bashunit_run_forks_test.sh) —unset -fadds no forks, so they must stay green as-is.--rerun-failed,--retry,--random-order,--shard, providers,--parallel(aggregation),--stop-on-failure.## Unreleased→### Changed(perf, no behavior change).Constraints
.claude/rules/bash-style.md);unset -fqualifiesmake sa,make lintcleanAcceptance criteria
./bashunit tests/unit/wall time drops to ~1/3 of baseline (~64s → ~22s on the reference machine)./bashunit --parallel tests/unit/similarly improved (~26s → ~7s)./bashunit tests/,./bashunit --parallel tests/,./bashunit --parallel --simple --strict tests/make sa,make lintpassVerification commands